JP6056172B2 - Data processing apparatus, data processing method, and program - Google Patents

Description

  The present technology relates to a data processing device, a data processing method, and a program, and in particular, for example, a data processing device that makes it possible to easily and accurately determine the power consumption of each of a plurality of electrical devices in a home, The present invention relates to a data processing method and a program.

  For example, as a method of realizing so-called “visualization” of power consumption, etc., presenting individual power consumption and current consumption of electrical equipment such as home appliances (home appliances) in the home to users in the home For example, there is a method of attaching a smart tap to each outlet.

  The smart tap has a measurement function for measuring the power consumed by an electrical outlet (home appliance connected to) the smart tap and a communication function with the outside.

  In the smart tap, the power consumption measured by the measurement function is transmitted to the display, etc. using the communication function, and the power consumption from the smart tap is displayed on the display. Can be realized.

  However, attaching smart taps to all outlets in the home is not easy in terms of cost.

  In addition, home appliances such as a so-called built-in air conditioner (air conditioner) may be directly connected to a power line without going through an outlet, and a smart tap is used for such home appliances. That is difficult.

  Therefore, for example, in homes, it is called NILM (Non-Intrusive Load Monitoring), which calculates the power consumption of individual household appliances connected to the home based on the current information measured by the switchboard (distribution panel). Technology is drawing attention.

  In NILM, for example, using the current measured at one place, the power consumption of each home appliance (load) connected from there can be obtained without separately measuring.

  For example, Patent Document 1 discloses NILM technology that identifies active devices by calculating active power and reactive power from current and voltage measured at a single location and clustering the respective amounts of change. Yes.

  In the technique described in Patent Document 1, in order to utilize changes in active power and reactive power when the home appliance is on and off, change points of active power and reactive power are detected. For this reason, if the change point detection fails, it becomes difficult to accurately identify the home appliance.

  In addition, it is difficult to represent the operating state in two states, on and off, for recent home appliances. By simply using changes in active power and reactive power when on and off, home appliances can be accurately identified. It has become difficult to do.

  Thus, for example, Patent Documents 2 and 3 describe NILM technology that uses an LMC (Large Margin Classifier) such as SVM (Support Vector Machine) as a home appliance identification model (discrimination model).

  However, in NILM using an identification model, unlike generation models such as HMM (Hidden Markov Model), prepared learning data for each home appliance and learning the identification model using that learning data in advance. It is necessary to finish it.

  For this reason, it is difficult for NILM using an identification model to deal with home appliances that have not learned the identification model using known learning data.

  Therefore, for example, Non-Patent Documents 1 and 2 describe NILM technology using an HMM that is a generation model, which is not an identification model that requires prior learning using known learning data.

  However, in NILM using a simple HMM, as the number of home appliances increases, the number of HMM states becomes enormous, making it difficult to implement.

Specifically, for example, when the operating state of each home appliance is two states, on and off, the number of HMM states necessary to represent the operating state (combination) of M home appliances is , becomes 2 ^M pieces, the number of state transition probability is the square of the number of states (2 ^M) becomes ^two.

Therefore, the number M of household appliances in the home is not very large recently, for example, even if it is 20, the number of HMM states is a huge number of 2 ²⁰ = 1,048,576, and the transition probability The number of is the enormous number of the tera order of 1,099,511,627,776 of the square.

US Pat. No. 4,854,141 JP 2001-330630 A International Publication No. 01/077696 Pamphlet

Bons M., Deville Y., Schang D. 1994. Non-intrusive electrical load monitoring using Hidden Markov Models. Third international Energy Efficiency and DSM Conference, October 31, Vancouver, Canada., P. 7 Hisae Nakamura, Koichi Ito, Tatsuya Suzuki, "Electrical Equipment Monitoring System Based on Hidden Markov Model", IEEJ Transactions B, Vol. 126, No. 12, pp. 12231229, 2006

  At present, individual appliances such as home appliances (power-changing home appliances) such as air conditioners whose power consumption (current) fluctuates depending on modes and settings, etc. There is a demand for a proposal for NILM technology that can easily and accurately determine the power consumption and other factors.

  The present technology has been made in view of such a situation, and makes it possible to easily and accurately obtain individual power consumption and the like of an electric device.

A data processing device or program according to one aspect of the present technology performs state estimation for obtaining a speech state of the speaker using data representing a sum of speech signals superimposed with speech uttered by a plurality of speakers. The state estimation unit and waveform separation learning for obtaining a specific waveform related to the voice unique to the speaker are performed by using the state estimation by the state estimation unit and limiting frequency components within a frequency band that the voice can take. A program for causing a computer to function as a data processing device or a data processing device including a waveform separation learning unit.

A data processing method according to one aspect of the present technology performs state estimation for obtaining a speech state of the speaker by using data representing a sum of speech signals superimposed with speech uttered by a plurality of speakers, and the speaker A data processing method including a step of performing waveform separation learning for obtaining a specific waveform related to a sound specific to a voice by using the state estimation to limit frequency components within a frequency band that can be taken by the voice .

In one aspect of the present technology, state estimation is performed to determine the utterance state of the speaker using data representing the sum of speech signals superimposed with speech uttered by a plurality of speakers. Waveform separation learning for obtaining a specific waveform related to the voice is performed by limiting the frequency components in the frequency band that the voice can take using the state estimation .

  Note that the data processing device may be an independent device or an internal block constituting one device.

  The program can be provided by being transmitted via a transmission medium or by being recorded on a recording medium.

  According to one aspect of the present technology, it is possible to easily and accurately obtain individual power consumption of an electrical device.

It is a figure explaining the outline | summary of one Embodiment of the monitoring system to which the data processor of this technique is applied. It is a figure explaining the outline | summary of the waveform separation learning performed by household appliance separation. It is a block which shows the example of a structure of 1st Embodiment of the monitoring system to which this technique is applied. It is a figure explaining FHMM. It is a figure explaining the outline | summary of the formulation of household appliance separation by FHMM. It is a flowchart explaining the learning process (learning process) of FHMM according to the EM algorithm which a monitoring system performs. It is a flowchart explaining the process of E step which a monitoring system performs by step S13. It is a figure explaining the relationship between the forward probability α _{t, z} and the backward probability β _{t, z of the} FHMM and the forward probability α _{t, i} and the backward probability β _{t, j} of the HMM. It is a flowchart explaining the process of the M step which a monitoring system performs by step S14. It is a flowchart explaining the information presentation process which presents the information of household appliance #m which a monitoring system performs. It is a figure which shows the example of a display of the power consumption U ^(m) performed by an information presentation process. It is a block which shows the structural example of 2nd Embodiment of the monitoring system to which this technique is applied. It is a flowchart explaining the process of E step which a monitoring system performs by step S13. It is a flowchart explaining the process of the M step which a monitoring system performs by step S14. It is a block which shows the structural example of 3rd Embodiment of the monitoring system to which this technique is applied. It is a figure explaining the method of calculating | requiring the forward probability (alpha) _{t, p} by applying a particle filter to the state combination z of FHMM. It is a figure explaining the method of calculating | requiring backward probability (beta) _{t, p} by applying a particle filter to the state combination z of FHMM. It is a figure explaining the method of calculating a posteriori probability (gamma) _{t, p} by applying a particle filter to the state combination z of FHMM. It is a flowchart explaining the process of E step which a monitoring system performs by step S13. It is a flowchart explaining the process of E step which a monitoring system performs by step S13. It is a block which shows the structural example of 4th Embodiment of the monitoring system to which this technique is applied. It is a flowchart explaining the process of M step of step S14 which a monitoring system imposes load restrictions. It is a figure explaining load restrictions. It is a figure explaining a base waveform restriction | limiting. It is a flowchart explaining the process of M step of step S14 which a monitoring system imposes base waveform restrictions. It is a block which shows the structural example of 5th Embodiment of the monitoring system to which this technique is applied. It is a figure explaining the outline | summary of speaker separation by the monitoring system which learns FHMM. It is a block which shows the structural example of 6th Embodiment of the monitoring system to which this technique is applied. It is a flowchart explaining the process (learning process) of the model learning which a monitoring system performs. And FIG. 18 is a block diagram illustrating a configuration example of an embodiment of a computer to which the present technology is applied.

[Outline of this technology]

  FIG. 1 is a diagram illustrating an overview of an embodiment of a monitoring system to which a data processing device of the present technology is applied.

  In each household, electric power provided by an electric power company is drawn into a switchboard (distribution panel), and is supplied from the switchboard to electrical equipment such as home appliances (connected to an outlet or the like) in the home.

  The monitoring system to which this technology is applied measures the sum of the current consumed by one or more home appliances in the home at one location such as the root of the switchboard that supplies power to the home. The household appliance separation which isolate | separates the electric power (electric current) consumed by each household appliances, such as an air conditioner and a vacuum cleaner in a household, is performed from the series (current waveform series) of the sum of currents.

  As input data to be input to the monitoring system, total data relating to the sum of currents consumed by each home appliance can be adopted in addition to the sum of currents consumed by each home appliance itself.

  As the summation data, the summation of values for which addition is established can be employed. Specifically, as the total data, in addition to the total current consumed by each home appliance itself, for example, the sum of the power consumed by each home appliance and the waveform of the current consumed by each home appliance The sum total of frequency components obtained by (Fast Fourier Transform) or the like can be adopted.

  Moreover, in household appliance isolation | separation, the information regarding the electric current consumed by each household appliance other than the electric power consumed by each household appliance can be isolate | separated from total data. Specifically, in home appliance separation, for example, current consumed by individual home appliances and frequency components thereof can be separated from the total data.

  In the following description, for example, the sum of current consumed by each home appliance is adopted as the sum data, and in the home appliance separation, for example, from the waveform of the sum of current as sum data, consumed by each home appliance. The current waveform is separated.

  FIG. 2 is a diagram for explaining an outline of waveform separation learning performed in home appliance separation.

As the waveform separation learning, the current waveform Y _t as the sum data for each time t is the sum of the current waveform W ^(m) of the current being consumed by each home #m (sum), the current waveform Y _t from the current waveform W being consumed by the individual consumer electronics #m ^(m) is obtained.

  In FIG. 2, there are five home appliances # 1 to # 5 in the home, and among the five home appliances # 1 to # 5, home appliances # 1, # 2, # 4, and # 5 are on. Is in a state (a state in which power is consumed), and home appliance # 3 is in an off state (a state in which power is not consumed).

Therefore, in FIG. 2, the current waveform Y _t as the sum data is the current consumption W ⁽¹⁾ , W ⁽²⁾ , W ⁽⁴⁾ , and # 5 of the home appliances # 1, # 2, # 4, and # 5, respectively. , W ^{(5) is} the sum (sum).

[First embodiment of a monitoring system to which the present technology is applied]

  FIG. 3 is a block diagram illustrating a configuration example of the first embodiment of the monitoring system to which the present technology is applied.

  In FIG. 3, the monitoring system includes a data acquisition unit 11, a state estimation unit 12, a model storage unit 13, a model learning unit 14, a label acquisition unit 15, and a data output unit 16.

  The data acquisition unit 11 acquires a time series (current time series) of the current waveform Y as total data and a time series (voltage time series) of the voltage waveform (voltage waveform) V corresponding to the current waveform Y. The state estimation unit 12, the model learning unit 14, and the data output unit 16 are supplied.

  That is, the data acquisition part 11 is comprised with the measuring device (sensor) which measures an electric current and a voltage, for example.

  The data acquisition unit 11 measures, for example, a current waveform Y as a sum of currents consumed by home appliances in a home where a monitoring system is installed on a switchboard or the like, and measures a corresponding voltage waveform V. The state estimation unit 12, the model learning unit 14, and the data output unit 16 are supplied.

  The state estimation unit 12 includes a current waveform Y from the data acquisition unit 11 and an overall model (model parameter) φ that is stored in the model storage unit 13 and is a model of the entire home appliance in the home where the monitoring system is installed. Is used to estimate the state of operation of each home appliance. Then, the state estimation unit 12 supplies the operating state Γ of each home appliance as the estimation result of the state estimation to the model learning unit 14, the label acquisition unit 15, and the data output unit 16.

  That is, in FIG. 3, the state estimation unit 12 includes an evaluation unit 21 and an estimation unit 22.

  The evaluation unit 21 is supplied from the data acquisition unit 11 (to the state estimation unit 12) in each combination of states of the plurality of home appliance models # 1 to #M constituting the overall model φ stored in the model storage unit 13. An evaluation value E that evaluates the degree to which the current waveform Y is observed is supplied to the estimation unit 22.

  The estimation unit 22 uses the evaluation value E supplied from the evaluation unit 21 and is in each of the plurality of home appliance models # 1 to #M constituting the overall model φ stored in the model storage unit 13, that is, the home appliance model The operating state Γ of the home appliance represented by #m (the home appliance modeled by the home appliance model #m) is estimated and supplied to the model learning unit 14, the label acquisition unit 15, and the data output unit 16.

  The model storage unit 13 stores an overall model (model parameter) φ that is a model of a plurality of home appliances as a whole.

  The overall model φ is composed of a plurality of M home appliance models # 1 to #M which are models of home appliances (representing current consumption).

  The overall model parameter φ includes a current waveform parameter that represents current consumption for each operating state of the home appliance represented by home appliance model #m.

  The overall model parameter φ includes, for example, a state variation parameter that represents the transition (variation) of the operating state of the home appliance represented by the home appliance model #m, and an initial state that represents the initial state of the home appliance operating state represented by the home appliance model #m. The parameter and a dispersion parameter related to the dispersion of the observed value of the current waveform Y observed (generated) in the overall model can be included.

  The model parameters φ of the entire model stored in the model storage unit 13 are referred to by the evaluation unit 21 and the estimation unit 22 of the state estimation unit 12, the label acquisition unit 15, and the data output unit 16. Updated by the waveform separation learning unit 31, the distributed learning unit 32, and the state variation learning unit 33.

  The model learning unit 14 uses the current waveform Y supplied from the data acquisition unit 11 and the state estimation estimation result (operating state of each home appliance) Γ supplied from the state estimation unit 12 (the estimation unit 22 thereof). Then, model learning for updating the model parameter φ of the entire model stored in the model storage unit 13 is performed.

  That is, in FIG. 3, the model learning unit 14 includes a waveform separation learning unit 31, a distributed learning unit 32, and a state variation learning unit 33.

  The waveform separation learning unit 31 uses the current waveform Y supplied from the data acquisition unit 11 (to the model learning unit 14) and the operating state Γ of each home appliance supplied from the state estimation unit 12 (the estimation unit 22 thereof). Then, the waveform separation learning for obtaining (updating) the current waveform parameter as the model parameter φ is performed, and the current waveform parameter stored in the model storage unit 13 is updated with the current waveform parameter obtained by the waveform separation learning.

  The distributed learning unit 32 uses the current waveform Y supplied from the data acquisition unit 11 (to the model learning unit 14) and the operating state Γ of each home appliance supplied from the state estimation unit 12 (the estimation unit 22 thereof). Then, dispersion learning for obtaining (updating) the dispersion parameter as the model parameter φ is performed, and the dispersion parameter stored in the model storage unit 13 is updated with the dispersion parameter obtained by the dispersion learning.

