JP7621874B2 - Automated Learning System - Google Patents

Description

This embodiment relates to an automatic learning system.

In railway and other vehicles, events that could not be anticipated at the time of vehicle manufacture may occur due to aging of the vehicle, changes in the environment, changes in passenger behavior, etc. For this reason, in autonomous driving technology, the control parameters used to control various equipment on the vehicle may not remain appropriate permanently after they are defined at the time of vehicle manufacture.

However, because the control parameters are defined in software built into the vehicle's control device, updating the control parameters is not easy.

JP 2016-19411 A

We provide an automatic learning system that can set vehicle control parameters to appropriate states depending on the vehicle's condition and the vehicle's usage environment.

The automatic learning system according to this embodiment is an automatic learning system having control parameters used to control a vehicle. The first automatic learning device includes a data storage unit, a learning execution unit, a parameter generation unit, and an output unit. The data storage unit accumulates data related to the first control parameters collected from the vehicle during operation. The learning execution unit analyzes the data to generate a regression model or a classification model. The parameter generation unit generates a new second control parameter using the regression model or the classification model. The output unit outputs the second control parameter. The second automatic learning device is mounted on the vehicle and includes a parameter storage unit, a parameter control unit, and a vehicle control unit. The parameter storage unit stores the second control parameter from the first automatic learning device. The parameter control unit updates the first control parameter with the second control parameter. The vehicle control unit controls the vehicle using the second control parameter.

FIG. 1 is a block diagram showing an example of the configuration of an automatic learning system according to a first embodiment. FIG. 11 is a conceptual diagram showing an example of learning item information. FIG. 4 is a flow chart showing an example of the operation of the automatic learning system according to the first embodiment. FIG. 4 is a conceptual diagram showing an example of binary data in a data storage unit. FIG. 11 is a conceptual diagram showing an example of data definition information of binary data. FIG. 4 is a diagram showing a specific example of control parameters stored in a control parameter storage unit of the vehicle. FIG. 2 is a diagram showing an outline of a part of a control program of the vehicle control device. 13 is a graph showing the relationship between a first model and performance evaluation data. 13 is a graph showing the relationship between a second model and performance evaluation data. 13 is a graph showing a second model and a second control parameter. FIG. 13 is a diagram showing an example of control parameters after being updated with a second control parameter. 13 is a diagram showing an example of a screen of a display unit that displays permission or prohibition of distribution of a second control parameter. FIG. FIG. 11 is a flow diagram showing a modified example of a distribution method for a train configuration. FIG. 4 is a conceptual diagram showing an update state of control parameters in a train configuration to be updated. FIG. 4 is a conceptual diagram showing an update state of control parameters in a train configuration to be updated. FIG. 4 is a conceptual diagram showing an update state of control parameters in a train configuration to be updated. FIG. 13 is a diagram showing the definition of the graphical representation of the control parameters. 13 is a graph showing parameter performance information of a first control parameter; 13 is a graph showing parameter performance information of the second control parameter; FIG. 13 is a diagram showing an example of a classification model that classifies a variable that can take three classes into three classes. FIG. 11 is a block diagram showing a configuration example of an automatic learning system according to a second embodiment.

Below, an embodiment of the present invention will be described with reference to the drawings. The present invention is not limited to this embodiment. The drawings are schematic or conceptual, and the ratio of each part is not necessarily the same as the actual one. In the specification and drawings, elements similar to those described above with respect to the previous drawings are given the same reference numerals, and detailed explanations are omitted as appropriate.

First Embodiment
Fig. 1 is a block diagram showing an example of the configuration of an automatic learning system according to a first embodiment. The automatic learning system 1 according to this embodiment includes a vehicle 100 and a server 200. The vehicle 100 constitutes, for example, a train configuration that runs on a railway. Although only one vehicle 100 is shown in Fig. 1, each train configuration is made up of a plurality of vehicles. Furthermore, a plurality of train configurations are operated on a railway.

The vehicle 100 includes a parameter control device 109, a parameter storage unit 110, and a vehicle control device 120. The parameter control device 109 as a parameter control unit receives control parameters and their unique names transmitted from the server 200. The parameter storage unit 110 stores various control parameters 111 used to control the vehicle 100. The parameter control device 109 may be, for example, a PC (Personal Computer) or the like. The parameter storage unit 110 may be, for example, a storage device such as a HDD (Hard Disk Drive) or SSD (Solid State Drive). The control parameters 111 are parameters that automatically control the drive device, air conditioning device, lighting device, etc. of the vehicle 100. Therefore, the control parameters determine the timing of speed switching, air conditioning switching, lighting switching, the speed of the vehicle 100, notch switching, braking, etc. The vehicle control device 120 controls the drive device, air conditioning device, lighting device, etc. of the vehicle 100 based on the control parameters. A wide variety of control parameters are set. To distinguish and update each control parameter, each control parameter is given a unique name as an identifier. The parameter control device 109, the parameter storage unit 110, and the vehicle control device 120 constitute a second automatic learning device 700.

