JP2021033629A - Control apparatus and control method - Google Patents

Abstract

To properly maintain accuracy of estimating a model.SOLUTION: A control apparatus includes a storage unit, a determination unit, and a setting unit. The storage unit stores parameters to be applied to a model generated by machine learning, in accordance with ranges of application of the parameters. A determination unit determines a range of application to which input data to be input to the model corresponds. The setting unit sets a parameter to the model in accordance with the range of application determined by the determination unit.SELECTED DRAWING: Figure 2

Description

The present invention relates to a control device and a control method.

Conventionally, there is a technique for re-learning a model when the estimation accuracy of the model generated by machine learning is lowered (see, for example, Patent Document 1).

International Publication No. 2016/152053

However, with the prior art, there is a risk that the estimation accuracy of the model cannot be appropriately ensured. Specifically, in the prior art, since re-learning is performed using all the input data, the estimation accuracy of the model for some input data is not always improved.

The present invention has been made in view of the above, and an object of the present invention is to provide a control device and a control method capable of appropriately maintaining the estimation accuracy of a model.

In order to solve the above-mentioned problems and achieve the object, the control device according to the embodiment includes a storage unit, a determination unit, and a setting unit. The storage unit stores parameters applied to the model generated by machine learning according to the applicable range of the parameters. The determination unit determines which application range the input data input to the model corresponds to. The setting unit sets the parameter in the model based on the applicable range determined by the determination unit.

According to the present invention, the estimation accuracy of the model can be appropriately maintained.

FIG. 1A is a diagram showing a mounting example of a control device. FIG. 1B is a diagram showing an outline of a control method. FIG. 2 is a block diagram of the control device. FIG. 3 is a diagram showing an example of the parameter DB. FIG. 4 is a diagram showing an example of processing by the search unit. FIG. 5 is a flowchart (No. 1) showing a processing procedure executed by the control device. FIG. 6 is a flowchart (No. 2) showing a processing procedure executed by the control device.

Hereinafter, the control device and the control method according to the embodiment will be described with reference to the accompanying drawings. The present invention is not limited to the embodiments shown below.

First, the outline of the control device and the control method according to the embodiment will be described with reference to FIGS. 1A and 1B. FIG. 1A is a diagram showing a mounting example of a control device. FIG. 1B is a diagram showing an outline of a control method.

As shown in FIG. 1A, the control device 1 according to the embodiment is mounted on the vehicle 100. The control device 1 generates output data Do by inputting input data Di based on sensor information input from the vehicle-mounted sensor of the vehicle 100 into the model M. Here, the model M has layers called an input layer, an intermediate layer (hidden layer), and an output layer, and transmits the input data Di input from the input layer to the output layer while propagating each connection path. ..

At this time, the output data Do corresponding to the input data Di is generated by executing the arithmetic processing based on the parameters set in each connection path. The parameters include weight parameters, bias parameters, and the like.

By the way, as described above, the control device 1 performs arithmetic processing using the model M based on the sensor information input from the in-vehicle sensor. Here, the in-vehicle sensor may have an installation error for each vehicle 100, or the parts of the vehicle 100 may deteriorate.

Therefore, it is necessary to optimize the model M according to the current state of the vehicle 100. As a method of optimizing the model M, for example, a method of re-learning the model M using learning data corresponding to all the conditions can be considered.

However, in this case, it is necessary to verify whether or not the model M after re-learning is effective under all conditions, which is not preferable because it causes an increase in cost. Further, in this case, if re-learning is performed using the learning data corresponding to all the conditions, the accuracy of the output data Do with respect to some input data Di may be deteriorated before and after the re-learning.

Therefore, in the control method according to the embodiment, the parameters applied to the model M are stored according to the applicable range of the parameters, and the parameters are used properly according to the input data Di.

In the example shown in FIG. 1B, the case where the applicable range R1 and the applicable range R2 are the ranges set based on the engine load factor which is the first condition and the engine speed which is the second condition, respectively, is shown.

The first parameter corresponds to the applicable range R1, and the second parameter different from the first parameter corresponds to the applicable range R2. In the following, when the applicable range R1 and the applicable range R2 are not distinguished, they are simply described as the applicable range R.

Then, in the control method according to the embodiment, it is determined which application condition the input data Di corresponds to, the parameters of the model M are set according to the determination result, and then the input data Di is input to the model M.

Specifically, if the input data Di is the applicable range R1, the first parameter is set in the model M, and if the input data Di is the applicable range R2, the second parameter is set in the model M.

As described above, in the control method according to the embodiment, a plurality of parameters are stored according to the application range R of the parameters, and optimization is performed so that the model M covers the entire range with the plurality of parameters.

