WO2012160062A1

WO2012160062A1 - Transition detection method for automatic-setup non-intrusive appliance load monitoring

Info

Publication number: WO2012160062A1
Application number: PCT/EP2012/059502
Authority: WO
Inventors: Quentin JOSSEN; Frédéric KLOPFERT
Original assignee: Universite Libre De Bruxelles
Priority date: 2011-05-23
Filing date: 2012-05-22
Publication date: 2012-11-29
Also published as: EP2715376A1

Abstract

According to a first aspect, the invention relates to a method (51) for detecting transitions and comprising the steps of: generating a residual signal, α, from a measured signal and providing rules concluding that transitions occur when the residual signal,α, is larger than a threshold value λ. The residual signal, α, has a high amplitude when transitions occur and a low amplitude in the other cases. The method (51) is characterized in that the threshold value λ is automatically defined from local values of the residual signal.

Description

Transition detection method for automatic-setup non-intrusive appliance load monitoring

Field of the invention

[0001] According to a first aspect, the invention relates to a method for detecting in a measured signal transitions that are induced by elements of a physical system. According to a second aspect, the invention relates to an automatic-setup non-intrusive appliance load monitoring method for identifying appliances energy consumption, said method using a step of detecting transitions according to the first aspect of the invention. According to a third aspect, the invention relates to a detector for detecting transitions in a measured signal. According to a fourth aspect, the invention relates to a device for automatic-setup non-intrusive appliance load monitoring.

Description of prior art

[0002] Detecting transitions in signals occupies researchers since the mid 1970, see for instance the article by A. Willsky et al. entitled "A generalized likelihood ratio approach to the detection and estimation of jumps in linear systems", in IEEE Transactions on Automatic Control, vol. 21 , n °1 , p108-1 12 (1976). However, there is neither a unique theory nor a rule of thumb that fits to any application.

[0003] Examples of transitions are ON or OFF switchings of electrical components. A method for detecting transitions aims at providing a binary output that equals one when a transition (or change) occurs and zero when nothing happens. A detection problem is generally divided into two steps, as described in the article by M. Basseville, entitled "Detecting changes in signals and systems-A survey", published in Automatica, vol. 24, n °3, p309-326, 1988. The first step is a generation of a signal to be monitored by detection functions or rules. This signal is often named residual signal. A second step provides these rules applied to the residual signal for detection purposes.

[0004] Transition detection is notably used with non-intrusive appliance load monitoring methods to provide consumers with energy consumption advice. These tools attempt to identify the appliances operating in a house in order to track down energy waste, see for instance WO2009/103998. This patent application proposes a method of inference of appliance from one point of measurement on a supply line. The first step of this method is the identification of events or transitions (step of detecting transitions). In this patent application, these events or transitions represent ON or OFF switchings of an appliance or a component of an appliance. Events are grouped into cycles that include a pair of events. These cycles can then be grouped in clumps and/or clusters that are used for identifying appliances. In such methods, rough signals such as active or reactive power are filtered and then used to create a residual signal that reflects transitions. A global threshold value (rules for detecting transitions) is then experimentally predetermined to detect transitions in the residual signal. Such a global threshold approach nevertheless does not appear to be robust for the variety of data to process as the threshold value is fixed a priori. That means that the results significantly change with varied processed data and/or parameter values.

[0005] In the article "n-lncLOF: a dynamic local outlier detection algorithm for data streams" by Ke Gao et al. in 2010 2^nd international conference on Signal Processing Systems, IEEE, a degree of being an outlier that is named Local Outlier Factor (LOF) is used. Such a method does not use a threshold but rather seeks samples having the N largest LOF in a predetermined temporal interval (or temporal window). This method has the following disadvantages. An initial value of N, n_int, needs to be fixed. One also needs to choose an appropriate size for the predetermined temporal interval (or temporal window). Another drawback of this method is that at least one transition is always detected in the beginning of a signal provided n_int≠0, even if no real transition occurs. This method indeed selects values with the N largest LOF, and it is always possible to define such N values.

[0006] The literature generally separates stochastic from thresholding type solutions for a second step that aims at providing rules applied to the residual signal for detection purposes. However, in stochastic methods such as the ones proposed in the article by M Basseville et al. published in IEEE Trans. on Acoustics, Speech and Signal Processing, vol. 31 , n °3, pp 521 -535, 1983, and entitled "Design and comparative study of some sequential jump detection algorithms for digital signals", a threshold value still has to be defined to take a decision according to a likelihood of change or transition. As stated above, in the patent application WO2009/103998, a global threshold value is also determined a priori. As a consequence, such solutions (stochastic and thresholding type solutions) for this second step need an a priori choice of a threshold value. Such solutions appear not to be robust for a variety of different data to process, namely with respect to signals having various signal to noise ratio's. That means that the results can significantly change with varied processed data and/or parameter values.

[0007] In the article entitled "Design and Comparative Study of Some Sequential Jump Detection Algorithms for Digital Signals" by M Basseville et al. in IEEE Trans on Acoustics, Speech, and Signal Processing, vol. ASSP-31 , No3, 1983, different approaches for detecting transitions are summarized. Classic GLR (p525), modified GLR (p529), as well as the Filtered Derivatives method (p530) use a threshold that is determined a priori resulting in problems of robustness. In page 530 of this article, a method named Hinkley's cumulative sum test is presented. In this method, a residual signal, S_n or U_n, is generated. As shown in figure 14, the residual signal has a high amplitude when a transition occurs and a low amplitude in the other cases. Differences including such residual signals are then compared to a threshold H whose expression is given in equation (20). Such a threshold H is not determined a priori. However, such an expression for the threshold H can lead to not detecting some transitions. The expression of this threshold H does indeed include a variance calculated from a sequence of variables Y_n that constitute the signal whose transitions are to be detected (Y_n is indeed a sequence of variables the mean of which changes from m₀ for n< Θ to m for η≥θ, where Θ is the time of transition to be estimated). Hence, values to detect (the transitions) can be included in the calculation of the threshold H through the variance σ%. In particular, as the variance σ is included in the threshold calculation, a transition of large amplitude in the signal in the vicinity of a transition of low amplitude leads to an increase of the threshold H that can induce not detecting the transition of small amplitude. Therefore, this method named Hinkley's cumulative sum test is not reliable enough. Therefore, there is a need of a method for detecting transitions in a measured signal that is more reliable. Summary of the invention

[0008] According to a first aspect, it is an object of the invention to provide a method for detecting transitions in a measured signal that is more reliable with respect to other known methods. To this end, according to a first aspect, the inventors propose a method for detecting in a measured signal transitions that are induced by elements of a physical system and comprising the steps of:

- generating a residual signal, x, from said measured signal, said residual signal, x, having a high amplitude when transitions occur and a low amplitude in the other cases;

- providing rules that conclude that transitions occur when said residual signal, x, is larger than a threshold value Λ;

and that is characterized in that said threshold value λ is automatically defined from local values of said residual signal, x.