The vehicle 100 includes, for example, an air conditioner 130, a temperature and humidity sensor 140, and a load compensation sensor 150. The air conditioner 130 controls the temperature and humidity inside the vehicle 100. The temperature and humidity sensor 140 measures the temperature and humidity inside the vehicle 100. The load compensation sensor 150 measures the load of the vehicle 100 and calculates the occupancy rate. The air conditioning and air volume setting data of the air conditioner 130, the temperature and humidity data measured by the temperature and humidity sensor 140, the load and occupancy rate data measured by the load compensation sensor 150, and other data are transmitted to the vehicle control device 120 and used to control the air conditioner 130, and are also transmitted to the server 200 via the network 10. The vehicle control device 120 feedback controls the air conditioner 130 based on the control parameter. For example, if the control parameter is a set value of temperature or humidity, the vehicle control device 120 cools the air conditioner 130 when the temperature or humidity becomes higher than the control parameter. When the control parameter is a set value of the occupancy rate, the vehicle control device 120 assumes that the temperature and humidity inside the vehicle 100 will increase when the occupancy rate becomes higher than the control parameter, and sets the air conditioning device 130 to high cooling mode. The vehicle control device 120 may be, for example, a PC.

The server 200 includes an automatic learning device 300 as a first automatic learning device, a control parameter generating device 400, a control parameter output device 500, and a display unit 600. The server 200 is provided on the ground and is communicatively connected to the vehicle 100 via the network 10.

The automatic learning device 300 is a device that automatically generates a model (equation) for calculating control parameters using data from the vehicle 100. The automatic learning device 300 includes a data storage unit 310, a learning execution unit 320, and a learning item storage unit 330.

The data storage unit 310 accumulates data on control parameters collected from the vehicle 100 during operation. For example, the setting data of the air conditioning device 130, temperature data, humidity data, load data, occupancy rate data, etc. in the vehicle 100 are received from the vehicle 100 via the network 10 by the automatic learning device 300 and stored in the data storage unit 310. The data may be all data measured by the vehicle 100. The data may be periodically sampled while the vehicle 100 is in operation and immediately transmitted to the server 200 each time it is sampled. Alternatively, the data may be stored in a memory unit (not shown) of the vehicle 100 for a predetermined period of time and transmitted to the server 200 in a lump after the predetermined period has elapsed. At this time, the data storage unit 310 compresses the data and accumulates it in binary format. Since the data in binary format has a smaller data capacity than text data such as a database or CSV (Comma Separated Values) format, the capacity of the data storage unit 310 can be reduced. This reduces the cost of data accumulation. Because the data capacity is small, the data storage unit 310 may store all data measured by the vehicle 100, and the data storage period may be arbitrary or indefinite.

The learning item storage unit 330 stores information (definitions) necessary to generate a regression model or a classification model from the data in the data storage unit 310. For example, the learning item storage unit 330 includes learning method information 341, model performance information 342, learning start trigger information 343, relationship information 344 between the model and the control parameters, and regression model or classification model (also simply called model) information 345.

The learning method information 341 is information about an algorithm that is selected and used from a regression model and a classification model. The regression model may be, for example, either linear regression or nonlinear regression. The learning algorithm may be an algorithm that is generally provided in a programming language such as Python. Alternatively, the learning algorithm may be a custom-defined one.

The model performance information 342 is information indicating the degree of fit between the regression model or classification model generated by the learning execution unit 320 and the data. The degree of fit can be, for example, the squared error used in the regression model, or the precision, recall, F-value, etc. used in the classification model, as shown in FIG. 2. Note that the smaller the squared error, the higher the degree of fit. Also, the larger the precision, recall, and F-value, the higher the degree of fit.

The learning start trigger information 343 is information indicating the conditions under which the learning execution unit 320 starts data analysis. For example, as shown in FIG. 2, the learning start trigger information 343 may be the date and time at which data analysis is to start, or the above-mentioned threshold value of compatibility. When the set date and time arrives, the learning execution unit 320 may automatically start data analysis using the data in the data storage unit 310 and generate a model. Alternatively, when the compatibility between an existing model and data decreases and becomes lower than the threshold value, the learning execution unit 320 may automatically start data analysis using the data in the data storage unit 310 and regenerate a model.

The relationship information 344 between the model and the control parameters (also simply referred to as the relationship information 344) is information that indicates the correspondence between the name of the control parameter and the calculation formula of the control parameter based on the model. The name of the control parameter is the same as the above-mentioned name that is preset in the control parameter of the vehicle 100, or is a name that is identifiable and associated with it. This makes it possible to update the control parameters on the vehicle 100 side with new control parameters generated on the server 200 side.

Model 345 is a regression model or classification model generated by learning execution unit 320. Model 345 is added and saved each time it is generated by learning execution unit 320.

Learning item information 340, including learning method information 341, model performance information 342, learning start trigger information 343, and relationship information 344, is provided for each control parameter. When a large number of control parameters are set in the parameter storage unit 110 of the vehicle 100, the same number of pieces of learning item information 340 as the number of control parameters can be defined.

