JP7384265B1 - Driving support device, driving support method and program - Google Patents

Abstract

An object of the present invention is to provide a technology that can present various near-optimal manipulated variable series.
[Solution] A driving support device according to an aspect of the present disclosure includes feature quantities of manipulated variables in a past time interval, feature quantities of state variables in the past time interval, and features of environmental variables in the past time interval. a learning unit configured to learn a prediction model that predicts the feature amount of the state variable from the feature amount of the manipulated variable and the feature amount of the environment variable using the amount; The prediction model is configured to calculate an optimal feature amount of the manipulated variable in the future time interval using a feature amount of the environmental variable and a target feature amount of the state variable in the future time interval. a time series configured to extract a plurality of time series data of the manipulated variable in the future time interval using a plurality of features at least similar to the optimal feature of the manipulated variable; It has an extraction part.
[Selection diagram] Figure 2

Description

The present disclosure relates to a driving support device, a driving support method, and a program.

In recent years, due to the shortage of skilled operators in the industrial field, there has been an increasing need for plant operation support systems that utilize AI (Artificial Intelligence), machine learning technology, etc. In general, in plants such as power plants, steel plants, chemical plants, etc., operators use past knowledge to plan optimal operation amounts based on past knowledge so that state variables that contribute to future product quality reach target values, depending on the environment at the time. After that, the plant is operated based on the manipulated variables. In response, plant operation support systems that utilize prediction techniques and the like have been proposed in the past (for example, Patent Documents 1 to 4).

Japanese Patent Application Publication No. 2012-74007 Japanese Patent Application Publication No. 2006-120131 Japanese Patent Application Publication No. 2020-140253 Patent No. 7081728

However, conventionally, when presenting a time series of manipulated variables in a future time interval to an operator or the like, it has not been possible to present a time series of various manipulated variables close to the optimum. Hereinafter, the time series of manipulated variables will also be referred to as a "manipulated variable series."

For example, in Patent Document 1, since time series data cannot be handled, a manipulated variable series cannot be obtained. Further, for example, in Patent Document 2, an optimization technique is not used, and an optimal manipulated variable series is not necessarily obtained. Further, for example, in Patent Document 3, when obtaining a time series of driving indicators for a certain period in the future, the driving indicators are calculated using an optimization technique at each time in the period, and the time series of the optimal driving indicators is calculated. can be presented, but it is not possible to present a variety of manipulated variable series. Furthermore, for example, in Patent Document 4, although an optimal manipulated variable series can be presented, it is not possible to present various manipulated variable series.

The present disclosure has been made in view of the above points, and provides a technology that can present various near-optimal manipulated variable series.

A driving support device according to an aspect of the present disclosure uses a feature amount of a manipulated variable in a past time interval, a feature amount of a state variable in the past time interval, and a feature amount of an environment variable in the past time interval. a learning unit configured to learn a prediction model that predicts the feature amount of the state variable from the feature amount of the manipulated variable and the feature amount of the environmental variable; and a target feature amount of the state variable in the future time interval, the optimization unit is configured to calculate an optimal feature amount of the manipulated variable in the future time interval using the prediction model. and a time series extraction unit configured to extract a plurality of time series data of the manipulated variable in the future time interval using a plurality of feature quantities that are at least similar to the optimal feature quantity of the manipulated variable; has.

A technique is provided that can present various near-optimal manipulated variable series.

1 is a diagram illustrating an example of a hardware configuration of a driving support device according to an embodiment. 1 is a diagram illustrating an example of a functional configuration of a driving support device according to an embodiment. 7 is a flowchart illustrating an example of data collection processing according to the present embodiment. 7 is a flowchart illustrating an example of model learning processing according to the present embodiment. It is a flowchart which shows an example of operation amount sequence output processing concerning this embodiment. It is a flowchart which shows an example of operation amount series extraction processing (Example 1) concerning this embodiment. FIG. 3 is a diagram illustrating an example of extraction of a neighborhood set. FIG. 3 is a diagram illustrating an example of clustering. FIG. 7 is a diagram illustrating an example of actual operation feature amounts to be extracted for each cluster. It is a flowchart which shows an example of operation amount series extraction processing (Example 2) concerning this embodiment. FIG. 7 is a diagram illustrating an example of tagging performance operation feature amounts. FIG. 7 is a diagram illustrating an example of actual operation feature quantities to be extracted for each tag.