[0009] The method of the invention uses a threshold value λ (or threshold signal) that is automatically defined from local values of the residual signal, x. This residual signal, x, has a high amplitude when transitions in the measured signal occur and a low amplitude in the other cases, ie when there is no transition in the measured signal. The term local means that a threshold value λ corresponding to a given time is determined from values of said residual signal that are neighbour to said given time. Here, the threshold value is not defined from a variance of the signal whose transitions are to be detected. Hence, the method of the invention that is an alternative with respect to Hinkley's cumulative sum test described in Bassevile's article published in IEEE Trans on

Acoustics, Speech, and Signal Processing, vol. ASSP-31 , No3, 1983, is more reliable. With the method of the invention, transitions to detect are not included in the threshold definition. Hence, large transitions do not lead to increasing the threshold value resulting thereafter in not detecting transitions of small amplitudes. Hinkley's cumulative sum test of Bassevile's article does not define a threshold value from local values of a residual signal that has high amplitude when transitions in the measured signal occur and a low amplitude in the other cases. [0010] The method of the invention has other advantages. As a threshold value λ (or threshold signal) is not predetermined, the method of the invention is robust, and in particular more robust than methods where a threshold value is fixed a priori. The inventors propose such an automatic threshold method because a number of transitions is expected to be constant over a range of threshold values. As a consequence, when a number of transitions does not depend on a threshold value, it might be concluded that noise is not responsible for transitions and a threshold value is then considered as optimal. With the method of the invention, one does not need to choose a predetermined threshold value that is not easy to find. Hence, the method of the invention is more simple with respect to methods where a threshold value has to be fixed a priori.

[0011] In the article "n-lncLOF: a dynamic local outlier detection algorithm for data streams" by Ke Gao et al. in 2010 2^nd international conference on Signal Processing Systems, IEEE, no residual signal that has a high amplitude when transitions occur and a low amplitude in the other cases is used. This document neither provides rules concluding that transitions occur when such a residual signal is larger than a threshold value. This document rather selects N samples having the N largest LOF. The method of the invention has not the drawbacks of this method using LOF: no initial value of N, n_int, needs to be determined, no appropriate size for the predetermined temporal interval (or temporal window) has to be chosen and no transition is detected if the residual signal is not larger than the threshold signal.

[0012] Preferably, the threshold value, 1, is defined as a function of a local background noise. This background noise is a background noise of the residual signal as the threshold value λ is automatically defined from local values of the residual signal, x. The inventors propose this preferred embodiment because the amplitude of spikes of a residual signal with respect to amplitude of neighboring noise is an important criterion for detection purposes. As the threshold value is defined as a function of a local background noise, it does not include transitions to be detected. Such a preferred embodiment thus also provides a method that is more reliable with respect to Hinkley's cumulative sum test of Bassevile's article. Preferably, when the measured signal is a discrete time signal, the threshold value corresponding to a time index k, _k, is

i I— ^N \

given by _k = an_k where n_k = -—∑ °° x_P , and where:

- N represents a number of values of the residual signal, x, that are temporal neighbours of a value of said residual signal corresponding to a time index k, x_k ,

- n_k represents a background noise determined by values of said residual signals that are temporal neighbours of said residual signal corresponding to a time index k, x_k

- p is a parameter;

- x_PiU represent samples of the residual signal, x, with values under a p-th percentile defined on a window of interest of size N (which means from these N values of the residual signal, x)

- a is a coefficient higher than or equal to a minimum value, a_min, and

increased from said minimum value, a_min, to an optimal value, a_opt, such that said optimal value, a_opt, corresponds to a value of a that leads to a number of transitions, N_trans _> equal to a number of transitions corresponding to said optimal value of a minus a given value, GV: N_trans ( C — < ₀pt) ^— ^trans (^{a —}

«opt - GV).

By using this definition of the background noise, that is itself used in the determination of the threshold value, Λ, one can control that peaks corresponding to transitions are not included in the determination of the threshold value, Λ, by an appropriate choice of the p parameter . Only values of the residual signal with values under a p-th percentile are indeed considered in this noise expression. Hence, this preferred embodiment is also more reliable with respect to Hinkley's cumulative sum test of Bassevile's article.

Preferably, said minimum value of a is equal to one, c½_in = 1, a is an integer, and the given value is equal to one, GV = 1. More preferably, the optimal value a_opt that leads to a number of transitions equal to a number of transitions corresponding to said optimal value of a minus said given value, N_trans (^ao_Pt) = Ntrans (^ao_Pt ^_ Gv), is a minimum value of a that leads to a number of transitions equal to a number of transitions corresponding to said optimal value of minus said given value, GV.

[0013] Preferably, the method according to the first aspect of the invention comprises a step of filtering the measured signal before generating the residual signal. More preferably, said step of filtering comprises an application of a median filter and an application of a Kalman filter.

[0014] Preferably, the residual signal is a transition likelihood. More preferably, this transition likelihood is given by a ratio between a likelihood of no change and a likelihood of change.

[0015] Preferably, the measured signal is a current waveform. The one skilled in the art generally uses active power for the measured signal when a method for detecting transitions is used for identifying appliances energy consumption. Active power contains information on the amplitude and the phase of the current that is absorbed by components. Nevertheless, components such as motors or devices with an electronic power supply can consume a variable power when they undergo a variable load. The inventors have found that a current waveform is more representative of running components than its amplitude. Even for a fixed set of components drawing power, variations in this power can occur. In order to cope with these fluctuations, the inventors propose to look at a waveform of a current drawn by the components. The idea is that even if the amplitude of the drawn current varies, its waveform should conserve its properties because they are linked to the running components. When the measured signal is a current waveform and when it is a discrete time signal, the residual signal for a time index k, x_k, is preferably given by:

where M is a parameter and i_k a value of current at time index k.

[0016] Preferably, the transitions are on or off switchings of components.