The learning item storage unit 330 may be a storage device separate from the data storage unit 310, or may be the same storage device as the data storage unit 310. The learning item storage unit 330 may be, for example, a storage device such as an HDD or SSD.

As described above, the learning execution unit 320 performs data analysis (also called learning) according to the definition of each learning item information, and generates a regression model or classification model suitable for each control parameter. At this time, the learning execution unit 320 restores only the data related to the control parameters from the binary format among the data stored in the data storage unit 310 to a text format such as CSV format. The following explanation uses CSV data as an example of text data. The data related to the control parameters (also called learning data) is, for example, explanatory variables, objective variables, etc. defined in the learning item information. After generating the model, the learning execution unit 320 may discard the restored CSV format data. This can reduce the capacity of the data storage unit 310 and the capacity of the learning item storage unit 330. The learning execution unit 320 may be, for example, a computing device such as a PC.

The control parameter generating device 400 is a device that generates a control parameter 420 using the model 345 generated in the automatic learning device 300. The control parameter generating device 400 includes a control parameter calculation unit 410. The control parameter calculation unit 410 calculates, for example, a numerical value obtained using the coefficient a of the model 345 and the relationship information 344 in FIG. 2 as the control parameter 420. At this time, the calculation method applied to the model 345 is defined in the relationship information 344. The unique name (identifier) of the control parameter 420 calculated at this time is also defined in the relationship information 344. The control parameter 420 obtained from the model 345 is obtained, for example, as a specific numerical value. The control parameter calculation unit 410 may be, for example, a calculation device such as a CPU.

The control parameter calculation unit 410 transfers the unique name of the control parameter 420 to the control parameter output device 500 together with the control parameter 420 as the second control parameter.

The control parameter output device 500 as an output unit includes a control parameter storage unit 510 and a control parameter transmission unit 520. The control parameter storage unit 510 stores the control parameters 420 and their unique names from the control parameter generation device 400. The control parameter storage unit 510 may be, for example, a storage device such as an HDD or SSD. The control parameter transmission unit 520 transmits the control parameters 420 and their unique names to the parameter storage unit 110 of the vehicle 100.

The control parameter transmission unit 520 may transmit the control parameters 420 and their unique names to the vehicle 100 online via the network 10. Alternatively, the control parameter transmission unit 520 may output the control parameters 420 and their unique names to an electronic recording medium (e.g., a USB memory or an SD card) communicatively connected to the control parameter transmission unit 520, and store the control parameters 420 and their unique names in the vehicle 100 via the electronic recording medium. In this case, even in an offline state in which the communication connection between the vehicle 100 and the server 200 is cut off, the control parameters in the parameter storage unit 110 can be updated via the electronic recording medium.

The parameter control device 109 receives the new control parameters transmitted from the server 200, and the parameter storage unit 110 stores the new control parameters. At this time, the parameter control device 109 updates the control parameters having the same unique name as the new control parameters with the new control parameters. The vehicle control device 120 controls the vehicle 100 using the new control parameters stored in the parameter storage unit 110.

The display unit 600 displays various information such as the control parameters 420, the model 345, and information about the vehicle 100. The display unit 600 may also display the control parameters used in each of the multiple vehicles 100 in multiple train configurations in comparison with data related to the control parameters. The display unit 600 may be, for example, a touch panel display that also serves as a user interface. Although not shown, a user interface may also be provided separately from the display unit 600.

Figure 2 is a conceptual diagram showing an example of learning item information. In this learning item information, the explanatory variable x is the occupancy rate, and the objective variable y is the humidity after 10 minutes. The learning algorithm is linear regression y = a x x, where a is a coefficient. The objective variable y of the learning algorithm indicates the humidity after 10 minutes for a certain occupancy rate x of the vehicle 100. The learning data period is the period from 30 days before to 1 day before the day (learning day) on which the learning execution unit 320 performs data analysis. The learning execution unit 320 applies the occupancy rate and humidity data actually measured during the learning data period to the learning algorithm and calculates the coefficient a.

In this case, the learning data filter condition is, for example, "the door is closed." The learning data filter condition is a condition for data used for analysis (learning) in the learning execution unit 320. For example, if the learning data filter condition is "the door is closed," the occupancy rate and humidity data measured when the door is closed is used. Conversely, the occupancy rate and humidity data measured when the door is open is removed. The open/closed state of the door (not shown) of the vehicle 100 can be determined by a command from the vehicle control device 120 to the door, or by a sensor installed in the door.

The learning algorithm indicates a specific model (formula) of a regression model or a classification model. In this embodiment, the learning algorithm is, for example, a linear regression model of y = a x x. The learning algorithm may be a linear regression model other than this model, or may be a nonlinear regression model. Furthermore, the learning algorithm may be a classification model.

The learning target train set indicates information about the train set that is the target of updating the control parameters 420. For example, if the learning target train sets are A001 and A002, the learning execution unit 320 calculates the coefficient a of the regression model using data (e.g., occupancy rate, humidity, time) obtained from A001 and A002 during the learning data period. At this time, data measured when the doors of A001 and A002 are closed is extracted.