  The state variation learning unit 33 obtains (updates) an initial state parameter and a state variation parameter as the model parameter φ using the operating state Γ of each home appliance supplied from the state estimating unit 12 (estimating unit 22 thereof). ) State variation learning is performed, and the initial state parameter and the state variation parameter stored in the model storage unit 13 are respectively updated with the initial state parameter and the state variation parameter obtained by the state variation learning.

The label acquisition unit 15 is obtained from the operating state Γ of each home appliance supplied from the state estimation unit 12 (the estimation unit 22), the overall model φ stored in the model storage unit 13, and the data output unit 16. Using the power consumption U ^{(m) of the} home appliance represented by the home appliance model #m as necessary, the home appliance label L ^(m) for identifying the home appliance represented by each home appliance model #m is acquired, and the data output unit 16 is supplied.

The data output unit 16 includes the voltage waveform V supplied from the data acquisition unit 11, the operating state Γφ of each home appliance supplied from the state estimation unit 12 (the estimation unit 22), and the whole stored in the model storage unit 13. Using the model, the power consumption U ^(m) of the home appliance represented by each home appliance model #m is obtained and displayed on the display ^(not shown ⁾ to the user.

In the data output unit 16, the power consumption U ^(m) of the home appliance represented by each home appliance model #m can be presented to the user together with the home appliance label L ^(m) supplied from the label acquisition unit 15.

  In the monitoring system configured as described above, for example, an FHMM (Factorial Hidden Markov Model) can be adopted as the overall model stored in the model storage unit 13.

  FIG. 4 is a diagram for explaining the FHMM.

  That is, FIG. 4A shows a graphical model of a normal HMM, and FIG. 4B shows a graphical model of an FHMM.

In a typical HMM, at time t, in one state S _t being in that time t, one observed value Y _t is observed.

On the other hand, in FHMM, at time t, one observation value Y _t is observed in a combination of a plurality of states S ⁽¹⁾ _t , S ⁽²⁾ _t , ..., S ^(M) _t at that time t. Is done.

  FHMM is a probability generation model proposed by Zoubin Ghahramani et al. For details, see, for example, Zoubin Ghahramani, and Michael I. Jordan, Factorial Hidden Markov Models', Machine Learning Volume 29, Issue 2-3, Nov./ Dec. 1997 (hereinafter also referred to as Document A).

  FIG. 5 is a diagram for explaining the outline of the formulation of home appliance separation by FHMM.

  Here, the FHMM has a configuration having a plurality of HMMs. Each HMM included in the FHMM is also referred to as a factor, and the m-th factor is also described as factor #m.

In the FHMM, a combination of a plurality of states S ⁽¹⁾ _t to S ^(M) _t at time t is a combination of the states of factor #m at time t (factor # 1 and factor # 2 State, ..., set of factor #M).

  FIG. 5 shows an FHMM having a factor number M of 3.

  In household appliance separation, for example, one factor corresponds to one household appliance (one factor corresponds to one household appliance). In FIG. 5, factor #m corresponds to home appliance #m.

  In the FHMM, the number of states constituting the factor is arbitrary for each factor, but in FIG. 5, the number of states of the three factors # 1, # 2, and # 3 are all four. .

  In FIG. 5, at time t = t0, the factor # 1 is in the state # 14 (indicated by a thick circle) among the four states # 11, # 12, # 13, and # 14. ), Factor # 2 is state # 21 (indicated by a bold circle) among the four states # 21, # 22, # 23, and # 24. At time t = t0, factor # 3 is state # 33 (indicated by a thick circle) among the four states # 31, # 32, # 33, and # 34.

  In home appliance separation, the state of factor #m corresponds to the operating state of home appliance #m to which factor #m corresponds.

  For example, in factor # 1 corresponding to home appliance # 1, state # 11 corresponds to the off state of home appliance # 1, and state # 14 corresponds to the so-called normal mode on state of home appliance # 1. Also, for example, in factor # 1 corresponding to home appliance # 1, state # 12 corresponds to a so-called sleep mode on state of home appliance # 1, and state # 13 represents a so-called energy saving mode on of home appliance # 1. Corresponds to the state.

In the FHMM, in the state #mi of the factor #m, a unique waveform W ^(m) _#mi, which is a unique waveform for each state of each factor, is observed (generated).

In FIG. 5, the characteristic waveform W ⁽¹⁾ _{# 14} is observed in the state # 14 at the time t = t0 for the factor # 1, and the natural waveform in the state # 21 at the time t = t0 for the factor # 2. W ⁽²⁾ _{# 21} is observed. Further, with the factor # 3, the characteristic waveform W ⁽³⁾ _{# 33} is observed in the state # 33 at time t = t0.

  In the FHMM, a combined waveform obtained by synthesizing the eigen waveform observed in each factor is generated as an observed value observed in the FHMM.

  Here, as the synthesis of the characteristic waveforms, for example, the total sum (addition) of the characteristic waveforms can be employed. In addition, for the synthesis of eigen waveforms, for example, weighted addition of eigen waveforms or logical sum of eigen waveforms (when eigen waveform values are 0 and 1) can be used. The sum of is adopted.

In the FHMM learning, the current waveform as the sum data at each time t = ..., t0, t1,..., Y _t0 , Y _{t0 + 1} ,. In addition, FHMM model parameters are obtained (updated).

  When the above FHMM is adopted as the overall model φ stored in the model storage unit 13 (FIG. 3), the home appliance model #m constituting the overall model φ corresponds to a factor #m.

  Note that as the factor M of the FHMM, a value larger by a predetermined number as a margin than the maximum number of home appliances assumed to be present in the home is adopted.

  In addition, as the FHMM as the overall model φ, an FHMM having a state where each factor is 3 or more is employed.

  When the factor has only two states, the home appliance operating state corresponding to the factor can express only two states, for example, an off state and an on state. This is because it is difficult to obtain accurate power consumption for home appliances such as air conditioners (hereinafter also referred to as fluctuating load home appliances) that fluctuate.

  That is, by adopting an FHMM having a factor of 3 or more as the FHMM as the overall model φ, it is possible to obtain accurate power consumption and the like for the variable load home appliance.

Overall model phi, when adopting the FHMM, simultaneous distribution of the sequence of the current waveform Y _t observed in FHMM, the combination S _t a sequence of states S ^(m) _t of each factor #m P ({S _t , Y _t }) is calculated by equation (1) by assuming Markov property.

・・・（１）

... (1)

Here, the joint distribution _{P ({S t, Y t} }) at time t, at each factor state S ^(m) (combination of the M factors for each state) a combination of _t of #m S _t, the current This represents the probability that the waveform Y _t is observed.

P (S ₁ ) represents the initial state probability of being in the combination S ₁ of the state S ^(m) ₁ of each factor #m at the first time t = 1.

_{P (S t | S t-} 1) is the time t-1, the state of the combination S _t-1 Niite, at time t, represents the transition probability of transition to a combination of state S _t.

P (Y _t | S _t) is the time t, the combination S _t state represents the observation probability of the current waveform Y _t is observed.

State combinations S _t of at time t, the M pieces of factor # 1 #M in each state S ⁽¹⁾ at time ^{_{t t, S (2) t}} , ···, a combination of S ^(M) _{t Yes} , it is expressed by the equation S _t = {S ⁽¹⁾ _t , S ⁽²⁾ _t ,..., S ^(M) _t }.

Note that the operating state of home appliance #m is assumed to fluctuate independently from other home appliances #m ′, and the state S ^(m) _t of factor #m is the state S ^{(m ′) of} other factor #m ′. Transitions are independent of _t .

In addition, as the number K ^(m) of the HMM states as the factor #m of the FHMM, a number independent of the number K ^{(m ′)} of the HMM states as the other factor #m ′ can be adopted. . However, here, for simplicity of explanation, the numbers K ⁽¹⁾ to K ^(M) of the states of factor # 1 to #M are expressed by the equation K ⁽¹⁾ = K ( ² ) = ... = K ^{( M) Assuming that} = K, the same number K is assumed.

In FHMM, initial state probability P (S ₁ ), transition probability P (S _t | S _t-1 ), and observation necessary for calculating simultaneous distribution P ({S _t , Y _t }) of equation (1) The probability P (Y _t | S _t ) can be calculated as follows.

That is, the initial state probability P (S ₁ ) can be calculated according to the equation (2).

・・・（２）

... (2)

Here, P (S ^(m) ₁ ) represents the initial state probability that the state S ^(m) ₁ of the factor #m is the first state (initial state) at time t = 1.

The initial state probability P (S ^(m) ₁ ) is, for example, K with the initial state probability of the k-th (k = 1, 2,..., K) state of factor #m as the k-th row component. A column vector of rows (a matrix of K rows and 1 column).

The transition probability P (S _t | S _t−1 ) can be calculated according to equation (3).

・・・（３）

... (3)

_{^{Here, P (S (m) t}} | S (m) t-1) , in factor #m, the time t-1 to the state S ^(m) _t-1 Niite, time t to the state S ^(m) _This represents the transition probability of transition to _t .

The transition probability P (S ^(m) _t | S ^(m) _t−1 ) is, for example, k′th (k ′ = 1, 2,..., K from the kth state #k of the factor #m. ) In the K-th row and K-th column as a component (square matrix).

The observation probability P (Y _t | S _t ) can be calculated according to equation (4).

・・・（４）

... (4)

  Here, a dash (') represents transposition, and a superscript -1 represents an inverse (inverse matrix). | C | represents the absolute value (determinant) of C (determinant operation).

Also, D is represents the dimension of the observed value Y _t.

For example, the data acquisition unit 11 in FIG. 3 sets a timing at which the zero crossing occurs when the voltage changes from a negative value to a positive value as a timing at which the current phase is 0 at a predetermined sampling interval. One cycle of current (1/50 or 1/60 second in Japan) is sampled, and a column vector whose component is the sampled value is output as a current waveform Y _t for one time.

Assuming that the number of times that the data acquisition unit 11 samples current for one period is D times, the current waveform Y _t is a column vector of D rows.

According to the observation probability P (Y _t | S _t ) in Expression (4), the observed value Y _t follows a normal distribution with an average value (average vector) of μ _t and a variance (covariance matrix) of C.

Mean value mu _t is, for example, a column vector D line similar to the current waveform Y _t, dispersion C, for example, a D line D columns of the matrix (a matrix of diagonal components dispersed).

Mean value mu _t uses a unique waveform W ^(m) described in FIG. 5, represented by the formula (5).

・・・（５）

... (5)

Here, if the eigen waveform of state #k with factor #m is expressed as W ^(m) _k , eigen waveform W ^(m) _k of state #k with factor #m is, for example, current waveform Y _t and It is a column vector of similar D rows.

The characteristic waveform W ^(m) is the characteristic waveform W ^(m) ₁ , W ^(m) ₂ , ..., W ^{(m )} A matrix of D rows and K columns having a column vector that is a collection of _K and a natural waveform W ^(m) _k of state #k with factor #m as a component of the kth column.

Further, S ^{* (m)} _t represents the state of factor #m at time t, and hereinafter, S ^{* (m)} _t is also referred to as the current state of factor #m at time t. Current state S ^{* (m)} _t factors #m of time t, for example, as shown in equation (6), in component 0 of only one row of the K line, the other line component is 0 It is a column vector of K rows.

・・・（６）

... (6)

If the state of factor #m at time t is state #k, the column vector S ^{* (m)} _t of K rows as the current state S ^{* (m)} _t of factor #m at time _t is k rows Only the eye component is set to 1, and the other components are set to 0.

According to the equation (5), the total sum of the characteristic waveforms W ^(m) _k of the state #k of each factor #m at time _t is obtained as the average value μ _t of the current waveform Y _t at time t.

The model parameters φ of the FHMM are the initial state probability P (S ^(m) ₁ ) in equation (2), the transition probability P (S ^(m) _t | S ^(m) _t-1 ) in equation ( ₃ ), dispersion C 4), and a formula (5 unique waveform W ^{in) (m) (= W (} m) 1, W (m) 2, ···, W (m) K), the model of Figure 3 The learning unit 14 obtains model parameters φ of these FHMMs.

That is, the waveform separation learning unit 31 obtains the unique waveform W ^(m) as a current waveform parameter by waveform separation learning. In the distributed learning unit 32, the distributed C is obtained as a distributed parameter by distributed learning. The state variation learning unit 33 uses the state variation learning to obtain the initial state probability P (S ^(m) ₁ ) and the transition probability P (S ^(m) _t | S ^(m) _t−1 ), respectively, as the initial state. Parameters and state variation parameters.

Here, as described above, for example, even when the operating state of each home appliance is two states of ON and OFF, when the operating state (combination) of 20 home appliances is expressed by a normal HMM The number of HMM states is 2 ²⁰ = 1,048,576, and the number of transition probabilities is 1,099,511,627,776 of the square.

  On the other hand, according to FHMM, M home appliances that have only two states of operation on and off can be represented by M factors each having two states. Therefore, in each factor, the number of states is two, and the number of transition probabilities is four of the squares. Therefore, the operating state of M = 20 home appliances (factors) is expressed by FHMM. In this case, the number of FHMM states (total number) may be as small as 40 = 2 × 20, and the number of transition probabilities may be as small as 80 = 4 × 20.

FHMM learning, that is, initial state probability P (S ^(m) ₁ ), transition probability P (S ^(m) _t | S ^(m) _t-1 ), variance C, and eigenvalue as model parameter φ of FHMM The update of the waveform W ^(m) can be performed according to an EM (Expectation-Maximization) algorithm as described in Document A, for example.

In FHMM learning using the EM algorithm, the processing of the E step and the processing of the M step are alternated in order to maximize the expected value Q (φ ^new | φ) of the conditional complete data log likelihood of Equation (7). Repeated.

・・・（７）

... (7)

Here, the expected value Q (φ ^new | φ) of the conditional complete data log likelihood is a ^new model parameter when the complete data {S _t , Y _t } is observed under the model parameter φ. Under φ ^new , it means the expected value of log likelihood log (P ({S _t , Y _t } | φ ^new )) at which complete data {S _t , Y _t } is observed.

In the processing of the E step of the EM algorithm, the expected value Q (φ ^new | φ) of the conditional complete data log likelihood of the equation (7) is obtained (the value corresponding to), and in the processing of the M step of the EM algorithm, A new model parameter φ ^new is obtained that increases the expected value Q (φ ^new | φ) obtained in the processing of the E step, and the model parameter φ is calculated from (expected value Q (φ ^new | φ) To a new model parameter φ ^new .

  FIG. 6 is a flowchart for explaining FHMM learning processing (learning processing) according to the EM algorithm performed by the monitoring system (FIG. 3).

In step S <b> 11, the model learning unit 14 has an initial state probability P (S ^(m) ₁ ) and a transition probability P (S ^(m) _t | S ^{(m )} as model parameters φ of the FHMM stored in the model storage unit 13. ⁾ _t-1 ), variance C, and natural waveform W ^(m) are initialized, and the process proceeds to step S12.