An embodiment of the present invention will be described below. In the following embodiments, various manipulated variable series that bring the state variable value of the target closer to the target value, such as a plant, equipment, equipment, device, etc., and that are close to the optimal manipulated variable series A driving support device 10 that can support the driving of the target by presenting the information to an operator or the like will be described. Here, the driving support device 10 according to the present embodiment includes (1) a data collection phase in which data is collected for learning a prediction model, (2) a model learning phase in which the prediction model is learned, and (3) a prediction model. There are three phases: the driving support phase, which uses optimization technology to output a variety of manipulated variable series that are close to the optimum as the manipulated variable series for the target.

Note that the target refers to a plant, equipment, equipment, device, etc. operated by an operator or the like. Various sensors are installed on the target, and these sensors measure the state and environment of the target, operations on the target, and the like. Specific examples of the state include temperature, pressure, flow rate, etc. at a certain point in time. Specific examples of the environment include, for example, the outside temperature at a certain point in time, the equipment number, and the like. Specific examples of the operations include, for example, the amount of opening and closing of a target air valve, fuel valve, etc., the amount of oxygen controlled by opening and closing of the air valve, and the like. Variables that represent such states are called state variables, variables that represent the environment are called environment variables, and variables that represent operations are called operation variables.

Hereinafter, the total number of operation variables is I, the total number of state variables is J, and the total number of environment variables is K. In addition, it is assumed that the manipulated variable value (the manipulated variable value is also called the manipulated variable) is obtained as time series data, and that i (1≦i≦I) in a certain time interval T = [t _s , t _e ] Let X _i ={X _it |t∈T} be the time series data of the th manipulated variable value. Here, t _s is the start time of time interval T, _te is the end time of time interval T, and X _it is the manipulated variable value at time t. Note that since the time series data of the manipulated variable values represents the time series of the manipulated variables, the time series data of the manipulated variable values will also be referred to as the "manipulated variable series" below. On the other hand, it is assumed that the state variable value is obtained as feature data in the time interval, and the j-th (1≦j≦J) state variable value in the time interval T is y _j . Similarly, it is assumed that the environmental variable value is obtained as feature amount data in that time interval, and the k-th (1≦k≦K) environmental variable value in the time interval T is _zk .

The above time interval T may be any predetermined time interval, but for example, if the target is a batch plant, T may be the time interval of one batch. In addition to this, for example, T can be a time interval in which one product is manufactured, a time interval representing one work process, a time interval used as a unit of analysis, etc. Below, when considering a plurality of time intervals T, the m-th time interval is expressed as T ^(m) . In addition, the time series data of the i-th manipulated variable value in the m-th time interval T ^(m) is expressed as X _i ^(m) = {X _it | t∈T ^(m) }, and the j-th state variable value is expressed as y _j ^(m) , and the k-th environment variable value is expressed as z _k ^(m) .

For example, assuming a steel plant as the target, the m-th time interval T ^(m) is the m-th batch, X _i ^(m) is the time series data of the oxygen amount in the m-th batch, and y _j ^(m) is the time series data of the oxygen amount in the m-th batch. It is conceivable that the molten steel temperature of the m-th batch, z _k ^(m), is the furnace number of the m-th batch, etc.

<Example of hardware configuration of driving support device 10>
FIG. 1 shows an example of the hardware configuration of a driving support device 10 according to this embodiment. As shown in FIG. 1, the driving support device 10 according to the present embodiment includes an input device 101, a display device 102, an external I/F 103, a communication I/F 104, a RAM (Random Access Memory) 105, and a ROM. (Read Only Memory) 106, an auxiliary storage device 107, and a processor 108. Each of these pieces of hardware is communicably connected via a bus 109.

The input device 101 is, for example, a keyboard, a mouse, a touch panel, a physical button, or the like. The display device 102 is, for example, a display, a display panel, or the like. Note that the driving support device 10 may not include at least one of the input device 101 and the display device 102, for example.

The external I/F 103 is an interface with an external device such as the recording medium 103a. Examples of the recording medium 103a include a CD (Compact Disc), a DVD (Digital Versatile Disk), an SD memory card (Secure Digital memory card), and a USB (Universal Serial Bus) memory card.

Communication I/F 104 is an interface for connecting driving support device 10 to a communication network. The RAM 105 is a volatile semiconductor memory (storage device) that temporarily holds programs and data. The ROM 106 is a nonvolatile semiconductor memory (storage device) that can retain programs and data even when the power is turned off. The auxiliary storage device 107 is, for example, a nonvolatile storage device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive), and stores programs and data. The processor 108 is, for example, various arithmetic devices such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit).

Note that the hardware configuration shown in FIG. 1 is an example, and the driving support device 10 may have other hardware configurations. For example, the driving support device 10 may include multiple auxiliary storage devices 107 and multiple processors 108, or may include various hardware other than the illustrated hardware.