As described above, the model performance information 342 indicates the degree of fit between the regression model or classification model and the data. For example, when the learning algorithm is a regression model, the model performance information 342 may be the squared error of each data with respect to the regression model (y = a x x). When the learning algorithm is a classification model, the model performance information 342 may be the matching rate, recall rate, and F value of each data with respect to the classification model. The two-row error, matching rate, recall rate, and F value are selectable by the user. The model performance information 342 is used to determine whether or not to apply the newly generated control parameters 420 to the vehicle 100 or train formation.

As described above, the learning start trigger information 343 indicates the conditions under which the learning execution unit 320 starts data analysis. The learning start trigger information 343 is, for example, a specific day or performance degradation. The specific day or performance degradation can be selected by the user. If a specific day is selected and it is "the first day of every month", the learning execution unit 320 automatically performs data analysis using the data in the data storage unit 310 and the learning method information 341 on the first day of every month to generate a new model. Alternatively, if performance degradation is selected and it is "P%; previous day's data", the learning execution unit 320 calculates the degree of conformity between the previous day's data and the existing model. If the previous day's data deviates from the existing model by P% or more, the learning execution unit 320 determines that the model performance has deteriorated. In this case, the learning execution unit 320 automatically performs data analysis using the data in the data storage unit 310 and the learning method information 341 to generate a new model.

As described above, the relationship information 344 indicates the correspondence between the name of the control parameter and the calculation formula of the control parameter based on the regression model or the classification model. For example, the relationship information 344 indicates the relationship between the name of the control parameter and the calculation method of the control parameter. In FIG. 2, the control parameter name is "param001" and the corresponding calculation method is "10/a". That is, the control parameter "param001" is expressed by Equation 1.
param001=10/a (Formula 1)
According to the definitions of explanatory variable x and objective variable y, this control parameter param001 represents the occupancy rate at which the humidity rises by 10% 10 minutes after the data is measured. That is, when the occupancy rate exceeds the control parameter param001, it is estimated that the humidity will rise by 10% 10 minutes later. This allows the control device 120 to switch the air conditioner 130 in advance so as to suppress the rise in humidity according to the occupancy rate.

The control parameter calculation unit 410 calculates Equation 1 from the model generated by the learning execution unit 320. This provides a specific value for the control parameter "param001". For example, if the coefficient a obtained by learning is 0.256, the control parameter calculation unit 410 substitutes 0.256 for the coefficient a in Equation 1 to calculate the control parameter param001. The control parameter param001 = 39%. This means that when the occupancy rate is 39%, the humidity inside the vehicle 100 will increase by 10% after 10 minutes. Note that since there may be multiple pieces of relationship information 344 between models and control parameters, it may be written in a program.

Next, the operation of the automatic learning system 1 according to this embodiment will be described.

Figure 3 is a flow diagram showing an example of the operation of the automatic learning system 1 according to the first embodiment. First, the server 200 receives various measurement data from the vehicle 100, compresses it in binary format, and stores it in the data storage unit 310 (S10). The data includes, for example, the temperature, humidity, occupancy rate, door open/close status, date and time, etc. of the vehicle 100. These data are sampled periodically, associated with the date and time, and stored in the data storage unit 310. The binary format data (hereinafter, binary data) is classified and organized in folders in the data storage unit 310, for example, by month or day. Furthermore, the binary data is classified and organized by folder for each data item or attribute of each data.

Figure 4A is a conceptual diagram showing an example of binary data in the data storage unit 310. Data is stored in a separate folder for each specific period during which it was measured. For example, data is stored in a separate folder for each month during which it was measured. Furthermore, data is stored in a separate folder for each device (attribute) to which the data relates. For example, data related to air conditioning, data related to brakes, data related to main circuits, and data related to doors are each stored in separate folders.

Figure 4B is a conceptual diagram showing an example of data definition information for binary data. The data items of the binary data and the devices (attributes) related to that data are defined in the data definition information so that they can be identified from outside the data storage unit 310. For example, the data items "outside temperature," "humidity," and "air conditioning operation mode" are related to the device "air conditioning." The data item "door opening/closing mode" is related to the device "door." Using the data definition information, the learning execution unit 320 can extract data related to each item of the learning method information 341 from the data storage unit 310.

The method of storing binary data is not limited to this. Therefore, the data storage unit 310 may store data by day or by hour in separate folders. The data storage unit 310 may also store data by data item that is more detailed than attributes in separate folders. In this way, storing data by folder makes it easier for the learning execution unit 320 to access the desired data.

Next, the learning execution unit 320 refers to the learning item information 340 in the learning item storage unit 330. At this time, if there is learning item information that matches the learning start trigger information 343 in FIG. 2 (YES in S20), the learning execution unit 320 automatically starts learning. For example, if it is the first day of a certain month, the learning execution unit 320 starts data analysis according to the learning item information in FIG. 2. If there is no learning item information that matches the learning start trigger information 343 (NO in S20), the learning execution unit 320 ends learning without performing data analysis.