Here, the k-th row component of the column vector of K rows as the initial state probability P (S ^(m) ₁ ), that is, the k-th initial state probability π ^(m) _k of the factor #m is, for example, 1 Initialized to / K.

The i-th row and j-th column components (i, j = 1,2,..., K) of the K-row and K-column matrix as the transition probability P (S ^(m) _t | S ^(m) _t−1 ), That is, in the factor #m, the transition probability P ^(m) _{i, j} for transition from the i-th state #i to the j-th state #j is expressed by the expression P ^(m) _{i, 1} + P using a random number, for example. ^(m) _{i, 2} +... + P ^(m) Initialized to satisfy _{i, K} = 1.

  The matrix of D rows and D columns as the variance C is initialized to a diagonal matrix of D rows and D columns in which diagonal components are set using, for example, random numbers and other components (components) are set to 0.

The column vector of the kth column of the matrix of D rows and K columns as the eigen waveform W ^(m) , that is, each component of the column vector of the D row as the eigen waveform W ^(m) _k of the state #k with factor #m For example, initialization is performed using a random number.

In step S12, the data acquisition unit 11 acquires a current waveform for a predetermined time T, and the current waveforms at each time t = 1, 2,..., T are measured waveforms Y ₁ , Y ₂ ,. , Y _T are supplied to the state estimation unit 12 and the model learning unit 14, and the process proceeds to step S13.

  Here, the data acquisition unit 11 acquires a voltage waveform as well as a current waveform at time t = 1, 2,... The data acquisition unit 11 supplies voltage waveforms at times t = 1, 2,..., T to the data output unit 16.

  In the data output unit 16, the voltage waveform from the data acquisition unit 11 is used for calculation of power consumption.

In step S13, the state estimation unit 12 performs an E step process using the measured waveforms Y ₁ to Y _T from the data acquisition unit 11, and the process proceeds to step S14.

That is, in step S13, the state estimation unit 12 uses the measured waveforms Y ₁ to Y _T from the data acquisition unit 11 to state probability of being in each state of each factor #m of the FHMM stored in the model storage unit 13. Etc., and the estimation result of the state estimation is supplied to the model learning unit 14 and the data output unit 16.

  Here, as described with reference to FIG. 5, in the home appliance separation, the state of factor #m corresponds to the operating state of home appliance #m to which factor #m corresponds. Therefore, the state probability of being in state #k with factor #m of FHMM represents the operating state of home appliance #m to some extent with state #k, so state estimation to determine such state probability requires the operating state of home appliances. (Estimated).

In step S14, the model learning unit 14 performs an M step process using the measurement waveforms Y ₁ to Y _T from the data acquisition unit 11 and the estimation result of the state estimation from the state estimation unit 12, and the process is as follows. The process proceeds to step S15.

That is, in step S <b> 14, the model learning unit 14 is stored in the model storage unit 13 using the measurement waveforms Y ₁ to Y _T from the data acquisition unit 11 and the estimation result of the state estimation from the state estimation unit 12. In addition, by learning an FHMM in which each factor has a state of 3 or more, an initial state probability π ^(m) _k and a transition probability P ^(m) _i as model parameters φ of the FHMM stored in the model storage unit 13 are _obtained . _{, j} , variance C, and natural waveform W ^(m) are updated.

  In step S15, the model learning unit 14 determines whether the convergence condition of the model parameter φ is satisfied.

Here, as the convergence condition of the model parameter φ, for example, the processing of the E step and the M step is repeated a predetermined number of times, or the measurement waveforms Y ₁ to Y _T are observed in the FHMM. It can be adopted that the amount of change in likelihood between before and after updating the model parameter φ is within a preset threshold.

  When it is determined in step S15 that the convergence condition of the model parameter φ is not satisfied, the process returns to step S13, and the same process is repeated thereafter.

  In step S15, when it is determined that the convergence condition of the model parameter φ is satisfied, the learning process ends.

  FIG. 7 is a flowchart for explaining the processing of step E performed by the monitoring system of FIG. 3 in step S13 of FIG.

In step S <b> 21, the evaluation unit 21 analyzes the variance C of the FHMM as the overall model φ stored in the model storage unit 13, the eigen waveform W ^(m) , and the measurement waveform Y _t = {from the data acquisition unit 11. _{Y 1, Y 2, ···,} Y using _T}, each time t = {1,2, ···, of the T}, for each combination S _t states, observation probability of formula (4) P ( Y _t | S _t ) is obtained as an evaluation value E, supplied to the estimation unit 22, and the process proceeds to step S22.

In step S22, the estimation unit 22 determines the observation probability P (Y _t | S _t ) from the evaluation unit 21 and the transition probability P ^(m) _i, FHMM as the overall model φ stored in the model storage unit 13 _{. j} (and initial state probability π ^(m) ) are used to observe the measured waveforms Y ₁ , Y ₂ ,..., Y _t , and at time t, the state combination (state of factor # 1 at time t) The combination of the state of factor # 2 and the combination of the state of factor #M) determines the forward probability α _{t, z in z} , and the process proceeds to step S23.

  Here, how to calculate the forward probability of HMM, for example, see page 336 of CM Bishop, “Statistical prediction based on Bayesian theory under pattern recognition and machine learning”, Springer Japan, 2008 (hereinafter also referred to as Reference B). Have been described.

Forward probability alpha _{t, z,} for example, recurrence formula _{α t, z = Σα t-} 1 using a one time before the forward probability _{α t-1, w, w} P (z | w) P (Y t | It can be determined according to z).

In the recurrence formula α _{t, z} = Σα _{t-1, w} P (z | w) P (Y _t | z), Σ represents a summation that changes w to all the combinations of states of FHMM. .

In the recurrence formula α _{t, z} = Σα _{t−1, w} P (z | w) P (Y _t | z), w represents a combination of states at time t−1 one time before. P (z | w) represents the transition probability of being in the state combination w at time t-1 and transitioning to the state combination z at time t, and P (Y _t | z) is the state at time t. Represents the observation probability of observing the measured waveform Y _t in the combination z.

The initial value of the forward probability α _{t, z} , that is, the forward probability α _{1, z at} time t = 1, is the initial state probability π of the state #k of each factor #m constituting the state combination z ^(m) The product of _k is adopted.

In step S 23, the estimation unit 22 determines the observation probability P (Y _t | S _t ) from the evaluation unit 21 and the transition probability P ^(m) _i, FHMM as the overall model φ stored in the model storage unit 13 _. with _j, at time t, the combination of the states z Niite then measured waveform Y _t, obtains a Y _{t + 1,} · · ·, backward probability observing Y _T beta _{t, z,} the process, step Proceed to S24.

  Here, how to obtain the backward probability of the HMM is described, for example, on page 336 of the above-mentioned document B.

The backward probability β _{t, z} is, for example, a recurrence formula β _{t, z} = ΣP (Y _t | z) P (w | z) β _{t + 1} using the backward probability β _{t + 1, w} after one time. _{, w} .

In the recursion formula β _{t, z} = ΣP (Y _t | z) P (w | z) β _{t + 1, w} , Σ represents a summation that takes w to be changed to all combinations of FHMM states .

In the recurrence formula β _{t, z} = ΣP (Y _t | z) P (w | z) β _{t + 1, w} , w represents a combination of states at time t + 1 one time later. P (w | z) represents the transition probability of being in state combination z at time t and transitioning to state combination w at time t + 1, and P (Y _t | z) is the state at time t. Represents the observation probability of observing the measured waveform Y _t in the combination z.

Note that 1 is adopted as the initial value of the backward probability β _{t, z} , that is, the backward probability β _{T, z at} time t = T.

In step S24, the estimation unit 22 uses the forward probability α _{t, z} and the backward probability β _{t, z} and is in the state combination z at time t in the FHMM as the overall model φ according to the equation (8). The posterior probability γ _{t, z} is obtained, and the process proceeds to step S25.

・・・（８）

... (8)

Here, Σ of the denominator of the right side of the equation (8) represents the summation of Nikki changing the w, all combinations S _t of possible states at time t.

According to equation (8), the posterior probability γ _{t, z} is obtained by calculating the product α _{t, z} β _{t, z} of the forward probability α _{t, z} and the backward probability β _{t, z} as the product α _{t, w} β _{t. ,} of _w, obtained by normalizing the sum Σα _{t, w} β _{t, w} of the combination W∈S _t of possible states is FHMM.

In step S25, the estimator 22 uses the posterior probability γ _{t, z} and, at the factor #m, at the time t, the posterior probability <S ^(m) _t > in the state S ^(m) _t and at the time t In factor #m, we are in state S ^(m) _t , and in other factor #n, we find the posterior probability <S ^(m) _t S ⁽ⁿ⁾ _t '> in state S ⁽ⁿ⁾ _t. The process proceeds to step S26.

Here, the posterior probability <S ^(m) _t > is obtained according to Equation (9).

・・・（９）

... (9)

According to Equation (9), the posterior probability <S ^(m) _t > in state S ^(m) _t at time t in factor #m is the posterior probability γ _{t in} state combination z at time _{t. , z} is obtained by marginalizing a state combination z not including the state of factor #m.

The posterior probability <S ^(m) _t > is, for example, the state probability (posterior probability) of the kth state among the K states of factor #m at time t as the component of the kth row. Let it be a column vector of K rows.

The posterior probability <S ^(m) _t S ⁽ⁿ⁾ _t ′> is obtained according to equation (10).

・・・（１０）

... (10)

According to equation (10), at time t, the factor #m, state S ^(m) _t Niite, in other factors #n, posterior probability being in state _{^{S (n) t <S (}} m) t S ⁽ⁿ⁾ _t ′> is the posterior probability γ _{t, z} that is in the state combination z at time t with respect to the state combination z that does not include both the state of factor #m and the state of factor #n. It is calculated by making it.

Note that the posterior probability <S ^(m) _t S ⁽ⁿ⁾ _t ′> is, for example, the state probability (at time t) in the state #k of factor #m and the state #k ′ of other factor #n ( The a posteriori probability) is a matrix of K rows and K columns with components of k rows and k 'columns.

In step S26, the estimation unit 22 calculates the forward probability α _{t, z} , the backward probability β _{t, z} , the transition probability P (z | w), and the observation probability P (Y _t | S _t ) from the evaluation unit 21. using, in the factor #m, at time t-1, state S ^(m) _t-1 Niite, the next time t, posterior probability being in state ^{_{S (m) t <S (}} m) t-1 S ^(m) _{Find t} '>.

Then, the estimation unit 22 calculates the posterior probabilities <S ^(m) _t >, <S ^(m) _t S ⁽ⁿ⁾ _t '>, and <S ^(m) _t-1 S ^(m) _t '>. The estimation result of the state estimation is supplied to the model learning unit 14, the label acquisition unit 15, and the data output unit 16, and the process returns from the E step.

Here, the posterior probability <S ^(m) _t−1 S ^(m) _t ′> is obtained according to Equation (11).

・・・（１１）

(11)

In calculating the posterior probability <S ^(m) _t-1 S ^(m) _t '> in equation (11), the transition probability P (z | w) of transition from the state combination w to the state combination z is expressed by the equation (11) 3) According to 3), the transition probability P ⁽¹⁾ _{i (1} ) from the state #i (1) of the factor # 1 constituting the state combination w to the state #j (1) of the factor # 1 constituting the state combination z _{), j (1)} , transition probability P ⁽²⁾ from state #i (2) of factor # 2 constituting state combination w to state #j (2) of factor # 2 constituting state combination z _{i (2), j (2} ), ···, and factors #M constituting the combination z from one state factor #M constituting the combination w state #i (M) state #j (M ) Transition probability P ^(M) _{i (M), j (M)} product P ⁽¹⁾ _{i (1), j (1)} × P ⁽²⁾ _{i (2), j (2)} × XP ^(M) _{Obtained as i (M), j (M)} .

The posterior probability <S ^(m) _t-1 S ^(m) _t '> is, for example, the state probability of being in state #i at time t-1 and in state j at the next time t in factor #m. It is assumed that (a posteriori probability) is a matrix of K rows and K columns with a component of i rows and j columns.

Figure 8 is a forward probability alpha _{t, z} of FHMM, and will be described with backward probability beta _{t, z,} forward probability of (normal) HMM alpha _{t, i,} and the relationship between the backward probability beta _{t, j} FIG.

  As for the FHMM, an HMM equivalent to the FHMM can be configured.

  An HMM equivalent to a certain FHMM has a state corresponding to a combination z of states of the factors of the FHMM.

Then, the forward probability α _{t, z} and the backward probability β _{t, z} of the FHMM coincide with the forward probability α _{t, i} and the backward probability β _{t, j} of the HMM equivalent to the FHMM.

  FIG. 8A shows an FHMM composed of factors # 1 and # 2 having two states # 1 and # 2.

  In the FHMM of FIG. 8A, the combination # of the factor # 1 state #k and the factor # 2 state #k ′ z = [k, k ′], the factor # 1 state # 1 and the factor # 2 state # 1 Combination [1,1], factor # 1 state # 1 and factor # 2 state # 2 combination [1,2], factor # 1 state # 2 and factor # 2 state # 1 There are four combinations, a combination [2,1] and a combination [2,2] of state # 2 of factor # 1 and state # 2 of factor # 2.

  FIG. 8B shows an HMM equivalent to the FHMM of FIG. 8A.

  The HMM in FIG. 8B has four states # (1), [1,2], [2,1], and [2,2] corresponding to the four combinations of the states of the FHMM in FIG. 8A, respectively. 1, 1), # (1, 2), # (2, 1), and # (2, 2).

Then, the forward probability α _{t, z} = {α _{t, [1,1]} , α _{t, [1,2]} , α _{t, [2,1]} , α _{t, [2,2]} of the FHMM in FIG. 8A } Is the forward probability α _{t, i} = {α _{t, (1,1)} , α _{t, (1,2)} , α _{t, (2,1)} , α _{t, (2,2 )} } Matches.

Similarly, the backward probability β _{t, z} = {β _{t, [1,1]} , β _{t, [1,2]} , β _{t, [2,1]} , β _{t, [2,2} of the FHMM in FIG. 8A _] } Is the forward probability β _{t, i} = {β _{t, (1,1)} , β _{t, (1,2)} , β _{t, (2,1)} , β _{t, (2, 2)} Matches}.

For example, the sum Σα _{t, w} β _{t, w} of the denominator of the right side of the above equation (8), that is, the combination w∈S _t of the states α _{T, w} β _{t, w} that the FHMM can take is For the FHMM of FIG. 8A, the formula Σα _{t, w} β _{t, w} = α _{t, [1,1]} β _{t, [1,1]} + α _{t, [1,2]} β _{t, [1,2]} + α _{t, [2,1]} β _{t, [2,1]} + α _{t, [2,2]} β _{t, [2,2]}

  FIG. 9 is a flowchart for explaining the processing of the M step performed in step S14 of FIG. 6 by the monitoring system of FIG.