Therefore, in the control method according to the embodiment, when the applicable range R of the parameter changes due to deterioration of parts or the like, it is not necessary to relearn all the parameters, and only some parameters corresponding to some conditions are selected. You can relearn.

That is, in the control method according to the embodiment, it is possible to relearn the parameters corresponding to a part of the applicable range by using the learning data of some conditions. In other words, the parameters of the condition that was originally good in accuracy are not changed, but only the parameters of the condition that is poor in accuracy are changed.

Therefore, in the control method according to the embodiment, the estimation accuracy of the model can be appropriately maintained. In the above-mentioned example, the applicable range R for the two conditions of the first condition and the second condition is shown, but the present invention is not limited to this, and the number of such conditions may be one or more. It may be.

Further, in the above-described embodiment, the case where the applicable range is two is shown, but the present invention is not limited to this, and the applicable range R may be three or more.

Next, a configuration example of the control device 1 according to the embodiment will be described with reference to FIG. FIG. 2 is a block diagram of the control device 1. Note that FIG. 2 shows a control system S including the control device 1 and the server device 50. Further, it is assumed that the control system S includes a plurality of control devices 1.

In addition, FIG. 2 also shows the in-vehicle sensors 101. The in-vehicle sensors 101 are, for example, sensors that detect the state of the internal combustion engine (engine) of the vehicle 100. The control device 1 can determine a state such as an engine misfire by inputting the sensor information input from the in-vehicle sensors 101 into the model M.

Further, the server device 50 is a server that collectively manages the model M owned by each control device 1. As will be described later, when the server device 50 provides the model M to the control device 1 and receives the relearning request of the model M from the control device 1, the server device 50 relearns based on the relearning request. It has a function to perform.

After that, the server device 50 transmits the parameter which is the learning result of the re-learning to the control device 1. As a result, the control device 1 can optimize the parameters and optimize the model M.

As shown in FIG. 2, the control device 1 according to the embodiment includes a storage unit 2 and a control unit 3. The storage unit 2 is realized by, for example, a semiconductor memory element such as a RAM or a flash memory, or a storage device such as a hard disk or an optical disk.

Further, as shown in FIG. 2, the storage unit 2 includes a model DB 21, a parameter DB 22, and a learning data DB 23. The model DB 21 is a database that stores the model M.

The parameter DB 22 is a database that stores the parameters applied to the model M. As described above, a plurality of parameters are stored in the parameter DB 22 according to the application range of the parameters.

FIG. 3 is a diagram showing an example of the parameter DB 22. As shown in FIG. 3, in the parameter DB 22, "parameter ID", "applicable range", "weight parameter", "bias parameter" and the like are stored in association with each other.

The parameter ID is an identifier that identifies each parameter. The applicable range indicates the applicable range of the corresponding parameter. In the present embodiment, the applicable range indicates the applicable range provided for the first condition and the second condition shown in FIG. 1B.

The weight parameter and the bias parameter are examples of the above-mentioned parameters. These parameters are values derived by machine learning. In the example shown in FIG. 3, the weight parameter and the bias parameter are schematically shown as "C01" and "D01", but specific information is described in "C01" and "D01". It shall be. Further, in "C01" and "D01", a number of parameters corresponding to the number of nodes of the model M are described, respectively.

Returning to the description of FIG. 2, the learning data DB 23 will be described. The learning data DB 23 is a database that stores data used during re-learning. As will be described later, the control device 1 according to the embodiment has a function of searching for a parameter application range, and stores data outside the application range as learning data in the learning data DB 23.

The control unit 3 is a controller, and is realized by, for example, a CPU, an MPU, or the like executing various programs stored in the storage unit 2 with the RAM as a work area. Further, the control unit 3 can be realized by, for example, an integrated circuit such as an ASIC or FPGA.

As shown in FIG. 2, the control unit 3 includes an acquisition unit 31, a determination unit 32, a setting unit 33, a calculation unit 34, a search unit 35, and an update unit 36. The acquisition unit 31 acquires the sensor information input from the vehicle-mounted sensors 101. For example, the in-vehicle sensors 101 are sensors provided in the internal combustion engine of the vehicle 100 to detect the state of the internal combustion engine.

The determination unit 32 determines which application range R the input data Di input to the model M corresponds to. Specifically, the determination unit 32 refers to the application range of the parameter DB 22 and determines which application range the above sensor information (input data) corresponds to.

The setting unit 33 sets the parameters based on the applicable range determined by the determination unit 32. Specifically, the setting unit 33 extracts the parameter of the applicable range determined by the determination unit 32 from the parameter DB 22 and sets such a parameter in the model M.