When data analysis is started (YES in S20), the learning execution unit 320 extracts binary data from the data storage unit 310 according to the definition of the learning method information 341 (S30). For example, the learning execution unit 320 searches the data definition information in FIG. 4B for "occupancy rate" as a data item of the explanatory variable x, "humidity" as a data item of the objective variable y, and "door opening/closing mode" as a data item of the learning data filter condition, and identifies the equipment (attributes) "air conditioning", "brakes", and "doors" corresponding to each data item. Furthermore, the learning execution unit 320 identifies the folders for the equipment "air conditioning", "brakes", and "doors" in FIG. 4A from the folders "A001 train set" and "A002 train set" and "30 days before the learning date to 1 day before the learning date" from the data storage unit 310 according to the learning target train set and learning data period in the learning method information 341. Then, the learning execution unit 320 extracts binary data from the folders for "air conditioning", "brakes", and "doors".

Next, the learning execution unit 320 expands the extracted binary data into CSV data (S40). In this embodiment, the learning execution unit 320 expands only the data necessary for learning. Therefore, the memory capacity required to temporarily store the CSV data can be significantly smaller than the capacity required to expand all the data. In addition, the memory capacity required to read the expanded CSV data is also relatively small. Note that other compression formats may be used for efficiency.

Next, the learning execution unit 320 analyzes (learns) the expanded CSV data using the learning method information 341 specified in FIG. 2 to generate a model (hereinafter, a second model) (S50). At this time, the learning execution unit 320 divides the expanded CSV data into learning data and performance evaluation data. The learning execution unit 320 calculates the second model and the coefficient a, etc., using the learning data and the learning method information 341. For example, the learning execution unit 320 first uses linear regression as a learning algorithm to model the relationship between the occupancy rate and the humidity data 10 minutes after the measurement as the equation y = a x x. y is the humidity 10 minutes after the measurement, x is the occupancy rate, and a is a coefficient. Note that an intercept usually exists, but is omitted here. For modeling, the learning execution unit 320 first calculates the humidity data predicted value vector Y' 10 minutes later as shown in Equation 2 using the vector X that holds the occupancy rate data and a weight parameter w determined by a random number, etc.
Y'=X×w (Equation 2)

Next, the learning execution unit 320 calculates the squared error E from the humidity data vector Y for 10 minutes later and the humidity data predicted value vector Y' for 10 minutes later.

Next, the learning execution unit 320 updates the value of the parameter w using the gradient descent method or the like so that the squared error E becomes smaller. The learning execution unit 320 repeats this calculation to find the parameter w that minimizes the squared error E, and sets this value as the coefficient a. If there is too little learning data, the estimation accuracy by a×x decreases, so it is better to have a certain amount of learning data. Once the coefficient a is determined, the second model is determined as y = a×x. For example, if a = 0.256 as above, the second model is y = 0.256×x.

Once the coefficient a is determined, the second model is determined as y = a x. For example, if a = 0.256 as above, the second model is y = 0.256 x.

Next, the learning execution unit 320 inputs the performance evaluation data into the generated second model, and calculates model performance information indicating the degree of fit between the second model and the data (S60). The model performance information 342 is selected from, for example, squared error, accuracy rate, precision rate, recall rate, F-value, etc. In the case of a regression model, squared error is used. The model performance information obtained at this time is associated with the generated second model and stored in the learning item storage unit 330.

In addition, a threshold value may be set in advance for the model performance information, and if the degree of fit between the second model and the data is lower than the threshold value, the learning execution unit 320 may discard the second model. The threshold value for the model performance information may be stored in the learning item storage unit 330.

Next, the learning execution unit 320 reads out the current version of the model (hereinafter, the first model) that is already applied to the vehicle 100 from the learning item storage unit 330. The learning execution unit 320 also calculates model performance information for the first model using the same performance evaluation data as the model performance information for the second model (S70). Furthermore, the learning execution unit 320 compares the first model performance information indicating the suitability of the first model to the performance evaluation data with the second model performance information indicating the suitability of the second model to the performance evaluation data (S80).

If the second model performance information is higher in the degree of fit than the first model performance information (YES in S80), the control parameter calculation unit 410 applies the second model to the model-control parameter relationship information 344 to generate new control parameters (second control parameters) (S90). In this case, the second control parameters are stored in the control parameter storage unit 510 and transmitted from the control parameter transmission unit 520 to the vehicle 100 (S100). The second control parameters are received by the parameter control device 109 and stored in the parameter storage unit 110. The parameter control device 109 updates the first control parameters with the second control parameters (S110). Note that a high degree of fit means that the squared difference of the regression model is small, or that the accuracy, recall, and F value of the classification model are large, but other indices may also be used. In addition, the generalization performance of the second model may be verified more strictly by determining the degree of fit of the second model to the performance evaluation data at the time of creating the first model and comparing it with the degree of fit of the first model at the time of creating the first model.