In step S31, the waveform separation learning unit 31 and the measured waveform Y _t from the data acquisition unit 11 and the posterior probabilities <S ^(m) _t > and <S ^(m) _t S ^{(n )} with _t '>, by performing waveform separation learning, obtains the updated value W ^{(m) new new} unique waveform W ^(m), by the update value W ^{(m) new new,} stored in the model storage unit 13 The natural waveform W ^(m) is updated, and the process proceeds to step S32.

That is, the waveform separation learning unit 31 obtains an updated value W ^{(m) new} of the natural waveform W ^(m) by calculating Expression (12) as waveform separation learning.

・・・（１２）

(12)

Here, W ^new is a D-row and K-column matrix, and the updated value W ^{(m) new of the characteristic} waveform W ^(m) of factor #m is arranged from left to right in the order of factor (index) #m. It is a matrix of D rows and K x M columns. The column vector of the (m-1) K + k columns of the eigen waveform (updated value) W ^new that is a matrix of D rows and K × M columns is the eigen waveform W ^(m) _k ( Update value).

<S _t '> is a K × M column vector in which the posterior probability <S ^(m) _t >, which is a column vector of K rows, is transposed from top to bottom in order of factor #m A row vector of columns. The (m-1) K + k column component of the posterior probability <S _t '>, which is a row vector of K × M columns, has the state probability of being in state #k with factor #m at time t. .

<S _t S _t '> is a matrix of K rows and K columns, the posterior probabilities <S ^(m) _t S ⁽ⁿ⁾ _t '> are arranged in the order of factor #m from top to bottom, and factor # A matrix of K × M rows and K × M columns arranged in order from left to right in n order. The component of the (m-1) K + k row (n-1) K + k 'column of the posterior probability <S _t S _t '> which is a matrix of K × M rows and K × M columns is at time t. It is the state probability of being in state #k with factor #m and state #k ′ with other factor #n.

  The superscript asterisk (*) represents an inverse matrix or a pseudo inverse matrix.

According to the waveform separation learning for calculating Equation (12), the error between the measured waveform Y _t and the average value μ _t = ΣW ^(m) S ^{* (m)} _{t of} Equation (5) is as small as possible. The measurement waveform Y _t is separated into the natural waveform W ^(m) .

In step S < _b > 32, the variance learning unit 32 measures the measured waveform Y _t from the data acquisition unit 11, the posterior probability <S ^(m) _t > from the estimation unit 22, and the unique waveform stored in the model storage unit 13. By performing distributed learning using W ^(m) , the update value C ^new of the distribution C is obtained, the distribution C stored in the model storage unit 13 is updated, and the process proceeds to step S33.

That is, the distributed learning unit 32 obtains an update value C ^new of the distributed C by calculating Expression (13) as distributed learning.

・・・（１３）

(13)

In step S33, the state variation learning unit 33 uses the posterior probabilities <S ^(m) _t > and <S ^(m) _t−1 S ^(m) _t ′> from the estimation unit 22 to perform state variation learning. To obtain the updated value P ^(m) _{i, j} ^new of the transition probability P ^(m) _{i, j} and the updated value π ^{(m) new} of the initial state probability π ^(m) , and the updated value P ^(m) _{i, j} ^new and π ^{(m) new} are used to update the transition probability P ^(m) _{i, j} and the initial state probability π ^(m) stored in the model storage unit 13, and M Return from step processing.

That is, the state change learning unit 33 calculates the expressions (14) and (15) as the state change learning, thereby _obtaining the updated values P ^(m) _{i of the} transition probabilities P ^(m) _{i and j} , respectively. _{, j} ^new and the updated value π ^{(m) new} of the initial state probability π ^(m) are obtained.

・・・（１４）

(14)

・・・（１５）

(15)

Where <S ^(m) _{t-1, i} S ^(m) _{t, j} > is _{the i} of the posterior probability <S ^(m) _t-1 S ^(m) _t '> which is a matrix of K rows and K columns. This is a component in the row j column, and represents the state probability of being in the state #i at the time t-1 and in the state #j at the next time t in the factor #m.

<S ^(m) _{t-1, i} > is a component of the i _-th row of column vector posterior probability <S ^(m) _t-1 > of K row, and at time t-1, state # of factor #m i represents the state probability.

π ^(m) (π ^{(m) new} ) is K with the initial state probability π ^(m) _k (updated value π ^(m) _k ^new ) of state #k with factor #m as the component in the k-th row A column vector of rows.

  FIG. 10 is a flowchart for explaining an information presentation process for presenting information on home appliance #m, which is performed by the monitoring system (FIG. 3).

In step S41, the data output unit 16, the posterior of the estimation result of the state estimation from (voltage waveform corresponding to the measured waveform Y _t is a current waveform) V _t, the state estimation unit 12 the voltage waveform from the data acquisition unit 11 Using the probability <S ^(m) _t > and the eigen waveform W ^(m) stored in the model storage unit 13, the power consumption U ^(m) of each factor #m is obtained, and the process proceeds to step S42. .

Here, home appliance data output unit 16, by using the voltage waveform V _t at time t, the current consumption A _t of the corresponding home appliance #m on factors #m time t, corresponding to a factor #m time t Obtain the power consumption U _(m) of #m.

In the data output unit 16, the current consumption At of the home appliance #m corresponding to the factor #m at time _t is obtained as follows.

That is, for example, the data output unit 16 uses the characteristic waveform W ^(m) of the state #k with the maximum posterior probability <S ^(m) _t > for the factor #m as the home appliance # corresponding to the factor #m at the time t. calculated as the current consumption a _t of m.

Also, the data output unit 16 uses, for example, a factor that uses a state probability of each state of factor #m at time t as a weight, which is a component of a posteriori probability <S ^(m) _t > that is a column vector of K rows. Current consumption of home appliance #m corresponding to factor #m at time t with the weighted addition value of eigen waveform W ^(m) ₁ , W ^(m) ₂ ,..., W ^(m) _K in each state of #m _Calculate as At.

  If FHMM learning progresses and factor #m becomes a home appliance model that appropriately represents home appliance #m, the state probability of each state of factor #m at time t is the operating state of home appliance #m at time t The state probability of the state corresponding to is approximately 1, and the state probability of the remaining K-1 states is approximately 0.

As a result, in the factor #m, the characteristic waveform W ^(m) of the state #k having the maximum posterior probability <S ^(m) _t > and the state probability of each state of the factor #m at the time t are used as weights. The weighted addition values of the characteristic waveforms W ^(m) ₁ , W ^(m) ₂ ,..., W ^(m) _K in each state of factor #m are substantially the same.

In step S42, the label acquisition unit 15 acquires the home appliance label ^{m (m)} identifying the home appliance #m represented by each home appliance model #m, that is, the home appliance #m corresponding to each factor #m of the FHMM, and outputs the data. Then, the process proceeds to step S43.

Here, the label obtaining unit 15, for example, consumption and current A _t appliances #m corresponding to each factor #m obtained in the data output unit 16, from the power consumption U ^(m), the power consumption U ^(m) displays used hours of recognized consumer electronics #m, the and the current consumption a _t, power U ^(m), the name of the home appliances that fall within the use time period, that get input from the user, input by the user The name of the home appliance thus obtained can be acquired as a home appliance label L ^(m) .

In addition, in the label acquisition unit 15, for example, for various home appliances, a database in which attributes such as power consumption, current waveform (current consumption), usage time zone, and the like are registered in association with home appliance names is prepared in advance. ; then, in the database consumer electronics, consumer and current a _t appliances #m corresponding to each factor #m sought data output unit 16, the power consumption U ^(m), are recognized from the power U ^(m) The name of the home appliance associated with the usage time zone of #m can be acquired as the home appliance label L ^(m) .

Note that the label acquisition unit 15 has already acquired the home appliance label L ^{(m), and} for the factor #m corresponding to the home appliance #m supplied to the data output unit 16, the process of step S42 may be skipped. it can.

In step S43, the data output unit 16 presents the power consumption U ^(m) of each factor #m to the user by displaying it on a display ^(not shown ⁾ together with the home appliance label L ^{(m) of the} factor #m. Then, the information presentation process ends.

FIG. 11 is a diagram illustrating a display example of the power consumption U ^(m) performed in the information presentation process of FIG.

In the data output unit 16, for example, as shown in FIG. 11, the time series of the power consumption U ^(m) of the home appliance #m corresponding to each factor #m is converted into the home appliance label L ^{(m ) And} displayed on a display (not shown ⁾ .

  As described above, in the monitoring system, the FHMM that models the operating state of each home appliance is learned by the FHMM having a factor of 3 or more, so the power consumption (current) varies depending on the mode and settings. Accurate power consumption and the like can be obtained for variable load home appliances such as air conditioners.

  In addition, in the monitoring system, the power consumption of each household appliance in the home can be obtained by measuring the total current consumed by each household appliance at one location such as a switchboard. Compared with the case where a smart tap is attached, “visualization” of the power consumption of each home appliance in the home can be easily realized from both the cost and labor aspects.

  According to “visualization” of the power consumption of each home appliance in the home, for example, it is possible to raise awareness of power saving at home.

  In addition, the power consumption of each household electrical appliance obtained by the monitoring system is collected by, for example, a server, and the usage time zone of the household electrical appliance and the life pattern are estimated from the power consumption of each household electrical appliance. Can be used for marketing, etc.

[Second embodiment of the monitoring system to which the present technology is applied]

  FIG. 12 is a block diagram illustrating a configuration example of the second embodiment of the monitoring system to which the present technology is applied.

  In the figure, portions corresponding to those in FIG. 3 are denoted by the same reference numerals, and description thereof will be omitted below as appropriate.

  The monitoring system of FIG. 12 is common to the monitoring system of FIG. 3 in that the data acquisition unit 11 to the data output unit 16 are included.

  However, the monitoring system of FIG. 12 is different from the monitoring system of FIG. 3 in that an individual distributed learning unit 52 is provided instead of the distributed learning unit 32 of the model learning unit 14.

In the distributed learning unit 32 in FIG. 3, a variance C of 1 is obtained for the FHMM as the overall model in the distributed learning. In the individual distributed learning unit 52 in FIG. 12, for the FHMM as the overall model in the distributed learning, An individual variance (hereinafter, also referred to as individual variance) σ ^(m) is obtained for each factor #m or for each state #k of each factor #m.

Here, when the individual variance σ ^(m) is an individual variance for each state #k of each factor #m, the individual variance σ ^(m) is, for example, the variance σ ^{(m (m )} A row vector of K columns, where _k is a component of the kth column.

In the following, in order to simplify the description, when the individual variance σ ^(m) is an individual variance for each factor #m, all components σ of the row vector of K columns as the individual variance σ ^{(m) (m)} ₁ , σ ^(m) ₂ ,..., σ ^(m) _K is the same value common to factor #m (the same scalar value as the variance for factor #m or (covariance ) Matrix), the individual individual variance σ ^(m) for each factor #m is treated the same as the individual individual variance σ ^(m) for each state #k of each factor #m.

Individual variance sigma variance sigma ^(m) _k states #k factors #m is k-th column component row vector of K rows of a ^(m) is a scalar value as a dispersion, or, as a covariance matrix D-by-D matrix.

When the individual variance σ ^(m) is obtained in the individual variance learning unit 52, the evaluation unit 21 uses the individual variance σ ^(m) and replaces the equation (4) with the observation probability P according to the equation (16). (Y _t | S _t ) is obtained.

・・・（１６）

... (16)

Here, sigma _t of formula (16) is determined according to equation (17).

・・・（１７）

... (17)

S ^{* (m)} _t represents the state of factor #m at time t, as shown in equation (6), where only one of the K rows has 0 component and the other row has components It is a column vector of 0 K rows.

In the individual variance learning unit 52, the individual variance σ ^(m) is obtained as follows.

That is, the expected value <σ> of the variance σ of the FHMM as the overall model φ with respect to the measurement waveform Y _t that is actually measured (observed) is the current waveform as the measurement waveform Y _t as shown in Equation (18). Expected value of square error | Y _t -Y ^ _t | ² <| Y _t -Y ^ _t | ² > with current waveform as generated waveform Y ^ _t generated (observed) in FHMM as overall model φ be equivalent to.

・・・（１８）

... (18)

Expected value <σ> of variance σ is the sum σ of expected values σ ^(m) <S ^(m) _t > of factor #m of individual individual variance σ ^(m) for each state #k of factor #m ⁽¹⁾ <S ⁽¹⁾ _t > + σ ⁽²⁾ <S ⁽²⁾ _t > +... + Σ ^(M) <S ^(M) _t > is equivalent.

Also, the expected value <| Y _t -Y ^ _t | ² > of the squared error | Y _t -Y ^ _t | ² between the measured waveform Y _t and the generated waveform Y ^ _t is the measured waveform Y _t and the individual waveform W the sum of the expected value W ^(m) factor #m of <S ^(m) _t> of ^{(m) W (1) <} S (1) t> + W (2) <S (2) t> + ··・ Square error with + W ^(M) <S ^(M) _t > | Y _t- (W ⁽¹⁾ <S ⁽¹⁾ _t > + W ⁽²⁾ <S ⁽²⁾ _t > + ... + ^{W (M) <S (M} ) t>) | is equivalent to ^2.

  Therefore, equation (18) is equivalent to equation (19).

・・・（１９）

... (19)

Individual variance σ ^(m) = σ ^{(m) new} satisfying equation (19) can be obtained according to equation (20) by constrained quadratic programming with a constraint that individual variance σ ^(m) is not negative. it can.

・・・（２０）

... (20)

The individual variance learning unit 52 obtains an updated value σ ^{(m) new} of the individual variance σ ^(m ) according to the equation (20), and the individual variance stored in the model storage unit 13 by the updated value σ ^{(m) new} . Update σ ^(m) .

As described above, in the individual variance learning unit 52, when obtaining the individual variance σ ^(m) for each factor #m or for each state #k of each factor #m, compared to the case where the variance C of 1 is used. Since the FHMM as the overall model φ has an improved ability to express home appliances and the accuracy of state estimation in the estimation unit 22 is improved, more accurate power consumption and the like can be obtained.

In particular, more accurate power consumption can be obtained by using the individual variance σ ^(m) for a variable load home appliance such as a vacuum cleaner whose current consumption varies depending on the suction state.

In the monitoring system of FIG. 12, processing similar to that of the monitoring system of FIG. 3 (learning processing of FIG. 6 and information presentation processing of FIG. 10 ⁾ is performed except that individual variance σ ^(m) is used instead of variance C. Is done.

  FIG. 13 is a flowchart for explaining the process of step E performed by the monitoring system of FIG. 12 in step S13 of FIG.

In step S <b> 61, the evaluation unit 21 determines the individual variance σ ^(m) and eigen waveform W ^(m) of the FHMM as the overall model φ stored in the model storage unit 13, and the measurement waveform from the data acquisition unit 11. _{Y t = {Y 1, Y} 2, ···, Y T} with, each time t = {1,2, ···, T } , and for each combination S _t state, (formula (4) Instead, the observation probability P (Y _t | S _t ) is obtained according to the equations (16) and (17), supplied to the estimation unit 22, and the process proceeds to step S62.