<Example of functional configuration of driving support device 10>
FIG. 2 shows an example of the functional configuration of the driving support device 10 according to this embodiment. As shown in FIG. 2, the driving support device 10 according to the present embodiment includes an input section 201, a feature value conversion section 202, a model learning section 203, an optimization section 204, a time series extraction section 205, and an output section 201. 206. Each of these units is realized, for example, by one or more programs installed in the driving support device 10 causing the processor 108 or the like to execute the process. Further, the driving support device 10 according to the present embodiment includes a storage unit 207. The storage unit 207 is realized by, for example, the auxiliary storage device 107 or the like. Note that the storage unit 207 may be realized by, for example, a storage device such as a database server connected to the driving support device 10 via a communication network.

In the data collection phase, the input unit 201 acquires a manipulated variable series, a state variable value, and an environment variable value in a past time interval from the target.

In addition, the input unit 201 acquires data used for learning the predictive model (learning data to be described later) from the storage unit 207 in the model learning phase.

Furthermore, in the driving support phase, the input unit 201 acquires from the storage unit 207 environmental variable values and target values of state variables in a future time period in which it is desired to present various near-optimal manipulated variable sequences.

The feature amount conversion unit 202 converts the operation amount series into feature amount data in the data collection phase. Specifically, the feature amount conversion unit 202 transforms the operation amount series X _i ^(m) (i=1,...,I) in the m-th time interval T ^(m) into feature amount data x _i ^(m) (i=1,...,I). Here, x _i ^(m) represents a feature quantity of X _i ^(m) , and is, for example, a statistical quantity such as an average value or a total value of X _it (t∈T ^(m) ). The manipulated variable series X _i ^(m) (i=1,...,I) in the m-th time interval T ^(m ) and its feature data x _i ^(m ) and (i=1,..., I ), the state variable value y _j ^(m) (j=1,..., J), and the environment variable value z _k ^(m) (k=1,..., K) are associated at the m-th time. This is stored in the storage unit 207 as performance data in the section T ^(m) . That is, the actual data in the m-th time interval T ^(m) is (X ₁ ^(m) , ..., X _I ^(m) , x ₁ (m ⁾ , ..., x _I ^(m) , y ₁ ^(m) ,..., _yJ ^(m) , _z1 ^(m) ,..., _zK ^(m) ). Hereinafter, the actual data in the m-th time interval T ^(m) will also be referred to as the m-th actual data. Furthermore, hereinafter, the feature amount data of the time series data of manipulated variables (manipulated amount series) will also be referred to as "operated feature amount."

In the model learning phase, the model learning unit 203 uses the learning data to learn a prediction model that predicts the state variable value from the operation feature amount and the environmental variable value. That is, the model learning unit 203 calculates the state variable by y _j = f _j (x ₁ , ..., x _I , z ₁ , ..., z _K ) for each j = 1, ..., J. A prediction model f=(f ₁ , . . . , f _J ) that predicts the value is learned. Such a predictive model may be learned by, for example, a method such as DL (Deep Learning) or PLS (Partial Least Squares Regression) using a neural network or the like. Here, the learning data is data obtained by extracting operation features, state variable values, and environment variable values from actual data, and the m-th learning data is (x ₁ ^(m) , ..., x _I ^{( m)} , y ₁ ^(m) , ..., y _J ^(m) , z ₁ ^(m) , ..., z _K ^(m) ).

In the driving support phase, the optimization unit 204 uses the prediction model f to determine the optimal operation feature amount (hereinafter also referred to as the optimal operation feature amount) from the environmental variable value and the target value of the state variable in a future time interval. calculate. That is, the optimization unit 204 sets the environment variable value in the future time interval to z' _k and the target value of the state variable in the time interval to y' _j , and calculates f _j (x ₁ , ..., x _I , z ' ₁ , . . . , z' _K ) are close to y' _j , so that the operation feature quantities x ₁ , . . . , x _I are calculated as the optimal operation feature quantities. More specifically, the optimization unit 204 calculates |f _j (x ₁ , ..., x _I , z' ₁ , ..., z' _K ) for j=1, ..., J, for example. -y' _j | (that is, the sum of absolute errors regarding j=1, . . . , J) is calculated as the optimal manipulation feature amount x ₁ , ..., x _I do. This can be calculated by, for example, general mathematical programming or metaheuristics. Note that the absolute error is just an example, and is not limited to this; for example, a squared error or the like may be used.

The time series extraction unit 205 uses the optimal operation feature amount calculated by the optimization unit 204 to generate a plurality of operation amount series (that is, various operation amount sequences close to the optimum) to be presented to the operator etc. Extract from a set of data.