[0017] According to a second aspect, the inventors propose an automatic-setup non-intrusive appliance load monitoring method for identifying appliances energy consumption that is more reliable with respect to other known methods. To this end, this method comprises the steps of:

- detecting transitions in a measured signal;

- characterizing differences between steady states before and after these transitions;

- characterizing transient states located between steady states;

- identifying components;

- identifying appliances from the components identified in the previous step;

- providing appliances energy consumption;

and is characterized in that the step of detecting transitions uses a method according to the first aspect of the invention.

For the same reasons as the ones given for the first aspect of the invention, the method of the invention according to this second aspect is more reliable with respect to other known methods.

[0018] Preferably, the step of identifying appliances uses criteria of:

- identifying groups of components corresponding to appliances;

- identifying sequences of use of components;

- identifying running durations of components;

- identifying times of use of components.

Then, the identification of appliances is independent of appliances programs. Indeed, with the method according to the second aspect of the invention, no predetermined sequence of functioning of components is used for identifying appliances from components. Moreover, the step of characterizing steady states preferably uses spectral properties of the measured signal. This allows one to avoid needing to have pairs of transitions (typically pairs of on and off switchings) for identifying components, contrary to the method proposed in WO2009/103998. Hence, the method of the invention according to the second aspect is more robust with respect to other known methods also for this reason. The impact of not detecting a transition is also less important with the method according to the second aspect of the invention.

[0019] Preferably, the transitions comprise ON or OFF switchings of components and the method further comprises after the step of detecting transitions in a measured signal the following steps: - calculating for each ON switching the following signal :

s(k) = std(p(k: L)), k £ [l, ^L/₂]

where time index k = 1 corresponds to a maximum value of the measured signal in an initial transient state, where time index k = L is an end of an initial steady state, where std designates a standard deviation, and where p k-. L) represents the measured signal between k and L time indices;

- defining a time index corresponding to an end of a transient state, k^*, that maximises a distance between s k) and a straight line passing through s(l) and s(^L/₂).

This allows improving the separation between transient and steady states

[0020] According to a third aspect, the invention relates to a detector for detecting in a measured signal transitions that are induced by elements of a physical system and comprising:

- generation means for generating a residual signal, x, from said measured signal, said residual signal, x, having a high amplitude when transitions occur and a low amplitude in the other cases;

- decision means for providing rules able to conclude that transitions occur when said residual signal, x, is larger than a threshold value Λ; characterized in that

said detector further comprises determination means for automatically defining said threshold value λ from local values of said residual signal, x.

Preferably, the detector is characterized in that said measured signal is a discrete time signal and in that the determination means are able to define the threshold value corresponding to a time index k, A_fe, from the following formula: where nk _E-_N ^u=l ^Xp,u ^>

100

and where :

- N is a number of values of said residual signal, x, that are temporal neighbours of a value of said residual signal corresponding to a time index k, x_k - p is a parameter;

- x_P represent samples of the residual signal, x, with values under a p-th percentile defined on a window of interest of size N;

- n_k represents a background noise determined by values of said residual signal that are temporal neighbours of said residual signal corresponding to a time index k, x_k;

- a is a coefficient higher than or equal to a minimum value, a_min, and increased from said minimum value, a_min, to an optimal value, a_opt, such that said optimal value, a_opt, corresponds to a value of a that leads to a number of transitions, N_trans _> equal to a number of transitions corresponding to said optimal value of a minus a given value, GV\

^trans (^aopt) ⁼ ^trans (^aopt ^— GV) .

[0021] According to a fourth aspect, the invention relates to a device for automatic-setup non-intrusive appliance load monitoring for identifying appliances energy consumption and comprising:

- a detector for detecting transitions in a measured signal;

- means for characterizing differences between steady states before and after these transitions;

- means for characterizing transient states located between steady states;

- means for identifying components;

- means for identifying appliances from the components identified with previous means;

- means for providing appliances energy consumption.

Preferably, the detector is such as described in previous paragraph for two preferred embodiments.

Short description of the drawings

[0022] These and further aspects of the invention will be explained in greater detail by way of example and with reference to the accompanying drawings in which: shows a plot representing a number of detections versus values of a scalar used for defining a threshold value with the method according to a first aspect of the invention;

shows, in a sliding window, a measured active power versus time, a corresponding filtered active power versus time when a mean filter is applied, and a corresponding filtered active power versus time when a median filter is applied;

shows in an upper part a plot of a measured current versus time, and in a lower part, a corresponding residual signal, both plots being shown in an observation window, said residual signal being obtained with a preferred embodiment of the method according to a first aspect of the invention;

shows, in a form of a flowchart, an example of an automated energy auditor;

shows, in a form of a flowchart, three possibilities for identifying appliances consumption;

shows, in a form of a flowchart, different steps of a method according to a second aspect of the invention;

schematically shows a detector for detecting transitions in a measured signal according to a third aspect of the invention;

schematically shows an example of a device according to a fourth aspect of the invention in relation to a display and a sensor.

Detailed description of preferred embodiments

[0023] A method for detecting transitions 51 typically aims at detecting in a single measured signal transitions that are induced by elements of a physical system. Examples of a physical system are: a nuclear power plant, a house, a vehicle. Examples of elements of such a physical system are: the different units of a nuclear power plant, the different appliances or the different components of appliances in a house, the different components of a vehicle. Examples of a measured signal are: temperature of a cooling fluid in a nuclear power plant, a total electric current in a nuclear power plant or in a house, an active or reactive power of a nuclear power plant or of a house, a power of propulsion of a vehicle. Output of such a method for detecting transitions 51 typically equals one when transition occurs and zero when nothing happens. This detection problem has two major steps. First, a signal (named residual signal) to be monitored by detection means is generated from a measured signal. Preferably, the measured signal is first filtered as described below. The residual signal, x, should have high amplitude when transitions occur and low amplitude in the other cases. Second, transitions are detected from said residual signal and decision rules or detection means. The binary output that is equal to one when transitions occur and to zero when there is no transition is named detection signal.