  Thereafter, in steps S62 to S66, the same processes as those in steps S22 to S26 of FIG. 7 are performed.

  FIG. 14 is a flowchart for explaining the process of the M step performed in step S14 of FIG. 6 by the monitoring system of FIG.

In step S71, the waveform separation learning unit 31 obtains an update value W ^{(m) new} of the eigen waveform W ^(m) by waveform separation learning, as in step S31 of FIG. 9, and the update value W ^{(m) new.} Thus, the natural waveform W ^(m) stored in the model storage unit 13 is updated, and the process proceeds to step S72.

In step S < _b > 72, the individual variance learning unit 52 performs the measurement waveform Y _t from the data acquisition unit 11, the posterior probability <S ^(m) _t > from the estimation unit 22, and the characteristic stored in the model storage unit 13. Using the waveform W ^(m) , an updated value σ ^{(m) new} of the individual variance σ ^(m) is obtained according to Equation (20) as variance learning.

Then, the individual variance learning unit 52 updates the individual variance σ ^(m) stored in the model storage unit 13 with the updated value σ ^{(m) new} of the individual variance σ ^(m) , and the processing starts from step S72. Proceed to step S73.

In step S73, the state variation learning unit 33 performs the state variation learning by using the state variation learning, as in step S33 of FIG. 9, and the updated value P ^(m) _{i, j} ^{new of} the transition probability P ^(m) _{i, j} and the initial state. An updated value π ^{(m) new} of the probability π ^(m) is obtained, and the transition probability P ^{(m )} stored in the model storage unit 13 by the updated values P ^(m) _{i, j} ^new and π ^{(m) new} is obtained. ⁾ _{i, j} and the initial state probability π ^(m) are updated, and the process returns from the M step process.

[Third embodiment of a monitoring system to which the present technology is applied]

  FIG. 15 is a block diagram illustrating a configuration example of the third embodiment of the monitoring system to which the present technology is applied.

  In the figure, portions corresponding to those in FIG. 3 are denoted by the same reference numerals, and description thereof will be omitted below as appropriate.

  The monitoring system in FIG. 15 is common to the monitoring system in FIG. 3 in that the data acquisition unit 11 to the data output unit 16 are included.

  However, the monitoring system of FIG. 15 is different from the monitoring system of FIG. 3 in that an approximate estimation unit 42 is provided instead of the estimation unit 22 of the state estimation unit 12.

The approximate estimation unit 42 obtains posterior probabilities (state probabilities) <S ^(m) _t > and the like under a state transition restriction that restricts the number of factors that cause a state transition at one time.

Here, for example, the posterior probability <S ^(m) _t > is obtained by peripheralizing the posterior probability γ _{t, z} according to the equation (9).

The posterior probability γ _{t, z} is obtained for all combinations z of the states of each factor of the FHMM using the forward probability α _{t, z} and the backward probability β _{t, z} according to equation (8).

Incidentally, the number of combinations z status of each factor FHMM is present only K ^M Street, with an increase in the number M of factors, increasing the order of the exponent.

Therefore, when the FHMM has a certain number of factors, the forward probability α _{t, z} and the backward probability β _{t, z} and further the posterior probability γ _{t, z for} all the state combinations z of each factor. Is strictly calculated as described in Document A, the calculation amount is enormous.

  Thus, for estimation of the state of the FHMM, for example, an approximation estimation method such as Gibbs Sampling, Completely Factorized Variational Approximation, or Structured Variational Approximation is proposed in Document A. However, in these approximate estimation methods, the amount of calculation is still large, and the accuracy may be significantly reduced due to the approximation.

  By the way, the possibility that the operating states of all household appliances in the home change (fluctuate) at each time is extremely low except for an abnormal case such as a power failure.

Therefore, the approximate estimation unit 42 obtains the posterior probability <S ^(m) _t > and the like under a state transition restriction that restricts the number of factors that cause a state transition at one time.

  As the state transition restriction, it is possible to adopt limiting the number of factors that cause a state transition at one time to one or several or less.

According to the state transition restriction, the posterior probability γ _{t, z} used for obtaining the posterior probability <S ^(m) _t >, etc., and hence the forward probability α _{t, z} and the backward probability β _{t, z} are strictly Since the number of state combinations z that need to be calculated in is greatly reduced, and the reduced state combinations are very unlikely to occur, posterior probabilities (state probabilities) <S ^{( m) The} amount of calculation can be greatly reduced without significantly degrading the accuracy such as _t >.

As a method of obtaining the posterior probability <S ^(m) _t > and the like under the state transition restriction that restricts the number of factors that cause state transition at one time in the approximate estimation unit 42, the particle filter is used in the processing of the E step. There is a method of applying a particle filter to a combination z of FHMM states.

  Here, the particle filter is described, for example, on page 364 of the above-mentioned document B.

  In the present embodiment, as the state transition restriction, it is assumed that the number of factors causing a state transition at one time is limited to, for example, one or less.

The combination z states FHMM, when applying the particle filter, p-th particle S _p of the particle filter, to represent some combination of state of the FHMM.

Here, the particle S _p, the particles S _p is expressed, factor # 1 in the state S ⁽¹⁾ _p, factor # 2 state S ⁽²⁾ _p, · · ·, the state of factor #M S ^{(M )} _p , which is expressed by the formula S _p = {S ⁽¹⁾ _p , S ⁽²⁾ _p ,..., S ^(M) _p }.

Further, the combination z states FHMM, when applying the particle filter, in combination S _t states, observation probability P of formula measurement waveform Y _t is observed (4) | instead of (Y _t S _t) , The observation probability P (Y _t | S _p ) at which the measurement waveform Y _t is observed is used in the particle S _p (a combination of states represented by).

The observation probability P (Y _t | S _p ) can be calculated, for example, according to the equation (21).

・・・（２１）

(21)

Wherein the observation probability P _{(21) (Y t | S} p) is the mean value (mean vector) is the mu _t rather, mu is a point _p, the observation probability P (Y _t of formula (4) | S _t ) is different.

I mean mu _p, as well as the average value mu _t, a column vector of D rows similar to the current waveform Y _t, using a unique waveform W ^(m), the formula (22).

・・・（２２）

(22)

Here, S ^{* (m)} _p is of the states constituting the combination of the states of the particles S _p, represents the state of the factor #m, following the state of factor #m particles S _p S ^{* (m )} Also called _p . State S ^{* (m)} _p factors #m particles S _p, for example, as shown in equation (23), with component 0 of only one row of the K line, the other line component is 0 It is a column vector of K rows.

・・・（２３）

(23)

If the state of factor #m that constitutes the combination of states as particles S _p is state #k, column vector S ^{* (} K rows as factors S ^{* (m)} _p of factor #m of particles S _p ^{m) For} _p , only the component in the k-th row is set to 1, and the other components are set to 0.

According to equation (22), the particles S sum of unique waveform W ^(m) _k states #k of each factor #m constituting the combination of the states of the _p has an average value of the current waveform Y _t at time t mu _Calculated as _p .

FIG. 16 is a diagram for explaining a method of obtaining the forward probability α _{t, p} by applying a particle filter to the combination z of FHMM states.

  In FIG. 16 (the same applies to FIGS. 17 and 18 to be described later), for example, the number of FHMM factors is four, and each factor represents two states representing an on state and an off state as operating states. I will have it.

The evaluation unit 21 includes the variance C of the FHMM as the overall model φ stored in the model storage unit 13, the natural waveform W ^(m) , and the measured waveform Y _t = {Y ₁ , Y from the data acquisition unit 11. ₂ ,..., Y _T } and each particle S _p = {S ₁ , S ₂ ,..., S _R } at each time t = {1, 2,. For the combination of states, the observation probability P (Y _t | S _p ) of Expression (21) is obtained and supplied to the approximate estimation unit 42.

Then, the approximate estimator 42 has the observation probability P (Y _t | S _p ) from the evaluation unit 21 and the transition probability P ^(m) _{i, j} of the FHMM as the overall model φ stored in the model storage unit 13. (And initial state probability π ^(m) ), the measured waveforms Y ₁ , Y ₂ ,..., Y _t are observed, and the forward probability α _{t in} the combination of states as particles S _{p at} time _{t , p} .

Forward probability alpha _{t, p} are, for example, recurrence formula _{α t, p = Σα t-} 1 using a one time before the forward probability _{α t-1, r, r} P (S p | r) P (Y t | S _p ).

In the recurrence formula α _{t, p} = Σα _{t-1, r} P (S _p | r) P (Y _t | S _p ), r limits the number of factors that change state at one time to 1 state or less under the state transition restrictions that represents the combination r state as time t-1 of particles can transition to a combination of states of the particles S _p of time t.

Therefore, can become a combination of state r is a combination of the same state and the combination of the states of the particles S _p, only one state is a combination of state different from the combination of the states of the particles S _p.

In the recurrence formula α _{t, p} = Σα _{t−1, r} P (S _p | r) P (Y _t | S _p ), Σ represents the summation taken for all the combinations r of states.

In addition, in the recurrence formula α _{t, p} = Σα _{t-1, r} P (S _p | r) P (Y _t | S _p ), P (S _p | r) is expressed as a particle at time t−1. state combinations r Niite represents the transition probability of transition to the combination of the states of the particles S _p at time t, using transition probability P ^(m) _{i, j} of the FHMM for the entire model phi, according to equation (3) Desired.

As the initial value of the forward probability alpha _{t, p,} for example, from combination of the states of all, as particles S _p, the state #k of each factor #m constituting the combination of conditions selected randomly initial A product of state probabilities π ^(m) _k is adopted.

Approximate estimation unit 42, not all combinations of possible states is FHMM, only the combination of the states of the particles S _p, obtains the forward probability alpha _{t, p.}

When the number of particles S _p present in the time t and the R, R-number of forward probability α _{t, p} is required.

When the approximate estimation unit 42 determines the forward probability α _{t, p} for the R particles S _p = {S ₁ , S ₂ ,..., S _R } at time t, the approximate estimation unit 42 obtains the forward probability α _{t, p} . Based on the R particles S _p = {S ₁ , S ₂ ,..., S _R }, the particles are sampled.

That is, the approximate estimation unit 42 has a predetermined number of P particles in the descending order of the forward probability α _{t, p} from among the _R particles S _p = {S ₁ , S ₂ ,..., S _R }. particle _{S p = {S 1, S} 2, ···, S P} is sampled, only that P-number of particles S _p, leaving as particles of time t.

In the case of P ≧ R, R-number of particles S _p all is sampled.

The number P of sampling particles S _p is a value smaller than the number of combinations of possible states is FHMM, for example, is set based on the calculated cost, and the like which are acceptable to the monitoring system.

Approximate estimation unit 42, P number of particles S _p = time _{t {S 1, S 2,} ···, S P} is supposed to be sampled, the P number of particles _{S p = {S 1, S} 2, · .., S _P } predicts a particle at time t + 1, which is one hour later, under state transition restrictions.

Under the state transition limit state 1 time limit the number of factors that transition below 1 state, the particles S _q at time t + 1, and the combination of the states combination identical to the state of the particle S _p , the combination of the states of the particles S _p only one state and different combinations of states are predicted.

As described above, a four number of factors FHMM, each factor is, if having two states representing an ON state and an OFF state as operating states, for particles S _p with time t , a combination of the states of the same types 1 and combination of the states of the particles S _p of the time t, and combination of the states of four types of only one state is different from the combination of the states of the particles S _p at the time t of the combination of 5 kinds in total, it is expected as a particle S _q at time t + 1.

After predicting the particle S _q = {S ₁ , S ₂ ,..., S _Q } at time t + 1, the approximate estimation unit 42 determines the particle S _{p at} time t for the particle S _{q at} time t + 1. In the same manner as described above, the forward probability α _{t + 1, q} is obtained, and thereafter the forward probability at each time t = 1, 2,.

As described above, the approximate estimation unit 42 predicts a particle after one time under the state transition restriction using a combination of states of each factor #m as a particle _, and based on the forward probability α _{t, z} , While repeatedly sampling a predetermined number of particles, the forward probability α _{t, z} of the state combination z as a particle is obtained for each time t.

FIG. 17 is a diagram for explaining a method of obtaining a backward probability β _{t, p} by applying a particle filter to a combination z of FHMM states.

The approximate estimation unit 42 uses the observation probability P (Y _t | S _t ) from the evaluation unit 21 and the FHMM transition probability P ^(m) _{i, j} as the overall model φ stored in the model storage unit 13. Te, the time t, and are in the combination of the states of the particles S _p, then the measurement waveform _{Y t, Y t + 1,} ···, the backward probability beta _{t, p} observing Y _T seek.

The backward probability β _{t, p} is, for example, a recurrence formula β _{t, p} = ΣP (Y _t | S _p ) P (r | S _p ) β _t using the backward probability β _{t + 1, r} after one time It can be found according to _{+ 1, r} .

In the recurrence formula β _{t, p} = ΣP (Y _t | S _p ) P (r | S _p ) β _{t + 1, r} , r limits the number of factors that change state at one time to 1 state or less under the state transition restrictions that represents the combination r state as time t + 1 of particles can transition from a combination of conditions as particles S _p of time t.

Therefore, can become a combination of state r is a combination of the same state and the combination of the states of the particles S _p, only one state is a combination of state different from the combination of the states of the particles S _p.

In the recurrence formula β _{t, p} = ΣP (Y _t | S _p ) P (r | S _p ) β _{t + 1, r} , Σ represents the summation taken for all the state combinations r.

Furthermore, a recurrence formula beta _{t, p} = .SIGMA.P in _{| | (S p r) β} t + 1, r, P (Y t S p) P (r | S p) is the time t as a particle S _p In the state combination, it represents the transition probability of transition to the state combination r as a particle at time t + 1, and uses the FHMM transition probability P ^(m) _{i, j} as the overall model φ, according to equation (3) Desired.

Note that 1 is adopted as the initial value β _{T, p} of the backward probability β _{t, p} .

Approximate estimation unit 42, like the forward probability alpha _{t, p,} not all combinations of possible states is FHMM, only the combination of the states of the particles S _p, determine the backward probability beta _{t, p.}

When the number of particles S _p present in the time t and the R, R-number of backward probability β _{t, p} is required.

When the approximate estimation unit 42 determines the backward probability β _{t, p} for the R particles S _p = {S ₁ , S ₂ ,..., S _R } at time t, the approximate estimation unit 42 determines the backward probability β _{t, p} . Based on the R particles S _p = {S ₁ , S ₂ ,..., S _R }, the particles are sampled.

That is, the approximate estimation unit 42 determines the backward probability β _{t, R} from the _R particles S _p = {S ₁ , S ₂ ,..., S _R }, as in the case of the forward probability α _{t, p . A} predetermined number of P particles S _p = {S ₁ , S ₂ ,..., S _P } are sampled in descending order of _{p, and} only the P particles Sp are left as particles at time t. .