The output unit 206 outputs the plurality of manipulated variable series extracted by the time series extraction unit 205 to any predetermined output destination (for example, the display device 102, the storage unit 207, a terminal used by an operator, etc.). . Hereinafter, as an example, it is assumed that the output destination of the output unit 206 is a terminal used by an operator or the like. Thereby, for example, an operator or the like can obtain various manipulated variable sequences that are close to the optimum for the time interval (that is, a plurality of manipulated variable sequences that bring the target state variable value closer to the target value). Therefore, an operator etc., for example, selects a reasonable manipulated variable series from these multiple manipulated variable series as candidates in light of his own empirical rules, etc., and then operates the target using the selected manipulated variable series. It is possible to formulate operation plans for

The storage unit 207 stores performance data collected in the data collection phase. The storage unit 207 also stores the prediction model f learned in the model learning phase, target values of state variables used in the driving support phase, and the like. In addition to these, the storage unit 207 may store various data (for example, intermediate calculation results when learning the prediction model f, intermediate calculation results when calculating the optimal operation feature amount, etc.). good.

<Data collection processing>
The data collection process for collecting data for learning a predictive model in the data collection phase will be described below with reference to FIG. 3. Steps S101 to S103 in FIG. 3 are repeatedly executed, for example, for each mεM (where M is a set of numbers of time intervals for which data is to be collected). In the following, a case will be described in which data for a certain m-th time interval is collected.

The input unit 201 inputs a manipulated variable series X _i ^(m) (i=1,...,I) and a state variable value y _j ^(m) (j=1,..., J) in the m-th time interval. and the environmental variable value z _k ^(m) (k=1, . . . , K) are acquired from the target (step S101).

Next, the feature amount conversion unit 202 converts the manipulated variable series X _i ^(m) (i=1,...,I) into the manipulated feature amount x _i ^(m) (i=1,..., I). Convert (step S102).

Then, the input unit 201 inputs (X ₁ ^(m) , ..., X _I ^(m) , x ₁ ^(m) , ..., x _I ^(m) , y ₁ ^(m) , ..., y _J ^(m) , z ₁ ^(m) , . . . , z _K ^(m) ) is stored in the storage unit 207 as m-th performance data (step S103).

By repeatedly executing the above steps S101 to S103 for each m∈M, the actual data set {(X ₁ ^(m) ,...,X _I ^(m) ,x ₁ ^(m) ,... . . , x _I ^(m) , y ₁ ^(m) , . . . , y _J ^(m) , z ₁ ^(m) , . . . , z _K ^(m) ) |m∈M} are obtained. Note that each operation feature quantity x _i ^(m) , each state variable value y _j ^(m) , and each environment variable value z _k ^(m) may be normalized to a value of 0 or more and 1 or less, for example. good.

<Model learning processing>
The model learning process for learning a predictive model in the model learning phase will be described below with reference to FIG. 4.

The input unit 201 acquires learning data from the storage unit 207 (step S201). That is, the input unit 201 inputs the learning data set {(x ₁ ^(m) , . . . , x _I ^(m) , y ₁ ^(m) , . . . , y _J ^(m) , z ₁ ^(m) ,...,z _K ^(m) ) |mεM ₁ } is acquired from the storage unit 207. Here, M ₁ is a set of numbers of performance data corresponding to learning data, and M ₁ ⊂M. Note that M ₁ =M may be satisfied.

Next, the model learning unit 203 learns the prediction model f using each learning data included in the learning data set (step S202). That is, the model learning unit 203 calculates that f _j (x ₁ ^(m) , ..., x _I ^(m) , z ₁ ^(m) , ..., z _K ^(m) ) is different from y _j ^(m). A prediction model f=(f ₁ , . . . , f _J ) is learned to make predictions with high accuracy.

Then, the model learning unit 203 stores the trained prediction model f in the storage unit 207 (step S203).

<Operation amount series output processing>
Hereinafter, the operation amount series output process of outputting various nearly optimal operation amount series and presenting them to the operator etc. in the driving support phase will be described with reference to FIG. 5. In the following, a case will be described in which various near-optimal manipulated variable sequences for a certain time interval T in the future are output.

First, the input unit 201 inputs the environmental variable value z' _k (k=1, . . . , K) and the state variable target value y' _j (j=1, . . . , J) in the time interval T. It is acquired from the storage unit 207 (step S301). Note that, for example, the operating feature quantity x _i ^(m) , the state variable value y _j ^(m) , and the environmental variable value z _k ^(m) included in each performance data are normalized to a value of 0 or more and 1 or less. , the environment variable value z' _k (k=1,...,K) and the state variable target value y' _j (j=1,..., J) are similarly normalized. shall be taken as a thing.