[0024] Various types of residual signals, x, can be generated from a measured signal when one aims at detecting transitions. Different examples are given below in preferred embodiments. For detecting transitions from a residual signal, rules must be defined. Preferably, these rules allow one to decide, at each time, whether a detection function outputs 0 (no transition) or 1 (transition). The inventors propose to use an automatic threshold method for detecting transitions in a residual signal. The inventors propose such a method because the number of transitions is expected to be constant over a range of threshold values, Λ. As a consequence, when the number of transitions, N_trans, does no longer depend on a threshold value λ (or threshold signal) when such a threshold value is increased from a minimum value (zero as an example), it might be concluded that noise is no longer responsible for the transitions and a threshold value (or threshold signal) is then considered as optimal. In other words, if a range of threshold values is found that leads to a stable number of transitions, then these transitions are unlikely to come from noise, and so a value from this range will provide a suitable threshold value. The inventors propose to use a threshold value λ (or threshold signal) that is automatically defined from local values of the residual signal, x. Hence, one could write λ = λ(χ). A threshold value defined according to background noise in the vicinity of spikes is preferably used. Actually, what really matters is the amplitude of spikes with respect to the amplitude of neighboring noise. Preferably, a short time window of interest is used to estimate background noise, and amplitude of the residual signal is compared to this noise. Preferably, the threshold value is defined as a function of this background noise.

[0025] In a preferred embodiment, when the measured signal is a discrete time signal, the threshold value (or threshold signal) corresponding to time index k, _k, is given by equation (Eq. 1 ):

k = a n_k (Eq. 1 ),

where nk — _P__N∑U=I ^xp.u^■ (Eq- 2)

100

N represents a size of a window of interest around a time index k, and preferably, a window of interest centered on a time index k (or, in an equivalent manner, N represents a number of samples or values of the residual signal, x, considered around a time index k). The sum of equation (Eq. 2) starts from u = 1 to u = ^ N The expression ^ N means 'floor function' of a ratio of

.100 .100

p N by 100. If y is a real number, m an integer, and TL a set of integers, the floor function of y can be defined by the following equation: LyJ = max{m £ TL \ ≤ y). As an example, the floor function of the ratio 10 by

10

3 is equal to 3: — = 3. The value of N depends on a time size of the window of interest during which different values of the residual signal x are registered and on a sampling frequency. Hence, the window of interest is a window in which a local background noise n_k is calculated (see equation (Eq. 2)). Preferably, the time size of the window of interest is equal to 200 ms. Preferably, the sampling frequency for generating the residual signal is comprised between 0.1 to 10 points per period and preferably, the period is equal to 0.02 s (50 Hz). More preferably, the sampling frequency for generating the residual signal, x, is equal to 1 point per period. Preferably, N is comprised between 10 and 1000, and more preferably, between 100 and 500. x_v>u represent samples of the residual signal, x, with values under a p-th percentile defined on the window of interest of size N. One definition of percentile is that a p-th percentile (0 < p < 100) of N ordered values (arranged from a smallest to a largest value) is obtained by first calculating an ordinal rank r: r = ^ x N + l/2, (Eq. 3)

rounding the result to the nearest integer, and then taking the value from the ordered values that corresponds to that rank. However, the 100^th percentile is defined as the largest value of the ordered values. For example, from this definition, given the following set of values: 15, 20, 35, 40, 50; the rank of the

30 1

30-th percentile (p = 30) is r =— x 5 + - = 2. Thus, the 30-th percentile is the second number in the sorted list, ie 20. The 40-th percentile would be the third number (since 2.5 rounds up to 3) that is 35. Hence, p entering equation (Eq. 2) is a parameter comprised between 0 and 100. Preferably, p is equal to 40 which means that a 40^th percentile is considered in equation (Eq. 2). The scalar represents an optimal multiple (or factor) of the background noise for defining the threshold value, λ. In the preferred embodiment corresponding to equations (Eq. 1 ) and (Eq. 2), the automatic threshold method searches, preferably in a window of time size comprised between one and ten minutes and more preferably in a window of time size equal to five minutes, a factor a of the locally defined background noise n_k that leads to a stable number of transitions. Hence, contrary to n_k, a is not preferably locally defined. Indeed, a is preferably globally determined, which means that is preferably determined in a window having a longer time size than the window of interest. In a general case, a is a scalar that is not necessarily an integer, a is determined from a minimum value, a_min and increased until an optimal value, a_opt. This optimal value of a leads to threshold values _k that induce a number of transitions equal to the number of transitions induced by threshold values '_k corresponding to an value that is equal to said optimal value minus a given value, GV. So, if the number of transitions is N_trans , this means that : N_trans (a = a_opt) = N_trans {a = a_opt - GV) .

Preferably, the minimum value, a_min, is equal to one, a is an integer, and the given value is equal to one, GV = 1. As a summary, if one plots a histogram representing number of detections versus a values, the optimal value of a, a_opt, corresponds to a stable region in the histogram. More preferably, the optimal value of a, a_opt, is a minimum value of a that leads to a number of transitions equal to a number of transitions corresponding to said optimal value of a minus said given value because over transition numbers are preferred to under transition numbers (it is preferred to detect more transitions than not enough). This is illustrated in figure 1 where the minimum value of a that corresponds to a stable value 70 of the number of transitions is chosen.

[0026] Data from a measurement step 10 such as active power can be very noisy. The variable power drawn by some components is responsible for this noisy behavior which is not linked to measurement noise. This noise is not useful for detection purposes as it is not related to transitions or changes of components. Therefore, in a preferred embodiment, it should be filtered out. Various types of filters can be used for filtering a measured signal. Preferably, the one skilled in the art will use a mean filter and more preferably a median filter. A mean filter replaces a central value of a sliding window with a mean value of samples in this sliding window. Hence, a sliding window is a window into which the filters are applied to data from a measurement step 10 (the measured signal). A mean filter is a weighted sum and requires only few calculations. A disadvantage of a mean filter is its incapability of keeping sharp changes unlike a median filter that filters out noise without distorting shapes of transitions or changes, as it is illustrated in figure 2. A median filter consists in replacing a central value of a sliding window with a median value of samples in this window. A median value is described as a numeric value separating a higher half of a sample, a population, or a probability distribution, from a lower half. A median value of a finite list of numbers can be found by arranging all observations from lowest value to highest value and picking a middle one (i.e a median value of a set of n values is the (— )^th value when the n values are

2 ^J

arranged from the lowest value to highest value). If there is an even number of observations, then there is no single middle value; the median is then usually defined to be the mean of the two middle values. In an embodiment corresponding to a situation where a measured signal is filtered by using a median filter, the size of the sliding window is preferably chosen equal to 140 periods which typically corresponds to 2.8 s.