At this time, if the previous application range and the current application range are the same, the setting unit 33 may keep the parameters of the model M as they are. That is, the setting unit 33 may set the parameter only when the applicable range R changes.

The calculation unit 34 generates output data Do by inputting the sensor information acquired by the acquisition unit 31 into the model M as input data Di. Further, the calculation unit 34 can also generate the output data Do for evaluation by inputting the evaluation data into the model M.

The search unit 35 searches the parameter application range R based on the output data Do obtained by inputting the evaluation data into the model M. Specifically, the search unit 35 searches for a range in which the estimation accuracy of the model M exceeds the threshold value as the application range R, and a range in which the estimation accuracy is lower than the threshold value as the non-applicable range Rn.

In the following, it is assumed that the model M is a model that generates the probability of knocking of each cylinder as output data Do. First, the search unit 35 notifies an engine control device (not shown) of an instruction to generate evaluation data.

Specifically, after designating the first condition and the second condition, for example, the engine control device is instructed to operate the engine under the condition where knocking occurs. As a result, the search unit 35 can acquire sensor information indicating knocking as evaluation data. The search unit 35 can obtain all the evaluation data by instructing the engine control device while sequentially changing the first condition and the second condition.

Subsequently, the search unit 35 acquires the output data Do obtained by inputting the evaluation data into the model M from the calculation unit 34, and searches the applicable range of the parameters based on the output data Do and the teacher label. ..

Here, the output data Do obtained from the evaluation data should indicate that the knocking is inherent.

Therefore, the search unit 35 can determine that the parameters match when the output data Do indicates knocking, and the parameters match when the output data Do does not indicate knocking. It can be determined that there is no such thing.

For example, the search unit 35 defines a condition in which the knocking probability (hereinafter referred to as knock probability) indicated by the output data Do is 80% or more as the applicable range R, and a condition in which the knock probability is less than 80% as the non-applicable range Rn. Perform a search.

Then, the search unit 35 stores the input data Di and the output data Do in the non-applicable range in the learning data DB 23. At this time, the search unit 35 does not have to search the entire range with one parameter, and may search only a part of the range.

For example, the search unit 35 searches the current appropriate range R + α using the corresponding parameters. As a result, the search range of one parameter can be reduced, so that the search efficiency can be improved.

In this way, the search unit 35 can identify the non-applicable range by searching the applicable range and the non-applicable range of each parameter. In other words, it is possible to respond immediately to the non-applicable range. The search unit 35 shall perform the search process at predetermined intervals.

Further, when the non-applicable range is adjacent to the plurality of applicable ranges R, the search unit 35 can perform the search again by using the parameters of the plurality of applicable ranges R.

FIG. 4 is a diagram showing an example of processing by the search unit 35. For example, the search unit 35 searches for the non-applicable range Rn1 using only the first parameter corresponding to the applicable range R1, and for the non-applicable range Rn2, the second parameter corresponding to the first parameter and the applicable range R2. The non-applicable range Rn1 is searched using each of the parameters of.

This is because the non-applicable range Rn1 is not adjacent to the applicable range R2, and therefore, even if the second parameter is used, it is unlikely that the non-applicable range Rn1 will be the applicable range R of the second parameter. On the other hand, the non-applicable range Rn2 is adjacent to each of the applicable range R1 and the applicable range R2.

Therefore, even if the non-applicable range Rn2 changes from the applicable range R of the first parameter to the non-applicable range Rn, it is assumed that the non-applicable range Rn2 becomes the applicable range R of the second parameter. In this way, the search unit 35 searches for the non-applicable range Rn using each of the parameters of the adjacent applicable range R, so that the re-learning described later can be omitted.

Returning to the description of FIG. 2, the update unit 36 will be described. The update unit 36 updates the parameters of the non-applicable range Rn to the parameters relearned based on the evaluation data and the output data Do of the non-applicable range Rn.

Specifically, the update unit 36 extracts the evaluation data of the non-applicable range Rn and the teacher label (label indicating whether or not knocking is actually performed) from the learning data DB 23, and sends the learning request to the server device 50. Send.

The server device 50 retrains using the model M based on the evaluation data and the teacher label to generate optimized parameters in the non-applicable range Rn.

After that, the server device 50 transmits the parameter after re-learning to the control device 1, and the update unit 36 updates the parameter DB 22 based on the parameter. That is, the update unit 36 can update the applicable range of the parameter and add or delete the parameter.