Next, the optimization unit 204 uses the learned prediction model f to calculate the optimal operation feature amount for the time interval T (step S302). For example, the optimization unit 204 calculates |f _j (x ₁ , ..., x _I , z' ₁ , ..., z' _K )−y' _j | for j=1, ..., J. The operation feature quantities x ₁ , . . . , x _I that minimize the sum are calculated as the optimum operation feature quantities. Hereinafter, the optimal operation feature amount calculated in this step is assumed to be (x' ₁ , . . . , x' _I ).

Next, the time series extraction unit 205 uses the optimal operation features (x' ₁ , _. . . X ₁ ^(m) , ..., X _I ^(m) , x ₁ ^(m) , ..., x _I (m ⁾ , y ₁ ^(m) , ..., y _J ^(m) , z ₁ ^(m) ,...,z _K ^(m) ) |m∈M} (step S303). Note that details of the process of this step (operated amount series extraction process) will be described later. Hereinafter, the plurality of manipulated variable sequences extracted in this step are assumed to be {(X ₁ ^(m) , . . . , X _I ^(m) )|m∈M ₂ }. Here, M ₂ is a set of numbers of performance data including the manipulated variable series extracted in this step, and M ₂ ⊂M.

Then, the output unit 206 allows an operator etc. to use the plurality of manipulated variable series {(X ₁ ^(m) , ..., X _I ^(m) ) | m∈M ₂ } extracted in step S303 above. output to the terminal (for example, display on the display of the terminal) (step S304). Thereby, the operator can obtain various manipulated variable sequences {(X ₁ ^(m) , ..., X _I ^(m) ) | m∈M ₂ } that are close to the optimum in the time interval T, for example. , it becomes possible to drive the object or formulate an operation plan for the object according to the operation amount series selected from among them.

<Manipulated variable series extraction process>
As an example of the manipulated variable series extraction process in step S303 of FIG. 5, Example 1 and Example 2 will be described below. Hereinafter, the operation amount series and the operation feature amount included in each performance data that constitute the performance data set will also be referred to as the "actual operation amount series" and the "actual operation feature amount", respectively.

≪Manipulated variable series extraction processing (Example 1)≫
The manipulated variable sequence extraction process in the first embodiment will be described with reference to FIG. 6.

First, the _time series extraction unit 205 selects the optimal operation feature amount (x _{' 1} ^, . . . ) from the actual operation feature amount (x ₁ ⁽ m) , . , x' _I ) is extracted as a neighborhood set (step S401).

For example, the time series extraction unit 205 selects N pieces in descending order of the sum of |x' _i −x _i ^(m) | for i=1,...,I, where N is a predetermined integer of 2 or more. Extract the actual operation features (x ₁ ^(m) ,..., x _I ^(m) ), and extract these N actual operation features (x ₁ ^(m) ,..., x _I ^(m) ) Let the set of be the neighborhood set. Hereinafter, the neighborhood set will be denoted by R. As an example, the neighborhood set R when I=2 and N=12 is shown in FIG. In the example shown in FIG. 7, the actual operation feature amount included in the portion surrounded by the broken line 1000 is the actual operation feature amount that belongs to the neighborhood set R.

Note that the method for extracting the neighborhood set R described above is one example, and is not limited to this. For example, the time series extraction unit 205 calculates |f _j (x' ₁ , . . . , x' _I , z' ₁ , . . . , z' _K ) - y _j for j=1, . ^(m) The sum of | and |x' _i -x _i ^(m) | for i=1,...,I and the sum of |z' _k -z _k for k=1,...,K ^(m) | Extract N actual operation features (x ₁ ^(m) , ..., x _I ^(m) ) in descending order of the total sum with the sum of | ₁ ^(m) ,..., _xI ^(m) ) may be set as the neighborhood set R.

Further, for example, the neighborhood set R may be extracted from the actual operation feature values (x ₁ ^(m) , . . . , x _I ^(m) ) that satisfy predetermined constraint conditions. As such a constraint condition, for example, a certain actual operation feature amount x _i ^(m) is smaller (or larger) than the optimal operation feature amount x′ _i .

Next, the time series extraction unit 205 clusters the actual operation features (x ₁ ^(m) , ..., x _I ^(m) ) included in the neighborhood set R and divides them into L clusters (step S402). Also, at this time, the time series extraction unit 205 calculates representative points of each cluster. Note that as a clustering method, for example, the k-means method may be used. Hereinafter, the representative point of each cluster is assumed to be μ ^(l) = (μ ₁ ^(l) , . . . , μ _I ^(l) ) (l=1, . . . , L). Further, a set of actual operation feature amounts (x ₁ ^(m) , . . . , x _I ^(m) ) belonging to the l-th cluster is defined as V _l . As an example, FIG. 8 shows the result of clustering the neighborhood set R indicated by the broken line 1000 in FIG. The example shown in FIG. 8 shows a case where L=3, and the actual operation feature amounts included in the portions surrounded by broken lines 1100, 1200, and 1300 are the actual operation feature amounts that belong to each cluster. In the example shown in FIG. 8, representative points of each cluster are also shown.