[0027] After applying a filter such as one of the ones described in the previous section, one skilled in the art can preferably apply a second filter for amplifying transitions. More preferably, a Kalman filter is used for amplifying transitions. Still more preferably, a model such as the one proposed in the article by M Basseville et al. published in IEEE Trans, on Acoustics, Speech and Signal Processing, vol. 31 , n °3, pp 521 -535, 1 983, and entitled "Design and comparative study of some sequential jump detection algorithms for digital signals" is used for such a Kalman filter. In a preferred embodiment, the inventors propose to use such a model for amplifying transitions when a measured signal is filtered before generating a residual signal. In order to detect transitions, a model of a measured signal that has been filtered with a common filter such as a mean filter or a median filter as an example is used to predict values of next samples. The gap between the prediction and the observation is quantified. This gap is an image of the transition that occurred. A Kalman filter allows calculation of such a gap. Such a filter is adaptive as the model is updated according to observations.

[0028] Following the approach proposed in the article by M Basseville et al. and in the article by A. Willsky et al. entitled "A generalized likelihood ratio approach to the detection and estimation of jumps in linear systems" published in I EEE transactions on automatic control, vol. 21 , n °1 , pp 108-1 1 2, 1976, a model corresponding to equations (Eq. 4) and (Eq. 5) is used with such a Kalman filter:

X_k+1 = Φ¾ + W_k (Eq. 4),

y_k = HX_k + e_k (Eq. 5),

where

* ⁼ (Δ) ' ^{Φ =} (1 J) ^{AND W} = (l 0) (Eq. 6).

W_k and e_k are two Gaussian white noise sequences whose variances are parameters to fix. Preferably, these variances are equal to 0.01 . They are usually named state noise and measurement noise. Δ is a slope and μ a mean value, preferably of a filtered active power, and index k relates to a time index. A Kalman filter corresponding to the model related to equations (Eq. 4) and (Eq. 5) is:

X_k+i,-k = Φ¾¾ (Eq. 7),

X_k._k = X_k._k_₁ + K_Yk (Eq. 8),

Yk = y_k - ¾;n (Eq. 9). K is a gain of the Kalman filter. It is designed in such a way that y_k named innovation is a Gaussian sequence with minimum variance. The symbol ^Λ refers to variables calculated via such a Kalman filter and means an estimation of X at time k knowing values of X until time k - 1. The innovation y_k reflects occurrence of transitions and is preferably used after in the step of generating a residual signal as it is described below. Preferably, the adaptive property of the Kalman filter is implemented through update equations as explained in the article by A. Willsky et al. entitled "A generalized likelihood ratio approach to the detection and estimation of jumps in linear systems" published in I EEE transactions on automatic control, vol. 21 , n °1 , pp 1 08-1 1 2, 1 976. Initial condition for solving equations (Eq. 7) to (Eq. 9) is: X_{0 0} = X₀ : initial value estimated by the Kalman filter is assumed to be equal to a first measured sample.

[0029] We now detail the step of generating a residual signal from the measured signal when said measured signal has been filtered according to the preferred embodiment described previously (for instance by applying a median filter followed by a Kalman filter). The inventors propose to use for the residual signal a transition likelihood. The important difference with respect to the state of the art (such as explained in articles by A. Willsky et al. and M. Basseville et al.) is that this transition likelihood is kept for a later decision about an occurrence of transitions of components (in the step of detecting transitions using the automatic threshold method). In stochastic methods such as Generalized Likelihood Ratio methods that are proposed in the articles by A. Willsky et al. and M. Basseville et al., transition likelihood is used in a step that aims at detecting transitions. In such methods, the transition likelihood is compared to a fixed predefined threshold in order to choose between no transition H₀ and transition Such a fixed threshold is not easy to find and not optimal for all data sets. Hence, in these methods, the residual signal is not a transition likelihood but a measured signal that is typically filtered. Various definitions of transition likelihood can be used for defining a residual signal according to a preferred embodiment of the method of the invention. Transition likelihood is a quantity or a function of time that predicts when transitions can occur, or in other words a probability value that predicts when a system passes from a given state to another given state. Preferably such a transition likelihood has high values when transitions occur and low values when no transition occurs. Preferably, when a Kalman filter is used in a preferred step of filtering the measured signal, the transition likelihood is estimated on a basis of a sequence of innovation y_fe. More preferably, the transition likelihood is then evaluated by using a stochastic algorithm that is typically used with Generalized Likelihood Ratio (GLR) methods. Then, the transition likelihood is given by equation (Eq. 10):

where likelihood functions

9 = θ, ν = v) can be calculated under Gaussian assumption because innovation y_k of the Kalman filter is a Gaussian sequence. More details about this method can be found in the article by A. Willsky et al. cited before. L(y_1} ... ,y_fe |H₀) represents likelihood of no transition whereas likelihood of having a transition is given by

= θ, ν = v), with Θ and v being maximum likelihood estimates of time and amplitude of transition respectively.

[0030] In a preferred embodiment, the measured signal that is chosen for detecting transitions is a current waveform. More preferably, the residual signal at time index k, x_k, is then given by:

xk -∑u=k-M (^Ecl- )' where 2M + 1 is a size of an observation window around time index k that is used for defining the residual signal at time index k, x_k (or, in an equivalent manner, 2M + 1 represents a number of samples that are considered for defining a residual signal when equation (Eq. 1 1 ) is used). i_k is a value of electrical current at time index k. The strategy of equation (Eq. 1 1 ) is to compute a normalized difference between successive periods of current. Current waveforms are normalized with respect to their amplitude and possibly with respect to voltage amplitude. Then, the shapes of two waveforms of which the amplitude equals 1 are subtracted. Figure 3 shows, in the lower part, the application of equation (Eq. 1 1 ) to a current depicted in the upper part of figure 3 for M = 3 = 96 (three periods with a sampling frequency of 1600 Hz, and a frequency of the measured signal equal to 50 Hz. From this figure, we see that the residual signal x_k is well synchronized with current variations.

[0031] Preferably, the transitions detected by the method according to the first aspect of the invention correspond to ON and/or OFF switchings of components but in a general case, transitions can correspond to any significant change in energy consumption of components for instance such as power increment of vacuum cleaners. Preferably the method of the first aspect of the invention for detecting in a measured signal transitions 51 that are induced by elements of a physical system is used for identifying appliances energy consumption. In this case, said elements are preferably components of appliances, and said physical system is preferably a house.

[0032] According to a second aspect, the invention relates to an automatic-setup non-intrusive appliance load monitoring method 50 for identifying appliances consumption and comprising the steps of:

a) detecting transitions 51 in a measured signal;

b) characterizing differences between steady states before and after these transitions 52;

c) characterizing transient states located between steady states 53; d) identifying components 54;

e) identifying appliances 55 from the components identified in the previous step 54;

f) providing appliances energy consumption 56;

where the step of detecting transitions 51 uses a method as described above according to the first aspect of the invention.