If the second model performance information has a lower degree of conformance than the first model performance information (NO in S80), the control parameter calculation unit 410 does not generate the second control parameters. In this case, the vehicle control device 120 controls the vehicle 100 using the existing first control parameters as they are without updating them (S120). At this time, the second model and the second control parameters may be discarded.
Thereafter, the automatic learning system 1 repeatedly executes the learning and control parameter update of Fig. 3. This enables automatic updating of the control parameters to those adapted to the aging of the vehicle 100, changes in the environment, changes in passenger behavior, and the like.

(Control parameters)
For example, Fig. 5A is a diagram showing a specific example of control parameters stored in the parameter storage unit 110 of the vehicle 100. Fig. 5B is a diagram showing an outline of a part of the control program of the vehicle control device 120.

The air conditioner 130 can automatically switch between a strong cooling operation mode, which provides strong cooling, and a weak cooling operation mode, which provides weak cooling. Based on existing knowledge, the conditions for switching between the strong cooling operation mode and the weak cooling operation mode are stipulated by the occupancy rate at which the humidity inside the vehicle increases by 10% or more after 10 minutes. This occupancy rate (first control parameter) is defined by the control parameter name "param001" shown in FIG. 5A, and is currently set to 40%. This occupancy rate of 40% is stored as the first control parameter in the parameter storage unit 110 of the vehicle 100 in association with the control parameter name "param001". Other control parameters "param002", ... "paramXXX" are set in FIG. 5A, but here we focus on the control parameter "param001".

Figure 5B describes that if the occupancy rate exceeds 40% set in "param001", the vehicle control device 120 will switch the air conditioning mode. Although the description of the air conditioning mode is omitted in Figure 5B, in this embodiment, if the occupancy rate exceeds 40%, the vehicle control device 120 will switch the air conditioning device 130 to strong cooling operation mode.

In step S110 of FIG. 3, when updating the existing control parameter (first control parameter) "param001" with the second control parameter, the occupancy rate "Value=40" defined in "param001" of the parameter storage unit 110 of FIG. 5A is updated to the value of the second control parameter.

The "max" set in "param001" in FIG. 5A indicates the upper limit of the occupancy rate, and "min" indicates the lower limit of the occupancy rate. The unit of occupancy rate is percentage, so "max = 100" and "min = 0". When updating the control parameters, if the second control parameter is an abnormal value other than 0 to 100%, the second control parameter is discarded. In this way, the parameter storage unit 110 also stores the settings of the upper and lower limits of the control parameter in order to eliminate abnormal values of the second control parameter. Note that it is preferable to fix the upper and lower limits of the control parameter so that they are not updated for safety reasons.

(Model performance information)
Fig. 6A is a graph showing the relationship between the first model M1 and the performance evaluation data D. In step S70, model performance information is calculated for the first model. At this time, the relationship between the first model M1 and the performance evaluation data D is as shown in the graph of Fig. 6A. The coefficient a of the first model M1 is, for example, 0.18, and the error rate as the first model performance information is, for example, 0.02.

Figure 6B is a graph showing the relationship between the second model M2 and the performance evaluation data D. In step S60, model performance information is calculated for the second model. At this time, the relationship between the second model M2 and the performance evaluation data D is as shown in the graph in Figure 6B. The coefficient a of the second model M2 is, for example, 0.256, and the error rate as the second model performance information is, for example, 0.01.

When the first and second model performance information are compared, the error rate of the second model performance information is lower than the error rate of the first model performance information. In other words, the second model performance information has a higher degree of conformance than the first model performance information. Therefore, in this case, in step S90, the control parameter calculation unit 410 generates new second control parameters, and the first control parameters are updated with the second control parameters.

(Update of control parameters)
Fig. 7A is a graph showing the second model M2 and the second control parameter, and Fig. 7B is a diagram showing an example of the control parameter after being updated with the second control parameter.

In step S110, when updating the first control parameter with the second control parameter, the parameter control device 109 searches the parameter storage unit 110 by the parameter name and updates the control parameter having the same parameter name. For example, if the parameter name of the second control parameter is "param001", the parameter control device 109 searches for "param001" in the parameter storage unit 110 and updates the corresponding existing first control parameter. If the first control parameter is 40% as shown in FIG. 5A and the second control parameter is 39% (=10/0.256) as shown in FIG. 7A, the vehicle control device 120 changes "param001" from 40% to 39% as shown in FIG. 7B. As a result, when the occupancy rate exceeds 39%, it is estimated that the humidity will increase by 10% 10 minutes later, so the vehicle control device 120 switches the air conditioner 130 to the strong cooling operation mode.

When the first control parameter and the second control parameter have the same value, the first control parameter does not need to be updated with the second control parameter.

(Approval of control parameter updates)
The second control parameter may be transmitted from the control parameter output device 500 to a plurality of vehicles 100. For example, Fig. 8 is a diagram showing an example screen of the display unit 600 displaying permission or prohibition of distribution of the second control parameter. The distribution target indicates the train formation to which the second control parameter is distributed. In this example, formations P4515 and P5442 are the distribution targets.