Next, for _each cluster, the time series extraction unit 205 selects ^the _{representative} point μ _l ⁽ The actual operation feature amount (x ₁ ^(m) , . . . , x _I ^(m) ) closest to l=1, . . . , L) is extracted (step S403).

For example, the time series _extraction unit ₂₀₅ extracts i=1, . . . out of (x ₁ ^{(m )} , . . . , the actual operation feature quantity (x ₁ ^{(m) ,} . . . , x _I (m)) with the smallest sum of |μ _i ^(l ) −x i ^{(m) |} _regarding I is extracted. As a result, L actual operation feature quantities having diversity within a range close to the optimal operation feature quantity (x' ₁ , . . . , x' _I ) are extracted. As an example, FIG. 9 shows actual operation feature amounts extracted from each cluster indicated by broken lines 1100 to 1300 in FIG. As shown in FIG. 9, the actual operation feature closest to the representative point of each cluster is the extraction target.

Note that the method for extracting the actual operation feature amount described above is an example, and is not limited to this. For example, the time series _extraction unit ₂₀₅ extracts j=1, . . . of (x ₁ ^{(m )} , . ..., the sum of |f _j (x' ₁ , ..., x' _I , z' ₁ , ..., z' _K ) - y _j ^(m) | for J, and i=1, ...・, the sum of |μ _i ^(l) −x _i ^(m) | for I and the sum of |z' _k −z _k ^(m) | for k=1, ..., K is the smallest Actual operation feature amounts (x ₁ ^(m) , . . . , x _I ^(m) ) may be extracted. Alternatively, for example, the time series extraction unit 205 extracts j=1 among (x ₁ ^(m) , ..., x _I ^(m) )∈V _l for l=1, ..., L. , ..., the sum of |f _j (μ ₁ ^(l) , ..., μ _I ^(l) , z' ₁ , ..., z' _K ) - y _j ^(m) | for J, and The sum of |μ _i ^(l) −x _i ^(m) | for i=1, ..., I and the sum of |z' _k -z _k ^(m) | for k=1, ..., K The actual operation feature amount (x ₁ ^(m) , . . . , x _I ^(m) ) having the smallest total sum may be extracted.

Then, the time series extraction unit 205 extracts L actual operation amount series corresponding to the L actual operation feature quantities extracted in step S403 above from the actual data set as operation amount series to be presented ( Step S404). As a result, L manipulated variable series {(X ₁ ^(m) , ..., X _I ^(m) ) | m∈M ₂ } (where, |M ₂ |=L) is obtained.

≪Manipulated variable series extraction processing (Example 2)≫
The manipulated variable sequence extraction process in the second embodiment will be described with reference to FIG. 10.

First, the time series extraction unit 205 attaches a tag to each performance data included in the performance data set (step S501). Here, the tag is identification information for identifying performance data based on some criteria. For example, it is possible to use an operator name, operator ID, etc. as a tag. In this case, the time series extraction unit 205 sets (adds) to each performance data a tag indicating the operator who was performing the operation when the operation amount series included in the performance data was obtained. At this time, the time series extraction unit 205 identifies the operator who was performing the operation when the operation amount series included in each performance data was obtained, for example, by referring to the operator's work information, login information, etc. do it. As an example, when I=2 and any one of the following tags: tag A indicating operator A, tag B indicating operator B, and tag C indicating operator C is attached to the record data, the actual operation feature amount is shown in FIG. Shown below. In the example shown in FIG. 11, the actual operation feature amount included in the actual performance data to which the tag A is attached, the actual operation feature amount included in the actual achievement data to which the tag B is attached, and the actual performance data to which the tag C is attached The included actual operation feature amounts and optimal operation feature amounts are shown.

Next, the time series extraction unit 205 selects a plurality of tags to be extracted from among the tags assigned in step S501 above (step S502). Note that the time series extraction unit 205 may select all of the tags assigned in step S501 above as extraction targets, or may select only some tags as extraction targets. In the following, the number of tags selected in this step is assumed to be W.

Next, for each tag selected in step S502 above, the time series extraction unit 205 extracts the actual operation feature amount closest to the optimal operation feature amount from among the actual operation feature amounts included in the performance data to which the tag is attached. is extracted (step S503).