[0033] Figure 4 shows a global flowchart with four main steps summarizing an example of method for automatically monitor energy consumption. The first step relates to a measurement step 10. In order to give useful consumption information to users, a complete disaggregated energy consumption of an household, i.e. a precise consumption of each appliance is desired. Therefore, the second main step consists in an identification of appliance consumption 20. Then, a third step of advice generation 30 is carried out. Last, information can be sent to a user through a user interface 40. Preferably, data recorded from the measurement step 10 are grid voltage and current consumed by appliances. Preferably, data is acquired with a 1600 Hz sampling frequency on a 16-bits analog-to-digital converter.

[0034] Identification of appliance consumption 20 can be carried out in three different ways as shown in the flowchart of figure 5. These three different ways can be classified in two families: sub-metering 21 or software disaggregation (22 and 50). Sub-metering 21 consists in installing separate meters for each energy-consuming appliances. Software disaggregation uses statistical and/or signal processing software to identify individual appliance consumption within the total consumption measured at one unique central point, such as a main energy meter. This technique is also called Non-Intrusive Appliance Load Monitoring (NIALM). NIALM is complex and can require a manual setup (MS-NIALM) 22 i.e. a learning phase during which users will need to turn each appliance on and off separately. Conversely, automated or automatic-setup monitoring (AS-NIALM) 50 is also possible, though more complex. The method of the invention according to a second aspect is an AS- NIALM method 50.

[0035] Figure 6 is a flowchart showing the different steps of a method according to the second aspect of the invention (AS-NIALM method 50). After acquiring data from the measurements 10, the first step of such a method is to detect significant changes in the electricity consumption pattern (in the field of electricity). The AS-NIALM method 50 can be used in other fields than electricity (for instance gas, fuel or water) though time constants need to be adapted. The detected changes are called transitions. These transitions are assumed to be linked to a switching (on or off) of a component of an appliance, or to a change of running speed of a component. Changes in a measured active power are indeed expected when an electric motor undergoes a greater torque or when it turns faster. Examples of appliances are: fridges, washing machine, television, computer, DVD player, clothes dryer. Examples of components are: motor, resistor, heater, pump. From the detection signal generated by the step of detecting transitions 51 , a measured signal (such as the active power) can be divided into transient and steady states. Both types of states can be mathematically characterized (steps 52 and 53 of figure 6) in order to perform component 54 and, thereafter, appliance 55 identification. For both states, features (such as magnitude of a third harmonic of a measured current for instance) are sought for discerning one component from another.

[0036] Steady states: as an example, active and reactive powers can be used for monitoring appliances consumption. For improving appliances identification, spectral features are preferably used: harmonic magnitude normalised with respect to a fundamental's amplitude and total harmonic distortion for instance. As regards Shannon's theorem, harmonics until 350Hz are preferably considered when a sampling frequency of 1600Hz is used during the measurement step 10. Features describing a current waveform are also preferably used. A ratio of maximal amplitude of the current and its mean value is preferably used to describe a current waveform; and time evolution features such as the time between the zero voltage and the maximum value of the current. Chosen features are considered to be well selected if they lead to separated clusters in a feature space. Features are preferably extracted from differences between steady states, for instance between a present state and a previous state. A cluster is a set of points that are similar and represents a sufficiently dense area in a feature space. Chosen features are then preferably computed on a reconstructed signal as follows: after reception of time indices of successive steady states, differences in the spectral features of both states are computed. The result is the spectrum of the signal drawn only by a new electric component; the latter can be reconstructed and the description is based on this new mathematical signal.

[0037] Transient states: transient states can correspond for instance to on or off switchings. We focus on describing on switchings; the off switchings mainly consist of down steps that contain less useful information. When looking at a typical consumption curve, transient states are reproductive shapes, even if sometimes complicated shapes. Therefore, one can follow the approach proposed in the article by S. Charbonnier et al., entitled "Trends extraction and analysis for complex system monitoring and decision support," in Engineering

Applications of Artificial Intelligence, vol. 18, n °. 1 , p. 21 -36, Feb. 2005. First, the signal within the transient state is decomposed in primitive functions. These primitive functions describe the evolution of the signal; they are constant, growing or decreasing and can be first or second order functions. Second, a letter is associated to each primitive function and the state is represented as a word. Other parameters such as the energy of the signal within the state, the ratio between maximum and mean power values or damping properties could also be used. The features are preferably computed from the shape of the active power.

[0038] After having characterized transient and steady states (steps 52 and 53 of figure 6), the AS-NIALM method 50 carries out a step of identifying components 54. The component identification 54 is a common classification problem; objects have to be classified according to their features. The objects are the signal states and the features are the signal properties within these states, as described in the previous section. The inventors aim at providing a nonintrusive method for identification of appliance consumption; therefore, unsupervised classification techniques are preferably implemented. These methods attempt to identify clusters from points in a feature space. The components are preferably identified with a two steps classification strategy. The first step consists in clustering and classifying the components according to their nature (motor, resistor, heater, television are examples of different natures). The second step classifies the components within their nature cluster; this is the component classification itself. As an example, the second step aims at deducing that a motor that undergoes a transition is a specific motor of 300 W and not a motor of 300 W of another type for instance.

Complementary using steady and transient characteristics

The description of steady and transient states is complementary. The steady state description as well as the description of an ON transient state characterise a component that has been switched ON for instance. Two component identifications are run in parallel, each one considering the features of one type of state. Their output can be compared in order to compute a reliability coefficient; this coefficient is high if the two outputs are similar, in the other case it gets a low value.

Identifying a component nature

Identifying a nature of a component is a classification problem with a predefined number of classes. Therefore, a K-means algorithm is preferably chosen. A K- means method is a clustering technique that partitions n observations into K clusters, where K is a predefined value. It classifies the components according to their nature; there will be as much classes as defined component natures. Three examples of different natures are resistors, motors and electronic devices. Not all features are relevant when looking only at a component nature. Preferably, when searching a nature of a component, one can consider THD (total harmonic distortion), or ratio Q/P (Q = reactive power and P = active power).

Identifying the components within their nature

The second classification step is far different because the number of clusters to identify is generally large and unknown; no prior information is generally available about a number of appliances contained in a house, so no prior information is available concerning the identification of components within their nature. Hence, a DBScan (Density Based Spatial Clustering for Applications with Noise) classification algorithm is preferably used. Looking at the steady states, the components will differ for instance by a magnitude of their consumption from others within a component nature cluster. As far as the transient state features are concerned, two nearly identical but different components could have different transient shapes.