The model performance information shows the model performance information of each of the first and second models. In Figures 6A and 6B, the model performance information is shown as an error rate, but in Figure 8, it is shown as an accuracy rate. The model performance information displayed on the display unit 600 is not particularly limited, and any of the model performance information 342 shown in Figure 2 may be displayed. Here, the accuracy rate of the second model is higher than the accuracy rate of the first model. Therefore, the fitness of the second model is higher than the fitness of the first model.

Furthermore, the display unit 600 displays an icon that allows the user to select whether or not to distribute the second control parameter to the cars of the train formation. When distributing the second control parameter, the user selects the "Distribute" icon. When not distributing the second control parameter, the user selects the "Do not distribute" icon. In this way, the user may be able to select whether to permit or prohibit the distribution of the second control parameter.

Generally, when a railway operator makes an update that changes the behavior of the vehicle, the railway operator must obtain approval from within the railway operator. Therefore, it is preferable that the automatic learning system 1 according to this embodiment displays a screen prompting the railway operator for approval when transmitting the second control parameters. This allows the control parameters to be updated only after obtaining approval from the railway operator. As a result, the safety of the operation of the vehicle 100 can be ensured.

The server 200 is also connected to a plurality of train configurations so that it can communicate with them. As shown in FIG. 8, the control parameter output device 500 may output the second control parameter to a preset portion of the plurality of train configurations. Alternatively, the control parameter output device 500 may output the second control parameter to all train configurations that are communicatively connected. In this case, when the parameter control device 109 of each train configuration receives the distributed second control parameter, it searches the parameter storage unit 110 for the unique name of the second control parameter. If there is a unique name that matches the unique name of the second control parameter, the first control parameter corresponding to the unique name is updated with the second control parameter. If there is no unique name that matches the unique name of the second control parameter, the control parameter of that vehicle is not updated.

Multiple control parameter updates may be performed at the same time. In this case, approval for many control parameter updates is required at the same time. It can be cumbersome for the railway operator to approve each control parameter update individually. Therefore, the display unit 600 may be configured to perform batch approval for multiple control parameter update processes that occur within a certain period of time.

Furthermore, even if the train formation is the same model, the train sections under its jurisdiction may be different. In this case, the display unit 600 may specify the train formation on a train section basis and approve the distribution of the control parameters. In this case, the display unit 600 may be able to select any train section, not just a specified train section, and approve the distribution of the control parameters.

Furthermore, depending on the control parameters, approval from the railway operator may be omitted. In this case, the display unit 600 may skip approval by omitting to display the icons "Distribute" and "Do not distribute" in FIG. 8.

In this way, according to this embodiment, even control parameters of devices that were not anticipated when the vehicle 100 was manufactured can be updated without being limited to specific control parameters by assigning a unique name (identifier) and registering it in the parameter storage unit 110. This allows all control parameters set in the parameter storage unit 110 to be updated with new control parameters from the external server 200. The new control parameters can be generated from a model obtained by automatic learning of the server 200 using measurement data from the vehicle. In other words, the control parameters can be automatically updated by automatic learning of the server 200 and searching for and updating the control parameters of the vehicle 100.

Control parameter updates may be performed automatically in this way, or manually by the user. The control parameter identifier may include a data format such as json or xml. A control parameter may have a one-to-one correspondence with an identifier, or multiple control parameters may correspond to one identifier.

Furthermore, it is common for the same equipment to be installed in multiple cars of the same train configuration. Therefore, the identifier for the control parameters may be, for example, a combination of the car number and a unique name, so that it can be updated for each car.

(Modification of the method of distributing to train configurations)
In general, it is possible to verify the suitability of a model generated by automatic learning to past data. However, the suitability to future data is unknown. Therefore, the performance of the second model and the second control parameters generated by the automatic learning system 1 cannot be known until they are applied to the vehicle 100. In such a situation, if the second control parameters are distributed simultaneously to all train configurations whose control parameters are to be updated, there is a risk that a major disruption will be caused to the train configuration if the performance of the second control parameters is poor.

Therefore, the automatic learning system 1 according to this modified example distributes the second control parameters only to some of the train configurations that have been registered in advance after approval from the railway operator. Then, after checking the performance of the second control parameters, it decides whether or not to distribute the second control parameters to the remaining train configurations to be updated.

Figure 9 is a flow diagram showing a modified example of a distribution method for a train configuration. Figures 10A to 10C are conceptual diagrams showing the update state of control parameters in a train configuration to be updated.

First, steps S10 to S120 in FIG. 3 are executed. At this time, in step S100, the control parameter output device 500 distributes the second control parameter to some of the train configurations to be updated. The first control parameters of some of the train configurations to be updated are updated with the second control parameter. For example, as shown in FIG. 10A, the first control parameter P1 of some of the train configurations T1 and T2 among the train configurations T1 to T5 is updated with the second control parameter P2.

Next, the data storage unit 310 collects and accumulates further update data from some of the train configurations that are controlled using the second control parameters (S130).