Approximate estimation unit 42, P number of particles S _p = time _{t {S 1, S 2,} ···, S P} is supposed to be sampled, the P number of particles _{S p = {S 1, S} 2, · .., S _P } predicts the particle at time t−1, which is one hour before, under the state transition restriction.

Under the state transition limit state 1 time limit the number of factors that transition below 1 state, as the time t-1 particle S _q, and the combination of the state combination identical to the state of the particle S _p , the combination of the states of the particles S _p of time t only one state and different combinations of states are predicted.

As described above, a four number of factors FHMM, each factor is, if having two states representing an ON state and an OFF state as operating states, for particles S _p with time t , a combination of the states of the same types 1 and combination of the states of the particles S _p of the time t, and combination of the states of four types of only one state is different from the combination of the states of the particles S _p at the time t of the combination of 5 kinds in total, it is expected as a time t-1 particle S _q.

Approximate estimation unit 42, the particles of time _{t-1 S q = {S} 1, S 2, ···, S Q} for, as in the case of particle S _p of time t, backward probability beta _{t + 1, q} is obtained, and thereafter, the backward probability at each time t = T, T-1,.

As described above, the approximate estimation unit 42 predicts a particle one time ago under the state transition restriction with a combination of states of each factor #m as particles _, and based on the backward probability β _{t, z} , While repeatedly sampling a predetermined number of particles, the backward probability β _{t, z} of the state combination z as a particle is obtained for each time t.

FIG. 18 is a diagram for explaining a method for obtaining a posteriori probability γ _{t, p} by applying a particle filter to a combination z of FHMM states.

Approximate estimation unit 42, at each time t, the particles forward probability alpha _{t, p} is obtained (hereinafter also referred to as forward particles) and combination of the states of the S _p, backward probability beta _{t, p 'is} obtained When the combination of states as particles (hereinafter also referred to as backward particles) S _{p ′} is the same combination z, the forward probability α _{t, z} and the backward probability β _{t, z} for that state combination _z Is used to obtain the posterior probability γ _{t, z} in the state combination z at time t in the FHMM as the overall model φ according to the equation (23).

・・・（２４）

... (24)

Here, Σ of the denominator of the right side of the equation (24), w a _'state that both the remaining common combination S _p ∩S _p' forward particles S _p and backward particles S _p of time t It represents a summation that takes everything.

Incidentally, forward probability alpha _{t, z} combinations z state does not remain as a positive particle S _p at time t, and, backward probabilities of combinations z state does not remain as a backward particles S _{p 'at} time t beta _{t, z} For example, both are set to 0.

Therefore, the posterior probability γ _{t, z} of the combination of the states other than the combination S _p ∩S _{p ′} of the state remaining in common to both the forward particle S _p and the backward particle S _{p ′ at} time t is 0. .

In the monitoring system of FIG. 15, the monitoring of FIG. 3 is performed except that the posterior probability <S ^(m) _t > is obtained under the state transition restriction by applying a particle filter to the state combination z of the FHMM. Processing similar to the system is performed.

  FIGS. 19 and 20 are flowcharts for explaining the processing of the E step performed in step S13 of FIG. 6 by the monitoring system of FIG.

  FIG. 20 is a flowchart following FIG.

  In step S81, the approximate estimation unit 42 initializes time (a variable representing) t to 1 as an initial value for the forward probability, and the process proceeds to step S82.

At step S82, the approximate estimation unit 42, from the combination of the states that can take the FHMM, a combination of a predetermined number of states, for example, randomly selects as particles S _p (sampling). In step S82, the all combinations of the possible states is FHMM, can be chosen as a particle S _p.

The approximate estimation unit 42 obtains the initial value α _{1, p} of the forward probability α _{t, p} for the state combination as the particle Sp, and the process proceeds from step S82 to step S83.

In step S83, it is determined whether or not the time t is equal to the last time (sequence length) T of the sequence of the measurement waveform Y _t = {Y ₁ , Y ₂ ,..., Y _T }.

If it is determined in step S83 that time t is not equal to the last time T of the series of measurement waveforms Y _t , that is, if time t is less than time T, the process proceeds to step S84 and approximate estimation is performed. part 42, under the state transition limit as a predetermined condition, the combination of the states of the particles S _p, predicts a time t + 1 the particle is after 1 time, the process proceeds to step S85.

  In step S85, the approximate estimation unit 42 increments the time t by 1, and the process proceeds to step S86.

In step S86, the evaluation unit 21 analyzes the variance C of the FHMM as the overall model φ stored in the model storage unit 13, the eigen waveform W ^(m) , and the measured waveform Y _t = {from the data acquisition unit 11. Y _1, Y _2, ···, with Y _T}, at time t, each particle _{S p = {S 1, S} 2, ···, the combination of the states of the S _R}, equation (21) Find the observation probability P (Y _t | S _p ). The evaluation unit 21 supplies the observation probability P (Y _t | S _p ) to the approximate estimation unit 42, and the process proceeds from step S86 to step S87.

In step S87, the approximate estimation unit 42, the observation probability P (Y _t | S _p ) from the evaluation unit 21, and the transition probability P ^(m) _i of the FHMM as the entire model φ stored in the model storage unit 13 _, with _j, measured waveform Y _1, Y _2, · · ·, observing Y _t, at time t, determine the forward probability alpha _{t, p} being in the combination of the states of the particles S _p, the process, Proceed to step S88.

In step S88, the approximate estimation unit 42, based on the forward probability alpha _{t, p,} from the particles S _p, in descending order of forward probability alpha _{t, p,} sampling a predetermined number of particles, leaving the particles at time t .

  Thereafter, the process returns from step S88 to step S83, and the same process is repeated thereafter.

If it is determined in step S83 that the time t is equal to the last time T in the series of measurement waveforms Y _t , the process proceeds to step S91 in FIG.

  In step S91, the approximate estimation unit 42 initializes the time t to T as an initial value for the backward probability, and the process proceeds to step S92.

In step S92, the approximate estimation unit 42, from the combination of the states that can take the FHMM, a combination of a predetermined number of states, for example, randomly selects as particles S _p. In step S92, all combinations of possible states is FHMM, may select as particles S _p.

The approximate estimation unit 42 obtains the initial value α _{T, p} of the backward probability β _{t, p} for the state combination as the particle Sp, and the process proceeds from step S92 to step S93.

In step S93, it is determined whether or not the time t is equal to the first time 1 of the series of the measurement waveform Y _t = {Y ₁ , Y ₂ ,..., Y _T }.

If it is determined in step S93 that the time t is not equal to the first time 1 of the series of the measured waveform Y _t , that is, if the time t is greater than the time 1, the process proceeds to step S94 and is approximated. estimation unit 42, under the state transition limit as a predetermined condition, the combination of the states of the particles S _p, predicts the time t-1 of the particle is one time before, the process proceeds to step S95.

  In step S95, the approximate estimation unit 42 decrements the time t by 1, and the process proceeds to step S96.

In step S <b> 96, the evaluation unit 21 analyzes the variance C of the FHMM as the overall model φ stored in the model storage unit 13, the eigen waveform W ^(m) , and the measured waveform Y _t = {from the data acquisition unit 11. Y _1, Y _2, ···, with Y _T}, at time t, each particle _{S p = {S 1, S} 2, ···, the combination of the states of the S _R}, equation (21) Find the observation probability P (Y _t | S _p ). The evaluation unit 21 supplies the observation probability P (Y _t | S _p ) to the approximate estimation unit 42, and the process proceeds from step S96 to step S97.

In step S97, the approximate estimation unit 42 observes the observation probability P (Y _t | S _t ) from the evaluation unit 21 and the transition probability P ^(m) _i of the FHMM as the entire model φ stored in the model storage unit 13. _, with _j, at time t, and are in the combination of the states of the particles S _p, then the measurement waveform _{Y t, Y t + 1,} ···, the backward probability beta _{t, p} observing Y _T determined The process proceeds to step S98.

In step S98, the approximate estimation unit 42, based on the backward probability beta _{t, p,} from the particles S _p, in order of backward probability beta _{t, p,} sampling a predetermined number of particles, leaving the particles at time t .

  Thereafter, the processing returns from step S98 to step S93, and the same processing is repeated thereafter.

If it is determined in step S93 that the time t is equal to the first time 1 of the series of measurement waveforms Y _t , the process proceeds to step S99.

In step S99, the approximate estimation unit 42 uses the forward probability α _{t, z} (= α _{t, p} ) and the backward probability β _{t, z} (= β _{t, p} ), according to the equation (24), and the overall model φ In the FHMM, the posterior probability γ _{t, z} in the state combination z is obtained at time t, and the process proceeds to step S100.

Thereafter, in steps S100 and S101, the same processing as in steps S25 and S26 of FIG. 7 is performed, and the posterior probabilities <S ^(m) _t >, <S ^(m) _t S ⁽ⁿ⁾ _t ′>, and <S ^(m) _t-1 S ^(m) _t > is obtained.
The

As described above, the approximate estimation unit 42 can obtain the posterior probability <S ^(m) _t > and the like under the state transition restriction by applying the particle filter to the state combination z of the FHMM. It is possible to omit the calculation for the combination of the states having low characteristics and to improve the calculation efficiency such as the posterior probability <S ^(m) _t >.

[Fourth embodiment of a monitoring system to which the present technology is applied]

  FIG. 21 is a block diagram illustrating a configuration example of the fourth embodiment of the monitoring system to which the present technology is applied.

  In the figure, portions corresponding to those in FIG. 3 are denoted by the same reference numerals, and description thereof will be omitted below as appropriate.

  The monitoring system in FIG. 21 is common to the monitoring system in FIG. 3 in that it has a data acquisition unit 11 to a data output unit 16.

  However, the monitoring system of FIG. 21 is different from the monitoring system of FIG. 3 in that a restricted separation learning unit 51 is provided instead of the waveform separation learning unit 31 of the model learning unit 14.

In the separation learning unit 31 in FIG. 3, the eigen waveform W ^(m) is obtained without any particular restriction. However, the restricted waveform separation learning unit 51 is unique under a predetermined restriction peculiar to home appliances (home appliance separation). Determine the waveform W ^(m) .

Here, in the separation learning unit 31 in FIG. 3, no particular restriction is imposed on the calculation of the eigen waveform W ^{(m) in} the M step.

Therefore, the separation learning unit 31 in FIG. 3 calculates the eigen waveform W ^(m) by waveform separation learning only considering that the measurement waveform Y _t is a waveform in which a plurality of current waveforms are superimposed.

That is, in the waveform separation learning of the separation learning unit 31 in FIG. 3, the measurement waveform Y _t is separated into eigen waveforms W ^(m) as the plurality of current waveforms on the assumption that the plurality of current waveforms are superimposed. .

If there are no restrictions on the calculation of the eigen waveform W ^(m), the degree of freedom (redundancy) of the solution for waveform separation learning, that is, the degree of freedom for multiple current waveforms that can be separated from the measured waveform Y _t is Since it is very high, a natural waveform W ^(m) as a current waveform that is inappropriate as the current waveform of home appliances may be required.

Therefore, the constrained waveform separation learning unit 51 performs waveform separation learning under a predetermined restriction specific to home appliances, so that a specific waveform W ^(m) as a current waveform inappropriate for the home appliance current waveform is obtained. The characteristic waveform W ^(m) as an accurate current waveform of the home appliance is obtained.

  Here, the waveform separation learning performed under a predetermined restriction is also referred to as restricted waveform separation learning.

  In the monitoring system of FIG. 21, processing similar to that of the monitoring system of FIG. 3 is performed except that constrained waveform separation learning is performed.

  In restricted waveform separation learning, examples of restrictions unique to home appliances include load restrictions and base waveform restrictions.

The load constraint is the multiplication of the intrinsic waveform W ^(m) as the current waveform of the home appliance #m and the voltage waveform V _t corresponding to the voltage waveform of the voltage applied to the home appliance #m, that is, the measurement waveform Y _t that is the current waveform. This is a restriction that the power consumption U ^(m) of the home appliance calculated by the above does not become a negative value (the home appliance #m does not generate power).

The base waveform constraint is the current waveform of the current consumption in each operating state of the home appliance #m. The unique waveform W ^(m) is a combination of one or more waveforms prepared in advance as the base waveform for the home appliance #m. It is a restriction that it is expressed by.

  FIG. 22 is a flowchart for explaining the process of step M14 in step S14 of FIG. 6 performed by the monitoring system of FIG.

In step S121, the constrained waveform separation learning unit 51 measures the measured waveform Y _t from the data acquisition unit 11, and the posterior probabilities <S ^(m) _t > and <S ^(m) _t S from the estimation unit 22. ^{(n) Using} _t ′> and performing waveform separation learning under load constraints, the updated value W ^{(m) new} of the eigen waveform W ^(m) is obtained, and by the updated value W ^{(m) new} , The natural waveform W ^(m) stored in the model storage unit 13 is updated, and the process proceeds to step S122.

That is, the constrained waveform separation learning unit 51 updates the eigen waveform W ^(m) by solving Equation (25) using constrained quadratic programming as waveform separation learning under load constraints. Find the value W ^{(m) new} .

・・・（２５）

... (25)

Here, in Formula (25), the unique waveform W to be updated value W ^{(m) new new} unique waveform W ^(m), constraints satisfying the expression 0 ≦ V'W as load constraints are imposed.

V is a column vector of D rows representing the voltage waveform V _t corresponding to the current waveform as the measurement waveform Y _t , and V ′ is a row vector obtained by transposing the column vector V.

The updated value W ^new , the posterior probability <S _t '>, and <S _t S _t '> in the equation (25) are the same as those described in the equation (12).

That is, the update value W ^new of the eigen waveform is a matrix of D rows and K × M columns, and the column vector of (m−1) K + k columns is the eigen waveform W ^{(m )} The updated value of _k .

The posterior probability <S _t '> is a row vector of K × M columns, and its (m-1) K + k column component is the state probability of being in state #k with factor #m at time t. It has become.

The posterior probability <S _t S _t '> is a matrix of K × M rows and K × M columns, and the (m−1) K + k rows (n−1) K + k ′ column components are time t is the state probability of being in state #k of factor #m and in state #k ′ of other factor #n.

The characteristic waveform W in Expression (25) is a matrix of D rows and K × M columns similar to the update value W ^{new of the characteristic} waveform, and minimizes X of argmin {X} on the right side of Expression (25). The natural waveform W is obtained by quadratic programming as the update value W ^new of the natural waveform.

  In steps S122 and S123, processing similar to that in steps S32 and S33 in FIG. 9 is performed.

  FIG. 23 is a diagram for explaining the load constraint.

That is, FIG. 23A shows an eigen waveform W ^(m) as a current waveform obtained by waveform separation learning without load restriction, and FIG. 23B shows a current obtained by waveform separation learning performed under load restriction. The characteristic waveform W ^(m) is shown as a waveform.

In FIG. 23, assuming that the number of factors M is two, the characteristic waveform W ⁽¹⁾ _{k of} state #k with factor # 1 and the characteristic waveform W ⁽²⁾ _{k of} state #k ′ with factor # 2 _' Is shown.