[0039] After identifying components (step 54), one needs to identify appliances (step 55). The inventors propose an original method for identifying the appliances from the components.

Component association: an appliance is considered as a system with one or several components. The previous step, the component identification 54, outputs components (nature and identifier) related to transitions. Therefore, components running together have to be identified as such from the time sequence of all the running components, as they may be running within the same appliance. An appliance is represented as a system of components whose sequences could vary. This would have the advantage of being independent of the appliance program ran by the user. The nature and the sequence of the components identified as running together will help identifying the appliance. Appliance identification (step 55): first, the method of the invention considers sets of components running together; the nature and an identifier are also used for each component. Second, generic information about the typical constitution of the main appliances has been collected; consequently, a generic database can be used to translate the components sets into running appliances. Using such a generic database relies on the three following arguments. First, similar types of appliances of a house involve similar components with typical power ranges. That is all the washing machine and clothes dryers own at least one engine and some resistors; the fridges use an engine as compressor; the lights have resistor or electronic behavior, etc. Second, the usage of the appliances can be generically described; the washing machine is used punctually whereas a fridge runs night and day with a cycling behaviour. Third, information on the time of use is valuable to help distinguishing different appliances using components with similar nature. From the identification of appliances 55, one can finally deduce appliances energy consumption 56. This can be achieved, as an example, by comparing the output of step 55 (appliances identification), and a measured signal such as an active power. As one knows, thanks to step 55, when each appliance is running, one can deduce from a measured active power as an example, the energy consumption of each appliance. This can be easily computed as the relation between appliance and components has been established by step 55 and both the drawn active power of each component and the moment of each transition are known thanks to steps 51 , 52, 54.

[0040] Preferably, after step a) and before steps b) and c) ie after the step of detecting transitions 51 and before the steps of steady state characterization 52 and transient states characterization 53, the following additional step is carried out. An event detection algorithm 51 allows detecting ON or OFF switchings of electrical components for instance. To improve the separation between transient and steady states, the inventors propose for each ON switching a knee in the measured signal that is here denoted by p (the measured signal, p, can be for instance an active power). More specifically, the following signal s(/c) is preferably built where

s(k) = std(p(k: L)), k E [l, ^L/₂] (Eq. 1 1 (a)) The value k = 1 corresponds to a maximum value of the measured signal in an initial transient state. Such an initial transient state is preferably determined from the detection signal of the method for detection transitions 51 that is equal to one when transitions occur and zero when there is no transition. Parameter L is a duration of an initial steady state, std stands for standard deviation, and p(k-. L) represents the measured signal from time index k to time index L. The initial steady state is, as the initial transient state, preferably determined from the detection signal of the method for detection transitions 51 . The end of a transient state is defined as the time k^* that maximizes the distance between s k) and a straight line passing through s(l) and s(^L/₂).

[0041] According to a third aspect, the invention relates to a detector 90 for detecting in a measured signal transitions that are induced by elements of a physical system. Such a detector 90 in relation with a sensor 200 is schematically shown in figure 7. Such a detector 90 comprises generation means 300 for generating a residual signal, x, from a measured signal. The measured signal can originate from a sensor 200 for instance. The residual signal, x, that is generated by the generation means 300 has a high amplitude when transitions occur in the measured signal and a low amplitude in the other cases. The detector 90 also comprises determination means 320 for automatically defining a threshold value λ from local values of the residual signal, x. Hence, the threshold value is written λ(χ) in figure 7. Last, the detector 90 comprises decision means 310 for providing rules able to conclude that transitions occur when the residual signal, x, is larger than the threshold value λ (or l(x)).

[0042] According to a fourth aspect, the invention relates to a device 80 for automatic-setup non-intrusive appliance load monitoring for identifying appliances energy consumption. Such a device 80 is schematically shown in figure 8 in relation with a display 150 and a sensor 200. This device 80 can be a computer and comprises a set of subunits or software modules that implement various steps of the method according to a second aspect of the invention. The computer 80 can be an ordinary, single processor personal computer. The different software modules described below can be included in different computers or different units rather than in a single computer 80. The computer 80 also includes an internal memory (not shown in figure 8) for storing computer program instructions which control how a processing unit within the computer 80 accepts, transforms, and outputs data. The internal memory includes both a volatile and a non-volatile portion. Those skilled in the art will recognize that the internal memory can be supplemented with computer memory media, such as compact disk, flash memory cards, magnetic disc drives. A detector 90 is able to detect transitions from a measured signal. The measured signal is typically provided by a sensor 200, as an example a power meter in the field of electricity. Means 100 and 1 10 are able to characterize steady states and transient states defined by these transitions. Means 120 are able to identify components of appliances from the characterization results provided by means 100 and 1 10. Means 130 can then identify appliances, and finally means 140 can provide appliances energy consumption. Optionally, the result is sent to a display 150 or any other type of man-machine interface.

[0043] We now turn to present some results obtained with the method of the invention for detecting transitions 51 . These results were obtained by using a filtered active power and a current waveform for the measured signal. In both cases, the automatic threshold method that uses equations (Eq. 1 ) and (Eq. 2) was used for detecting transitions. From experimental data traces induced by three types of appliances, 205 transitions where manually identified: twenty in fridge tracks, 138 in washing machine tracks and 47 in clothes dryer tracks. When a current waveform is used for the measured signal and when equation (Eq. 1 1 ) is used for generating a residual signal, the method of the invention leads to 193 correct detections, whereas 1 1 detections are missed and 12 over detections are provided. When an active power is used for the measured signal and when a median filter followed by a Kalman filter are applied with equation (Eq. 10) for providing the residual signal, the method of the invention leads to 192 correct detections, whereas 13 detections are missed and 15 over detections are generated. These results were compared with those obtained with a method such as the one described in the article by A. Willsky et al. entitled "A generalized likelihood ratio approach to the detection and estimation of jumps in linear systems" published in IEEE transactions on automatic control, vol. 21 , n , pp 108-1 12, 1976, applied to the same active power signal that has been filtered by a median filter. Then, the number of detections falls to 178, a number lower than 192, and much less than the correct number of changes: 205. With such a method, there are 24 missed detections and 23 over detections. Hence, the method of the invention for detecting transitions 51 is better.