For example, if the set of performance operation features included in performance data to which tag r is attached is _Vr , the time series extraction unit 205 extracts |x for i=1, . . . , I for each tag r. It is sufficient to extract the actual operation feature quantity ( _{x 1} ^{(m) ,} . . . , x _I ^(m) )∈V _{r for} which the sum of ' i −x i (m) | is the smallest. As an example, in the example shown in FIG. 11, when tag A, tag B, and tag C are selected in step S502 described above, the actual operation feature amounts to be extracted are shown in FIG. As shown in FIG. 12, for each tag, the actual operation feature closest to the optimal operation feature is the extraction target. As a result, for example, when the tag indicates an operator name, an operator ID, etc., W various performance operation feature amounts are extracted in consideration of the preferences of W operators.

Note that the method for extracting the actual operation feature amount described above is an example, and is not limited to this. For example, the time series extraction unit 205 extracts |f _j (x' ₁ , . . . , x' _I , z' 1 , . _. . , z' for j=1, . . . , J for each tag r. The _sum of _K ) _{−y j} ^{(m) |} and the sum ^of |μ The actual operation feature amount (x ₁ ^{(m) ,} . . . , x _I ^(m) ₎ ∈V _r that has the smallest sum with the sum of |z′ k −z k (m) _| may be extracted.

Then, the time series extraction unit 205 extracts W actual operation amount series corresponding to the W actual operation feature quantities extracted in step S503 above from the actual data set as operation amount series to be presented ( Step S504). As a result, W manipulated variable sequences having diversity within a range close to the optimal manipulated variable sequence {(X ₁ ^(m) , ..., X _I ^(m) ) | m∈M ₂ } (however, |M ₂ |=W) is obtained.

<Summary>
As described above, the driving support device 10 according to the present embodiment learns a prediction model that predicts a state variable value from the feature amount of the manipulated variable and the environment variable value, and then uses this prediction model to set the state variable value to the target value. Calculate the optimal feature quantity of the manipulated variable that approaches . The driving support device 10 according to the present embodiment performs various operations within a range close to the manipulated variable series (that is, the optimal manipulated variable series) when the feature amount of the optimal manipulated variable is converted into a time series. The quantity series is output to a user such as an operator. As a result, the optimal feature value of the manipulated variable that is not found in past performance data is obtained as the optimal solution, and various manipulated variable series corresponding to the feature value within a range similar to the optimal solution are presented to users such as operators. can do. Therefore, a user such as an operator selects a reasonable manipulated variable series from these various manipulated variable series as candidates, in light of his or her own empirical rules, and operates the target with the selected manipulated variable series. It is possible to formulate driving plans.

Note that, for example, when the object is operated according to a manipulated variable series selected by a user such as an operator, the operation results are obtained as performance data. Similarly, for example, when a driving plan is formulated with reference to a manipulated variable series selected by a user such as an operator, the object is driven according to this driving plan, and the driving results are obtained as performance data. Therefore, the above-described model learning process may be executed using such performance data to re-learn the predictive model f.

<Modified example>
Some modifications of the above embodiment will be described.

≪Modification 1≫
In the above embodiment, various manipulated variable series close to the optimal manipulated variable series are presented to users such as operators. The operation amount series corresponding to the amount) may also be presented to a user such as an operator.

≪Modification 2≫
In the second embodiment of the operation amount series extraction process described above, for each tag selected in step S502, the actual operation feature closest to the optimal operation feature from among the actual operation features included in the actual performance data to which that tag is attached Although the amounts were extracted, clustering may be performed as in the first embodiment. That is, for each tag selected in step S502, the performance operation feature amount included in the performance data to which the tag is attached is divided into a plurality of clusters, and then, as in the first embodiment, the representative point of the cluster is The closest actual operation feature quantity may be extracted. As a result, when the tag selected in step S502 is {1, ..., W}, and the performance operation feature amount included in the performance data to which the tag r∈{1, ..., W} is clustered, If the number of clusters is n _r , then n ₁ +...+n _W actual operation feature quantities are extracted, and as a result, n ₁ +...+n _W pieces corresponding to these actual operation feature quantities are extracted. The actual manipulated variable series is obtained as the manipulated variable series to be presented.

≪Modification 3≫
If the accuracy of the optimal manipulated variable series (that is, the manipulated variable series corresponding to the optimal manipulated feature) is sufficiently high, various manipulated variable series are presented to users such as operators, and the optimal manipulated variable series The object may be driven automatically.

The present invention is not limited to the above-described specifically disclosed embodiments, and various modifications and changes, combinations with known techniques, etc. can be made without departing from the scope of the claims. be.