In Figure 23A and 23B, the current waveform is unique waveform W ⁽¹⁾ _k factor # 1, both have become the voltage waveform V _t and phase.

However, in FIG. 23A, since the load is not restricted, the current waveform that is the intrinsic waveform W ⁽²⁾ _{k ′} of factor # 2 is in reverse phase to the voltage waveform V _t .

On the other hand, in FIG 23B, a result of the load are imposed, the current waveform is a factor # 2 specific waveform W ⁽²⁾ _{k ',} which is a voltage waveform V _t and phase.

The measured waveform Y _t can be separated into the characteristic waveform W ⁽¹⁾ _{k of} factor # 1 shown in FIG. 23A and the characteristic waveform W ⁽²⁾ _{k ′ of} factor # 2 as shown in FIG. 23B. It is also possible to separate the characteristic waveform W ⁽¹⁾ _k of factor # ¹ and the characteristic waveform W ⁽²⁾ _{k ′ of} factor # 2.

However, when the measurement waveform Y _t is separated into the characteristic waveform W ⁽¹⁾ _k of factor # ¹ and the characteristic waveform W ⁽²⁾ _{k ′ of} factor # 2 shown in FIG. 23A, the characteristic waveform W ⁽ Appliances # ^{2 2)} _{k 'corresponds} to a factor # 2 are opposite phase and voltage waveform V _t is will be doing power generation, it may become difficult to perform the appropriate consumer electronics separation .

On the other hand, when a load constraint is imposed, the measurement waveform Y _t is separated into the characteristic waveform W ⁽¹⁾ _k of factor # ¹ and the characteristic waveform W ⁽²⁾ _{k ′ of} factor # 2 shown in FIG. 23B. Is done.

Both the natural waveform W ⁽¹⁾ _k and the natural waveform W ⁽²⁾ _{k ′} in FIG. 23B are in phase with the voltage waveform V _t, and therefore correspond to factor # 1 according to the load constraint. The home appliance # 1 and the home appliance # 2 corresponding to the factor # 2 are both loads that consume power, and appropriate home appliance separation can be performed.

  FIG. 24 is a diagram for explaining the base waveform restriction.

In FIG. 24, Y is a matrix of D rows and T columns in which measured waveforms Y _t that are column vectors of D rows are arranged from left to right in the order of time t. Column vector of t th column of D rows and T is a matrix of rows measurement waveform Y is adapted to measure the waveform Y _t at time t.

In FIG. 24, W is a matrix of D rows and K × M columns in which the intrinsic waveform W ^{(m) of} factor #m, which is a matrix of D rows and K columns, is arranged in the order of factor #m from left to right. The column vector of (m−1) K + k columns of the eigen waveform W that is a matrix of D rows and K × M columns is the eigen waveform W ^(m) _k of the state #k with the factor #m.

  In FIG. 24, F represents K × M.

In FIG. 24, <S _t > is a column vector of F = K × M rows in which posterior probabilities <S ^(m) _t > at time t, which are column vectors of K rows, are lined up from top to bottom in the order of factor #m. Is a matrix of K × M rows and T columns arranged from left to right in order of time t. The component of the (m-1) K + k rows and t columns of the posterior probability <S _t > that is a matrix of K × M rows and T columns becomes the state probability of being in state #k with factor #m at time t. ing.

In the waveform separation learning, as shown in FIG. 24, in the FHMM, the product W × <S _t > of the eigen waveform W and the posterior probability <S _t > is observed as the measured waveform Y. Desired.

As described above, the base waveform constraint is that the specific waveform W ^(m) , which is the current waveform of the consumption current of each operating state of the home appliance #m, has a plurality of waveforms prepared in advance as base waveforms for the home appliance #m. In the waveform separation learning that imposes the base waveform restriction, as shown in FIG. 24, the eigen waveform W includes a predetermined number N of base waveforms B, a predetermined number of combinations, and the like. The characteristic waveform W is obtained as represented by the product B × A with the coefficient A.

Here, among the base waveform B of a predetermined number N, the n-th base waveform B, When be expressed as B _n, B _n, for example, a column vector of D rows and component sample values of the waveform The base waveform B is a matrix of D rows and N columns in which the base waveform B _n, which is a column vector of D rows, is arranged from left to right in the order of index n.

The coefficient A is a matrix of N rows and K × M columns, and the component in the nth row (m−1) K + k column is the eigen waveform W ^(m) _k and the N base waveforms B ₁ and B _2. ,..., B _N are coefficients that are multiplied by the _nth base waveform B _n to represent the combination (superimposition).

Here, the unique waveform W ^(m) _k, to express a combination of N base waveform B ₁ to B _N, as coefficients to the N base waveform B ₁ not being multiplied with the B _N, for example, An N-row column vector is represented as a coefficient A ^(m) _k, and the coefficient A ^(m) _k is arranged in N rows in the order of K states # 1 to #K in factor #m from left to right. a matrix of K rows, when it represents a coefficient a ^(m), the coefficient a is a coefficient a ^(m), factor #m turn, is a matrix arranged from left to right.

  The base waveform B may be prepared (obtained) by performing base extraction such as ICA (Independent Component Analysis) and NMF (Non-negative Matrix Factorization) used in image processing, etc. on the measurement waveform Y. it can.

  In addition, when a manufacturer or the like publishes a base waveform of a home appliance on a site or the like, the base waveform B can be prepared by accessing the site.

  FIG. 25 is a flowchart for explaining the process of step M in step S14 in FIG. 6 performed by the monitoring system in FIG.

In step S131, the constrained waveform separation learning unit 51, the measured waveform Y _t from the data acquisition unit 11, and the posterior probabilities <S ^(m) _t > and <S ^(m) _t S from the estimation unit 22 are obtained. ⁽ⁿ⁾ _t '> is used to perform waveform separation learning under the base waveform constraint, thereby obtaining an updated value W ^{(m) new} of the eigen waveform W ^(m) , and using the updated value W ^{(m) new} Then, the natural waveform W ^(m) stored in the model storage unit 13 is updated, and the process proceeds to step S132.

That is, the constrained waveform separation learning unit 51 solves the equation (26) using the constrained quadratic programming method as the waveform separation learning under the base waveform constraint, thereby obtaining the base waveform as the eigen waveform W. The coefficient A ^new for expressing with the combination of B is obtained.

・・・（２６）

... (26)

  Here, in the equation (26), the natural waveform W is expressed by a product B × A of a predetermined number N of base waveforms B and a predetermined coefficient A, and the coefficient A is minimized ( minA) is imposed as a base waveform constraint.

  Minimizing the coefficient A (minA) means minimizing the value (magnitude) of each component of the matrix of N rows and K × M, which is the coefficient A, and hence N that is the coefficient A. This means that the matrix of rows K x M is made as sparse as possible.

By making the matrix of N rows and K × M columns, which is the coefficient A, as sparse as possible, the natural waveform W ^(m) _k is represented by a combination of as few base waveforms B _n as possible.

When the constrained waveform separation learning unit 51 obtains the coefficient A ^new according to the equation (26), the updated value W of the eigen waveform W ^(m) is obtained using the coefficient A ^new and the base waveform B according to the equation (27). ^{(m) Find new} .

・・・（２７）

... (27)

  In steps S132 and S133, processing similar to that in steps S32 and S33 in FIG. 9 is performed.

  When the current consumed by home appliances is a combination of one or more of the base waveforms, by imposing a base waveform constraint, the waveform of the current that cannot be consumed by home appliances is an intrinsic waveform. Therefore, it is possible to obtain a proper waveform suitable for home appliances.

  In the above description, the load constraint and the base waveform constraint are imposed separately. However, the load constraint and the base waveform constraint can be imposed simultaneously.

When imposing both the load constraint and the base waveform constraint at the same time, the constrained waveform separation learning unit 51 performs Equation (28) as waveform separation learning under the load constraint and the base waveform constraint in the M step. By solving using constrained quadratic programming, the coefficient A ^new is a coefficient for expressing the eigen waveform W as a combination of the base waveform B, and the power consumption of the home appliance calculated using the eigen waveform W is negative. The coefficient A ^new that is not set to the value of is obtained.

・・・（２８）

... (28)

Then, the constrained waveform separation learning unit 51 uses the coefficient A ^new obtained according to the equation (28) and the base waveform B, and the updated value W ^{(m) of the} eigen waveform W ^(m) according to the equation (29 ^{). Find new} .

・・・（２９）

... (29)

[Fifth embodiment of a monitoring system to which the present technology is applied]

  FIG. 26 is a block diagram illustrating a configuration example of the fifth embodiment of the monitoring system to which the present technology is applied.

  In the figure, portions corresponding to those in FIG. 3 are denoted by the same reference numerals, and description thereof will be omitted below as appropriate.

  The monitoring system of FIG. 26 is common to the monitoring system of FIG. 3 in that the data acquisition unit 11 to the data output unit 16 are included.

  First, however, the monitoring system of FIG. 26 differs from the monitoring system of FIG. 3 in that an individual distributed learning unit 52 of FIG. 12 is provided instead of the distributed learning unit 32 of the model learning unit 14. .

  Second, the monitoring system of FIG. 26 is different from the monitoring system of FIG. 3 in that an approximate estimation unit 42 of FIG. 15 is provided instead of the estimation unit 22 of the state estimation unit 12.

  Third, the monitoring system of FIG. 26 is different from the monitoring system of FIG. 3 in that a constrained separation learning unit 51 of FIG. 21 is provided instead of the waveform separation learning unit 31 of the model learning unit 14. .

Therefore, in the monitoring system of FIG. 26, individual variance learning for obtaining individual variance (individual variance) σ ^(m) is performed for each factor #m or for each state #k of each factor #m.

Furthermore, in the monitoring system of FIG. 26, approximate estimation is performed to obtain posterior probabilities (state probabilities) <S ^(m) _t > and the like under a state transition restriction that restricts the number of factors that cause a state transition at one time. .

In the monitoring system of FIG. 26, the natural waveform W ^(m) is obtained by constrained waveform separation learning.

  In the monitoring system, each of individual dispersion learning, approximate estimation, and constrained waveform separation learning is performed independently, or individual dispersion learning, approximate estimation, and constrained waveform separation learning are all performed. Any two of individual distributed learning, approximate estimation, and constrained waveform separation learning can be performed.

  [Application of monitoring system other than home appliance separation]

  As described above, in the monitoring system that learns FHMM, the case where the current waveform as the total data is monitored (monitored) and the household appliances are separated has been described. However, the monitoring system that learns FHMM has one or more other signals. Can be applied to any application that monitors a superimposed signal and separates the signal superimposed on the superimposed signal.

  FIG. 27 is a diagram for explaining the outline of speaker separation by the monitoring system that performs FHMM learning.

According to the monitoring system that performs FHMM learning, instead of the current waveform used as the measurement waveform Y _t in the home appliance separation, the voice signal on which the utterances of a plurality of speakers are superimposed is used. Thus, speaker separation can be performed to separate the voices of the speakers from the speech signal on which the utterances are superimposed.

  In a monitoring system that performs FHMM learning, individual distributed learning, approximate estimation, and constrained waveform separation learning can be performed when speaker separation is performed as in the case of household appliance separation.

In speaker separation, when using individual variance σ ^(m) for each factor #m or each state #k of each factor #m, the FHMM's ability to express is improved compared to using variance C of 1. Therefore, the accuracy of speaker separation can be improved.

  In addition, since it is very unlikely that all utterance states of multiple speakers will change at each time, even if the number of speakers whose utterance state changes at one time is limited to one or several people There is no significant effect on the accuracy of speaker separation.

Therefore, in speaker separation, the number of factors that change state at one time is limited to, for example, one or less (restricting the number of speakers whose utterance state changes to one or less), Approximate estimation for obtaining the posterior probability <S ^(m) _t > and the like can be performed. Then, under the state transition limit, according to the approximate estimate to determine the posterior probability <S ^(m) _t> like, the posterior probability gamma _{t, z} used for determining the posterior probability <S ^(m) _t> etc. Since the number of state combinations z that need to be strictly calculated is greatly reduced, the amount of calculation can be greatly reduced.

Further, in the speaker separation, the individual waveform W ^(m) is a human voice waveform, and therefore the frequency component exists in a frequency band that can be taken by the human voice. Therefore, in constrained waveform separation learning performed by speaker separation, the frequency component of the individual waveform W ^(m) is limited to the frequency component within the frequency band that can be taken by human speech. Can be adopted. In this case, an appropriate natural waveform W ^(m) can be obtained as the waveform of human speech.

  In the speaker separation, as the constraints imposed by the constrained waveform separation learning, for example, a base waveform constraint can be adopted as in the case of home appliance separation.

[Sixth embodiment of a monitoring system to which the present technology is applied]

  FIG. 28 is a block diagram illustrating a configuration example of the sixth embodiment of the monitoring system to which the present technology is applied.

  In the figure, portions corresponding to those in FIG. 3 are denoted by the same reference numerals, and description thereof will be omitted below as appropriate.

  The monitoring system of FIG. 28 is common to the monitoring system of FIG. 3 in that the data acquisition unit 11 to the data output unit 16 are included.

  However, firstly, the monitoring system of FIG. 28 is different from the state estimation unit 12 in that an evaluation unit 71 and an estimation unit 72 are provided instead of the evaluation unit 21 and the estimation unit 22, respectively. Different from the monitoring system.

  Secondly, in the monitoring system of FIG. 28, the model learning unit 14 is provided with a constrained separation learning unit 81 instead of the waveform separation learning unit 31, and a distributed learning unit 32 and a state variation learning unit 33 are provided. This is different from the monitoring system of FIG.

In the monitoring system of FIG. 3, household appliances are separated using the FHMM as the overall model φ. However, in the monitoring system of FIG. 28, a model that is not an FHMM, that is, only the eigen waveform W ^(m) , for example, is used as a model parameter. The home appliance separation is performed using the model (hereinafter also referred to as a waveform model) having the model as the overall model φ.

Therefore, in FIG. 28, the waveform model is stored in the model storage unit 13 as the overall model φ. Here, the waveform model has an eigen waveform W ^(m) as a model parameter. In such a waveform model, one eigen waveform W ^(m) corresponds to the home appliance model #m.

The state estimation unit 12 includes an evaluation unit 71 and an estimation unit 72, and uses the measured waveform Y _t as the sum data supplied from the data acquisition unit 11 and the waveform model stored in the model storage unit 13, State estimation is performed to obtain the operating states of M home appliances # 1, # 2,.

  That is, the evaluation unit 71 evaluates the degree to which the current waveform Y supplied from the data acquisition unit 11 is observed in each home appliance model #m constituting the waveform model as the overall model φ stored in the model storage unit 13. The evaluated value E is obtained and supplied to the estimation unit 72.

The estimation unit 72 uses the evaluation value E supplied from the evaluation unit 71 to estimate the operating state C ^(m) _{t, k} at each time t of the home appliance represented by each home appliance model #m by, for example, integer programming. The model learning unit 14, the label acquisition unit 15, and the data output unit 16 are supplied.