[0044] The present invention has been described in terms of specific embodiments, which are illustrative of the invention and not to be construed as limiting. More generally, it will be appreciated by persons skilled in the art that the present invention is not limited by what has been particularly shown and/or described hereinabove. The invention resides in each and every novel characteristic feature and each and every combination of characteristic features. Reference numerals in the claims do not limit their protective scope. Use of the verbs "to comprise", "to include", "to be composed of", or any other variant, as well as their respective conjugations, does not exclude the presence of elements other than those stated. Use of the article "a", "an" or "the" preceding an element does not exclude the presence of a plurality of such elements. Also, the method according to a first aspect of the invention is able to detect only one transition. Such a transition can be induced by only one element of a physical system.

[0045] Summarized, the invention may also be described as follows. According to a first aspect, the invention relates to a method 51 for detecting transitions and comprising the steps of: generating a residual signal from a measured signal and providing rules concluding that transitions occur when the residual signal is larger than a threshold value Λ. The residual signal, x, has a high amplitude when transitions occur and a low amplitude in the other cases. The method 51 is characterized in that the threshold value λ is automatically defined from local values of the residual signal, x.

Claims

1 . A method (51 ) for detecting in a measured signal transitions that are induced by elements of a physical system and comprising the steps of:

- generating a residual signal, x, from said measured signal, said residual signal, x, being a signal having a high amplitude when transitions occur and a low amplitude in the other cases;

and characterized in that said threshold value λ is automatically defined from local values of said residual signal, x.

Method (51 ) according to claim 1 characterized in that said threshold value λ is defined as a function of a local background noise.

Method (51 ) according to claim 1 or 2 characterized in that said measured signal is a discrete time signal and in that the threshold value corresponding to a time index k, _k, is given by:

A_k = C L_K

where

and where :

- N is a number of values of said residual signal, x, that are temporal neighbours of a value of said residual signal corresponding to a time index k, x_k

- p is a parameter;

- n_k represents a background noise determined by values of said residual signal that are temporal neighbours of said residual signal corresponding to a time index k, x_k - a is a coefficient higher than or equal to a minimum value, a_min, and increased from said minimum value, a_min, to an optimal value, a_opt, such that said optimal value, a_opt, corresponds to a value of a that leads to a number of transitions, N_trans _> equal to a number of transitions corresponding to said optimal value of a minus a given value, GV\

^trans (^{a =} opt) ⁼ ^trans (^{a =} opt ^~ GV^.

Method (51 ) according to claim 3 characterized in that said minimum value is equal to one, a_min = 1, in that a is an integer, and in that said given value is equal to one, GV = 1.

Method (51 ) according to claim 3 or 4 characterized in that the optimal value, a_opt, that leads to a number of transitions equal to a number of transitions corresponding to said optimal value of a minus said given value, ^wtrans (^ao t) = ^trans (^ao_Pt ^{~ Gv}) > is a minimum value of a that leads to a number of transitions equal to a number of transitions corresponding to said optimal value of a minus said given value, GV.

Method (51 ) according to any of previous claims further comprising a step of filtering said measured signal before generating said residual signal.

Method (51 ) according to claim 6 characterized in that said step of filtering said measured signal comprises an application of a median filter and a Kalman filter.

Method (51 ) according to any of previous claims characterized in that said residual signal is a transition likelihood.

Method (51 ) according to claim 8 characterized in that said transition likelihood is given by a ratio between a likelihood of no change and a likelihood of change.

10. Method (51 ) according to any of claims 1 to 7 characterized in that said measured signal is a current waveform.

1 1 . Method (51 ) according to any of claims any of claim 3 to 5 characterized in that:

- said measured signal is a current waveform; and in that

- said residual signal for a time index k, x_k, is given by:

where M is a parameter and i_k a value of current at time index k.

12. Method (51 ) according to any of previous claims characterized in that said transitions are ON or OFF switchings of components.

13. An automatic-setup non-intrusive appliance load monitoring method (50) for identifying appliances energy consumption and comprising the steps of:

- detecting transitions (51 ) in a measured signal;

- characterizing differences between steady states before and after these transitions (52);

- characterizing transient states located between steady states (53);

- identifying components (54);

- identifying appliances (55) from the components identified in the previous step (54);

- providing appliances energy consumption (56);

characterized in that

the step of detecting transitions (51 ) uses a method according to any of claims 1 to 12.

14. An automatic-setup non-intrusive appliance load monitoring method (50) according to claim 13 characterized in that the step of identifying appliances (55) uses criteria of :

- identifying groups of components corresponding to appliances;

- identifying sequences of use of components;

- identifying running durations of components;

- identifying times of use of components.

15. Automatic-setup non-intrusive appliance load monitoring method (50) according to claim 13 or 14 characterized in that these transitions comprise ON or OFF switchings of components and in that the method further comprises after the step of detecting transitions (51 ) in a measured signal the following steps:

- calculating for each ON switching the following signal :

s(k) = std(p(k: L)), k £ [l, ^L/₂]

16. Detector (90) for detecting in a measured signal transitions that are induced by elements of a physical system and comprising:

- generation means (300) for generating a residual signal, x, from said measured signal, said residual signal, x, having a high amplitude when transitions occur and a low amplitude in the other cases;

- decision means (310) for providing rules able to conclude that transitions occur when said residual signal, x, is larger than a threshold value Λ; characterized in that said detector (90) further comprises determination means (320) for automatically defining said threshold value λ from local values of said residual signal, x.

17. Detector (90) according to claim 16 characterized in that said measured signal is a discrete time signal and in that the determination means (320) are able to define the threshold value corresponding to a time index k, _k, from the following formula:

A_k = C L_k

where

∑LlOO

n_k = -N ' u=l

100

and where :

- p is a parameter;

- n_k represents a background noise determined by values of said residual signal that are temporal neighbours of said residual signal corresponding to a time index k, x_k

^trans (^aopt) ⁼ ^trans (^aopt ^— GV) .

18. Device (80) for automatic-setup non-intrusive appliance load monitoring for identifying appliances energy consumption and comprising:

- a detector (90) for detecting transitions in a measured signal; - means (100) for characterizing differences between steady states before and after these transitions;

- means (1 10) for characterizing transient states located between steady states;

- means (120) for identifying components;

- means (130) for identifying appliances from the components identified with means (120);

- means (140) for providing appliances energy consumption. 19. Device (80) for automatic-setup non-intrusive appliance load monitoring for identifying appliances energy consumption according to claim 18 characterized in that the detector (90) is one of the detectors (90) of claim 16 or 17.