In this way, the update unit 36 can reduce the non-applicable range Rn by updating the parameter of the non-applicable range Rn. Further, in the control device 1 according to the present embodiment, re-learning is performed based on the input data Di (evaluation data) of the non-applicable range Rn and the corresponding teacher label. Therefore, it is possible to change the parameter of the non-applicable range Rn without changing the parameter of the applicable range R.

That is, it is possible to secure the estimation accuracy of the model M in the applicable range R and then newly secure the estimation accuracy of the non-applicable range Rn. Therefore, according to the control device 1 according to the embodiment, the estimation accuracy of the model can be appropriately maintained.

Next, the processing procedure executed by the control device 1 according to the embodiment will be described with reference to FIGS. 5 and 6. 5 and 6 are flowcharts showing a processing procedure executed by the control device 1.

First, a control procedure in a normal state will be described with reference to FIG. As shown in FIG. 5, first, when the control device 1 acquires the input data Di (step S100), the control device 1 determines the application range R corresponding to the input data Di (step S101).

Subsequently, the control device 1 sets the parameters of the model M based on the application range R determined in step S101 (step S102), inputs the input data to the model M (step S103), and ends the process. To do.

Subsequently, with reference to FIG. 6, a series of processes associated with the parameter update process will be described. As shown in FIG. 6, the control device 1 determines whether or not the update timing is reached (step S111). Here, the control device 1 can determine the update timing by using, for example, a predetermined cycle or the timing of exchanging parts as a trigger and the update timing as a trigger.

When the control device 1 determines in step S111 that the search timing is reached (step S111, Yes), the control device 1 acquires evaluation data (step S112) and performs search processing (step S113).

Subsequently, the control device 1 determines whether or not there is a non-applicable range Rn as a result of the search process (step S114), and if there is a non-applicable range Rn (step S114, Yes), the data of the non-applicable range Rn. At the same time, a learning request is transmitted to the server device 50 (step S115).

After that, the control device 1 acquires the parameter after re-learning from the server device 50 (step S116), updates the parameter DB 22 based on the acquired parameter (step S117), and ends the process.

As described above, the control device 1 according to the embodiment includes a storage unit 2, a determination unit 32, and a setting unit 33. The storage unit 2 stores the parameters applied to the model M generated by machine learning according to the parameter application range R. The determination unit 32 determines which application range R the input data input to the model M corresponds to.

The setting unit 33 sets parameters in the model M based on the application range R determined by the determination unit 32. Therefore, according to the control device 1 according to the embodiment, the estimation accuracy of the model M can be appropriately maintained.

By the way, in the above-described embodiment, the case where the control device 1 is mounted on the vehicle 100 has been described, but the present invention is not limited to this. That is, the present invention can be applied to various devices such as personal computers.

Further, in the above-described embodiment, the case where the server device 50 relearns the parameters in the non-applicable range has been described, but the present invention is not limited to this. That is, the control device 1 (update unit 36) may perform re-learning.

Further effects and variations can be easily derived by those skilled in the art. For this reason, the broader aspects of the invention are not limited to the particular details and representative embodiments expressed and described as described above. Therefore, various modifications can be made without departing from the spirit or scope of the general concept of the invention as defined by the appended claims and their equivalents.

1 Control device 31 Acquisition unit 32 Judgment unit 33 Setting unit 34 Calculation unit 35 Search unit 36 Update unit M model

Claims (6)

A storage unit that stores parameters applied to the model generated by machine learning according to the applicable range of the parameters, and a storage unit.
A determination unit that determines which of the applicable ranges the input data input to the model corresponds to, and
A control device including a setting unit for setting the parameters in the model based on the application range determined by the determination unit. The control device according to claim 1, further comprising a search unit for searching the applicable range of the parameter based on output data obtained by inputting evaluation data into the model. The search unit
A range in which the estimation accuracy of the model exceeds the threshold value is set as the applicable range, and a range in which the estimated accuracy is lower than the threshold value is set as the non-applicable range.
2. Claim 2 comprising an update unit for updating the parameters in the non-applicable range to parameters relearned based on the evaluation data in the non-applicable range and the teacher label of the evaluation data. The control device described in. The search unit
The control device according to claim 3, wherein the range searched as the non-applicable range is searched again using other parameters of the applicable range adjacent to the range. The control device according to any one of claims 1 to 4, further comprising an acquisition unit that acquires the input data from a sensor that detects the state of the vehicle. A storage process that stores the parameters applied to the model generated by machine learning according to the applicable range of the parameters, and
A determination step of determining which said application range the input data input to the model corresponds to, and
A control method including a setting step of setting the parameter based on the applicable range determined by the determination step.

Images