Here, for example, the estimation unit 72 estimates the operating state C ^(m) _{t, k} of the home appliance #m by solving the integer programming problem of Expression (30) by the integer programming method.

・・・（３０）

... (30)

Here, in Expression (30), E is an error between the measured waveform Y _t and the current waveform ΣΣW ^(m) _k C ^(m) _{t, k} as the total data observed in the waveform model as the overall model φ. In the estimation unit 72, an operating state C ^(m) _{t, k} that minimizes the error E is obtained.

Further, in Equation (30), the specific waveform W ^(m) _k represents a specific waveform that is a current waveform specific to the operating state C ^(m) _{t, k} of the home appliance #m, and is similar to the measurement waveform Y _t. , D column vector.

Furthermore, in the equation (30), K (m) represents the number of operating states C ^(m) _{t, k} (number of types) of the home appliance #m.

Operating state C ^(m) _{t, k} is an integer greater than or equal to 0, is a scalar value, and represents the operating state of home appliance #m at time t. The operating state C ^(m) _{t, k} corresponds to, for example, the mode (setting) of the household appliance #m, and the mode that can be turned on in the household appliance #m is limited to, for example, one or less. And

  In household appliance #m, the fact that the number of modes that can be turned on is limited to one or less is expressed by equation (31).

・・・（３１）

... (31)

According to the equation (31), the possible value of the operating state C ^(m) _{t, k} that is an integer of 0 or more is 0 or 1.

  For estimating the operating status of home appliances using integer programming, for example, “Non-intrusive operating status monitoring system for electrical equipment”, Shinkichi Inagaki, Tsukasa Egami, Tatsuya Suzuki (Nagoya University), Hisae Nakamura, Ito Koichi (Toenec Co., Ltd.), 42nd Society of Instrument and Control Engineers Discrete Event System Study Group, pp.33-38, Dec 20, 2008, Osaka University.

As described above, when the estimation unit 72 estimates the operating state C ^(m) _{t, k} of the home appliance #m by solving the equation (30), the evaluation unit 71 is supplied from the data acquisition unit 11. Using the measured waveform Y _t and the inherent waveform W ^(m) _k of the waveform model stored in the model storage unit 13, the error E of Expression (30) is obtained and supplied to the estimation unit 72 as the evaluation value E. .

In the model learning unit 14, the constrained waveform separation learning unit 81 is similar to the constrained waveform separation learning unit 51 in FIG. 21, and performs waveform separation learning under a predetermined restriction specific to home appliances (constrained waveform separation learning). Is performed to prevent the eigen waveform W ^(m) _{k from} being obtained as an inappropriate current waveform for the home appliance, and to obtain the eigen waveform W ^(m) _k as an accurate current waveform of the home appliance. .

That is, the constrained waveform separation learning unit 81 uses the measured waveform Y _t supplied from the data acquisition unit 11 and the operating state C ^(m) _{t, k} of the home appliance #m supplied from the estimation unit 72, for example, , under load constraints, by performing waveform separation learning, obtains the updated value W ^{(m) new} _k unique waveform W ^(m) _k, by its updated value W ^{(m) new} _k, in the model storage unit 13 The stored natural waveform W ^(m) _k is updated.

That is, the constrained waveform separation learning unit 81 performs, as the waveform separation learning under the load restriction, for example, by solving the quadratic programming problem of Expression (32) by the quadratic programming method, thereby generating the eigen waveform W ^(m) _k of obtaining the updated value W ^{(m) new} _k.

・・・（３２）

... (32)

Here, in the equation (32), a constraint that satisfies the equation 0 ≦ V′W as a load constraint is imposed on the eigen waveform W that becomes the updated value W ^{(m) new} _{k of} the eigen waveform W ^(m) _k . .

In Equation (32), V is a column vector of D rows representing the voltage waveform V _t corresponding to the current waveform as the measurement waveform Y _t , and V ′ is a row vector obtained by transposing the column vector V.

Further, assuming that the mode of home appliance #m indicating that the operating state C ^(m) _{t, k} is on or off is mode #k, in formula (32), W is the (m−1) K + k sequence. Is a matrix of D rows and K × M columns, which is an eigen waveform of mode #k of home appliance #m (when ON). Here, K represents, for example, the maximum value among K (1), K (2),..., K (M).

The updated value W ^new of the eigen waveform is a matrix of D rows and K × M columns, and the column vector of (m−1) K + k columns is the eigen waveform W ^(m) _k of state #k with factor #m. Is the updated value.

In Expression (32), C _t is a column vector of K × M rows in which (m−1) K + k rows of components are in the operating state C ^(m) _{t, k} .

  The restricted waveform separation learning unit 81 can impose other restrictions. That is, in the constrained waveform separation learning unit 81, for example, in the same manner as the constrained waveform separation learning unit 51 in FIG. 21, a base waveform constraint is imposed instead of the load constraint, and both the load constraint and the base waveform constraint are performed. Can be imposed.

  FIG. 29 is a flowchart for explaining the learning process (learning process) of the waveform model performed by the monitoring system of FIG.

In step S151, the model learning unit 14 initializes the eigen waveform W ^(m) _k as the model parameter φ of the overall model stored in the model storage unit 13, and the process proceeds to step S152.

Here, each component of the column vector of D rows as the eigen waveform W ^(m) _k is initialized using, for example, a random number.

In step S152, the data acquisition unit 11 acquires a current waveform for a predetermined time T, and the current waveforms at the respective times t = 1, 2,..., T are measured waveforms Y ₁ , Y ₂ ,. , Y _T are supplied to the state estimation unit 12 and the model learning unit 14, and the process proceeds to step S153.

  Here, the data acquisition unit 11 acquires a voltage waveform as well as a current waveform at time t = 1, 2,... The data acquisition unit 11 supplies voltage waveforms at times t = 1, 2,..., T to the data output unit 16.

  In the data output unit 16, in the information presentation process (FIG. 10), the voltage waveform from the data acquisition unit 11 is used for calculation of power consumption.

In step S153, the evaluation unit 71 of the state estimation unit 12 uses the measurement waveforms Y ₁ to Y _T from the data acquisition unit 11 and the characteristic waveform W ^(m) _k of the waveform model stored in the model storage unit 13. Thus, the error E as the evaluation value E of the equation (30) for obtaining the operating state C ^(m) _{t, k} of each home appliance #m is _obtained .

  Furthermore, the evaluation unit 71 supplies the error E to the estimation unit 72, and the process proceeds from step S153 to step S154.

In step S154, the estimation unit 72 estimates the operating state C ^(m) _{t, k} of each home appliance #m by minimizing the error E of the expression (30) from the evaluation unit 71, and the model learning unit 14 The label acquisition unit 15 and the data output unit 16 are supplied, and the process proceeds to step S155.

In step S155, the constrained waveform separation learning unit 81 of the model learning unit 14 measures the measured waveform Y _t supplied from the data acquisition unit 11 and the operating state C ^{(m) of the} home appliance #m supplied from the estimation unit 72. An updated value W ^{(m) new} _k of the eigen waveform W ^(m) _k is obtained by performing waveform separation learning under predetermined constraints such as load constraints using _{t, k} .

That is, the constrained waveform separation learning unit 81 solves the equation (32), for example, as waveform separation learning under a load constraint, thereby obtaining the updated value W ^{(m) new} _k of the eigen waveform W ^(m) _k. Ask.

Then, the constrained waveform separation learning unit 81 updates the eigen waveform W ^(m) _k stored in the model storage unit 13 with the updated value W ^{(m) new} _k of the eigen waveform W ^(m) _k. The process proceeds from step S155 to step S156.

  In step S156, the model learning unit 14 determines whether the convergence condition of the model parameter φ is satisfied.

Here, the convergence condition of the model parameter φ includes, for example, the estimation of the operating state C ^(m) _{t, k} in step S154 and the update of the eigen waveform W ^(m) _k by constrained waveform separation learning in step S155. It has been repeated a predetermined number of times set in advance, the change amount of the error E in the equation (30) between before and after the update of the model parameter φ is within a preset threshold value, etc. Can be adopted.

  If it is determined in step S156 that the convergence condition of the model parameter φ is not satisfied, the process returns to step S153, and thereafter, the processes of steps S153 to S156 are repeated.

By repeatedly performing the processing of steps S153 to S156 _, the estimation of the operating state C ^(m) _{t, k} in step S154 and the update of the eigen waveform W ^(m) _k by constrained waveform separation learning in step S155 are alternately performed. The operation state C ^(m) _{t, k} obtained in step S154 and the natural waveform W ^(m) _k (updated value) obtained in step S155 increase in accuracy.

  Thereafter, when it is determined in step S156 that the convergence condition of the model parameter φ is satisfied, the learning process ends.

  In addition, in the state estimation part 12, the operation state of a household appliance can be estimated by the integer programming method and the arbitrary other methods which do not use FHMM.

  [Description of computer to which this technology is applied]

  Next, the series of processes described above can be performed by hardware or software. When a series of processing is performed by software, a program constituting the software is installed in a general-purpose computer or the like.

  Therefore, FIG. 30 illustrates a configuration example of an embodiment of a computer in which a program for executing the series of processes described above is installed.

  The program can be recorded in advance on a hard disk 105 or a ROM 103 as a recording medium built in the computer.

  Alternatively, the program can be stored (recorded) in the removable recording medium 111. Such a removable recording medium 111 can be provided as so-called package software. Here, examples of the removable recording medium 111 include a flexible disk, a CD-ROM (Compact Disc Read Only Memory), an MO (Magneto Optical) disk, a DVD (Digital Versatile Disc), a magnetic disk, and a semiconductor memory.

  In addition to installing the program from the removable recording medium 111 as described above, the program can be downloaded to the computer via a communication network or a broadcast network, and can be installed in the built-in hard disk 105. That is, for example, the program is wirelessly transferred from a download site to a computer via a digital satellite broadcasting artificial satellite, or wired to a computer via a network such as a LAN (Local Area Network) or the Internet. be able to.

  The computer incorporates a CPU (Central Processing Unit) 102, and an input / output interface 110 is connected to the CPU 102 via a bus 101.

  The CPU 102 executes a program stored in a ROM (Read Only Memory) 103 according to a command input by the user by operating the input unit 107 or the like via the input / output interface 110. . Alternatively, the CPU 102 loads a program stored in the hard disk 105 to a RAM (Random Access Memory) 104 and executes it.

  Thus, the CPU 102 performs processing according to the above-described flowchart or processing performed by the configuration of the above-described block diagram. Then, the CPU 102 outputs the processing result as necessary, for example, via the input / output interface 110, from the output unit 106, transmitted from the communication unit 108, and further recorded in the hard disk 105.

  The input unit 107 includes a keyboard, a mouse, a microphone, and the like. The output unit 106 includes an LCD (Liquid Crystal Display), a speaker, and the like.

  Here, in the present specification, the processing performed by the computer according to the program does not necessarily have to be performed in time series in the order described as the flowchart. That is, the processing performed by the computer according to the program includes processing executed in parallel or individually (for example, parallel processing or object processing).

  Further, the program may be processed by one computer (processor) or may be distributedly processed by a plurality of computers. Furthermore, the program may be transferred to a remote computer and executed.

  Furthermore, in this specification, the system means a set of a plurality of components (devices, modules (parts), etc.), and it does not matter whether all the components are in the same housing. Accordingly, a plurality of devices housed in separate housings and connected via a network and a single device housing a plurality of modules in one housing are all systems. .

  The embodiments of the present technology are not limited to the above-described embodiments, and various modifications can be made without departing from the gist of the present technology.

  For example, the present technology can take a configuration of cloud computing in which one function is shared by a plurality of devices via a network and is jointly processed.

  In addition, each step described in the above flowchart can be executed by being shared by a plurality of apparatuses in addition to being executed by one apparatus.

  Further, when a plurality of processes are included in one step, the plurality of processes included in the one step can be executed by being shared by a plurality of apparatuses in addition to being executed by one apparatus.

  DESCRIPTION OF SYMBOLS 11 Data acquisition part, 12 State estimation part, 13 Model memory | storage part, 14 Model learning part, 15 Label acquisition part, 16 Data output part, 21 Evaluation part, 22 Estimation part, 31 Waveform separation learning part, 32 Distributed learning part, 33 State variation learning unit, 42 estimation unit, 51 constrained waveform separation learning unit, 52 individual distributed learning unit, 71 evaluation unit, 72 estimation unit, 81 constrained waveform separation learning unit, 101 bus, 102 CPU, 103 ROM, 104 RAM , 105 hard disk, 106 output unit, 107 input unit, 108 communication unit, 109 drive, 110 input / output interface, 111 removable recording medium

Claims (7)

A state estimation unit that performs state estimation to obtain the speech state of the speaker, using data representing the sum of speech signals superimposed with speech uttered by a plurality of speakers;
A waveform separation learning unit that performs waveform separation learning for obtaining a unique waveform related to speech unique to the speaker , using the state estimation by the state estimation unit to limit frequency components within a frequency band that the speech can take ; A data processing apparatus comprising: The data according to claim 1, wherein the waveform separation learning unit obtains the natural waveform under a base waveform constraint in which the natural waveform is expressed by one or more combinations of a plurality of base waveforms prepared for the speech. Processing equipment. The state estimation unit solves the integer programming problem for obtaining the utterance state that minimizes an error with respect to the data of the weighted addition value of the eigen waveform weighting the utterance state of the plurality of speakers. The data processing apparatus according to claim 1 or 2 , wherein an utterance state is obtained. The waveform separation learning unit solves a quadratic programming problem for obtaining the characteristic waveform that minimizes an error with respect to the data of the weighted addition value of the characteristic waveform weighted by the utterance state of the plurality of speakers. the data processing apparatus according to any one of claims 1 to 3 determine said specific waveform. In the state estimation unit, by solving the integer programming problem for obtaining the utterance state that minimizes an error with respect to the data of the weighted addition value of the eigen waveform weighting the utterance state of the plurality of speakers, the Asking for utterance status,
In the waveform separation learning unit, by solving the quadratic programming problem for obtaining the characteristic waveform that minimizes the error with respect to the data of the weighted addition value of the characteristic waveform weighted by the utterance state of the speaker, wherein the determining the specific waveform data processing apparatus according to any one of claims 1 to 4 carried out alternately. Using data representing the sum of speech signals superimposed with speech uttered by a plurality of speakers, performing state estimation to determine the speech state of the speakers,
A data processing method including a step of performing waveform separation learning for obtaining a specific waveform related to a voice specific to the speaker by using the state estimation to limit frequency components within a frequency band that the voice can take . A state estimation unit that performs state estimation to obtain the speech state of the speaker, using data representing the sum of speech signals superimposed with speech uttered by a plurality of speakers;
A waveform separation learning unit that performs waveform separation learning for obtaining a unique waveform related to speech unique to the speaker , using the state estimation by the state estimation unit to limit frequency components within a frequency band that the speech can take ; Program to make the computer function.

Images