10 Driving support device 101 Input device 102 Display device 103 External I/F
103a Recording medium 104 Communication I/F
105 RAM
106 ROM
107 Auxiliary storage device 108 Processor 109 Bus 201 Input section 202 Feature conversion section 203 Model learning section 204 Optimization section 205 Time series extraction section 206 Output section 207 Storage section

Claims (9)

a storage unit that stores a set of performance data in which feature quantities of manipulated variables in a past time interval are associated with time-series data of manipulated variables in the past time interval;
Using the feature amount of the manipulated variable in the past time interval, the feature amount of the state variable in the past time interval, and the feature amount of the environment variable in the past time interval, a learning unit configured to learn a prediction model that predicts the feature amount of the state variable from the feature amount of the environmental variable;
Calculating the optimal feature amount of the manipulated variable in the future time interval using the prediction model using the feature amount of the environment variable in the future time interval and the target feature amount of the state variable in the future time interval. an optimization unit configured to:
Among the features of the manipulated variables included in the performance data stored in the storage unit, n (where n is a predetermined integer of 2 or more) features that are at least similar to the optimal feature of the manipulated variables. n' corresponding to any one of the n feature quantities (however, n' is n' a time series extraction unit configured to extract time series data (an integer of 2 or more satisfying ≦n) from the storage unit ;
A driving support device with The time series extraction unit is
extracting n feature quantities similar to the optimal feature quantity of the manipulated variable from the performance data;
Divide the set of n features into n' clusters using a clustering method,
For each cluster, the n' time series data are extracted by extracting time series data associated with a feature most similar to a representative point of the cluster from the performance data. The driving support device according to claim 1 . Any one of n' tags is set in the performance data,
The time series extraction unit is
For each tag set in the performance data, by extracting time-series data associated with the feature most similar to the optimal feature of the manipulated variable from the performance data, The driving support device according to claim 1 , configured to extract series data. The driving support device according to claim 3 , wherein the tag is information that identifies an operator who operated an operation sequence represented by time series data included in the performance data. The driving support according to any one of claims 1 to 4 , further comprising an output section configured to output the n' time series data extracted by the time series extraction section to a predetermined output destination. Device. It is configured to collect time-series data of manipulated variables in the past time interval, feature quantities of state variables in the past time interval, and feature quantities of environmental variables in the past time interval from the target. collection department and
Claim further comprising: a feature converter configured to calculate an average value or a total of time-series data of the manipulated variable in the past time interval as a feature of the manipulated variable in the past time interval. 5. The driving support device according to 5 . The learning department is
The feature quantity of the manipulated variable in the past time interval and the feature quantity of the environmental variable in the past time interval are input, and the prediction is made so that the prediction accuracy of the feature quantity of the state variable in the past time interval is high. The driving assistance device according to claim 6 , configured to learn a model. a storage procedure for storing in a storage unit a set of performance data in which feature quantities of manipulated variables in a past time interval are associated with time-series data of manipulated variables in the past time interval;
Using the feature amount of the manipulated variable in the past time interval, the feature amount of the state variable in the past time interval, and the feature amount of the environment variable in the past time interval, a learning procedure configured to learn a prediction model that predicts the feature amount of the state variable from the feature amount of the environmental variable;
Calculating the optimal feature amount of the manipulated variable in the future time interval using the prediction model using the feature amount of the environment variable in the future time interval and the target feature amount of the state variable in the future time interval. an optimization procedure configured to
Among the features of the manipulated variables included in the performance data stored in the storage unit, n (where n is a predetermined integer of 2 or more) features that are at least similar to the optimal feature of the manipulated variable. n' corresponding to any one of the n feature quantities (however, n' is n' a time series extraction procedure configured to extract time series data (an integer of 2 or more satisfying ≦n) from the storage unit ;
A driving assistance method performed by a computer. a storage procedure for storing in a storage unit a set of performance data in which feature quantities of manipulated variables in a past time interval are associated with time-series data of manipulated variables in the past time interval;
Using the feature amount of the manipulated variable in the past time interval, the feature amount of the state variable in the past time interval, and the feature amount of the environment variable in the past time interval, a learning procedure configured to learn a prediction model that predicts the feature amount of the state variable from the feature amount of the environmental variable;
Calculating the optimal feature amount of the manipulated variable in the future time interval using the prediction model using the feature amount of the environment variable in the future time interval and the target feature amount of the state variable in the future time interval. an optimization procedure configured to
Among the features of the manipulated variables included in the performance data stored in the storage unit, n (where n is a predetermined integer of 2 or more) features that are at least similar to the optimal feature of the manipulated variables. n' corresponding to any one of the n feature quantities (however, n' is n' a time series extraction procedure configured to extract time series data (an integer of 2 or more satisfying ≦n) from the storage unit ;
A program that causes a computer to execute.

Images