Next, the model performance information of the second model is calculated using the update data (S140). The method of calculating this model performance information may be the same as that of step S60 in FIG. 3. At this time, the learning execution unit 320 calculates the model performance information (also called updated model performance information) using the second model and the update data.

Next, the learning execution unit 320 compares the degree of conformance between the second model and the update data (i.e., the updated model performance information) with a threshold value (S150). If the updated model performance information is higher than the threshold value (YES in S150), this indicates that the performance of the second control parameters is good. Therefore, the control parameter output device 500 outputs and distributes the second control parameters to other train formations (train formations other than the updated train formation) among the train formations to be updated (S160). As a result, the first control parameters of all train formations to be updated are updated to the second control parameters. For example, as shown in FIG. 10B, all first control parameters P1 of train formations T1 to T5 are updated to the second control parameters P2.

On the other hand, if the degree of conformity between the second model and the update data (i.e., the update model performance information) is lower than the threshold value (NO in S150), this indicates that the performance of the second control parameters is poor. Therefore, the control parameter output device 500 does not output the second control parameters to other train configurations (train configurations other than the updated train configuration) among the train configurations to be updated. In addition, the parameter control device 109 of the vehicle 100 returns the second control parameters of the updated vehicle to the first control parameters (S170). For example, as shown in FIG. 10C, the second control parameters P2 of the train configurations T1 and T2 are returned to the first control parameters P1, and the first control parameters P1 are maintained for the train configurations T3 to T5. The information in FIG. 10A to FIG. 10C may be displayed on the display unit 600. This allows the user to know at a glance which train configurations or vehicles have updated or not updated control parameters.

(Visualization of parameter performance information)
The display unit 600 may display the parameter performance information in a graph for each of the first and second control parameters. For example, FIG. 11A is a diagram showing a definition of a graph display for the control parameter "param001". The display unit 600 displays the unique name of the control parameter, the X-axis of the graph, the Y-axis of the graph, and an input screen in a graph format. The user inputs these items. In FIG. 11A, for example, the relationship between the occupancy rate and humidity for the control parameter "param001" is input to be displayed in a scatter diagram.

As a result, as shown in Figures 11B and 11C, the display unit 600 displays a graph showing the relationship between occupancy rate and humidity for each of the first and second control parameters P1, P2 of the control parameter "param001". Figure 11B is a graph showing parameter performance information for the first control parameter P1. Figure 11C is a graph showing parameter performance information for the second control parameter P2. The data in Figure 11B is measurement data when the first control parameter P1 is used, and the data in Figure 11C is measurement data when the second control parameter P2 is used.

As shown in FIG. 11B, with the first control parameter P1, as the occupancy rate increases, the humidity also increases. This means that the passengers feel uncomfortable, and that the control of the air conditioning device 130 is not appropriate. In other words, this means that the first control parameter P1 is not suited to the current state of the vehicle 100. On the other hand, as shown in FIG. 11C, with the second control parameter P2, even if the occupancy rate increases, the humidity does not change much and remains nearly constant. This means that the passengers do not feel uncomfortable, and that the control of the air conditioning device 130 is appropriate. In other words, this means that the second control parameter P2 is suited to the current state of the vehicle 100.

In this way, the display unit 600 visualizes the data corresponding to each version of the control parameter by displaying it in a graph. This allows the user to easily visually grasp the performance of the control parameter. The arrangement and layout of the graphs may be changed as desired. In addition to scatter plots, the graphs may be frequency distribution tables, box plots, line graphs, etc. Furthermore, the display unit 600 may display auxiliary lines, etc., superimposed on the graphs.

(Classification model)
In the above embodiment, the automatic learning system 1 uses a regression model as the learning algorithm. However, the automatic learning system 1 may use a classification model. For example, FIG. 12 is a diagram showing an example of a classification model that classifies variables that can take three classes C0 to C2 into three classes. In this example, a decision tree model is adopted as the classification model. Of course, this embodiment may adopt a classification model other than a decision tree model.

The conditions from the starting node to class C2 are taken as control parameters. In this case, the learning execution unit 320 and the control parameter calculation unit 410 extract the branch conditions present on the path from the starting node to the terminal node class C2 as a series of control parameters. The model and control parameter relationship information 344 is a program that extracts such control parameters. If the unique name of this control parameter is, for example, "param999" and the delimiter for each element is ":", the control parameter is expressed by Equation 3.
param999 = a>3:b>20:a>10 (Equation 3)
In a decision tree model, the same variables may be used repeatedly in multiple layers, so control parameters cannot be expressed using a normalized table. Therefore, it is necessary to store the branch elements on the route from the start node to the terminal node as an array.

As described above, the automatic learning system 1 according to this embodiment automatically learns data collected while the train formation is running, generates a regression model or a classification model, and uses the model to generate new second control parameters. By periodically and continuously generating these second control parameters and updating the existing first control parameters, the equipment of the vehicle 100 can be controlled with control parameters that are suited to the environment and situation at that time.

The control parameters are set in advance in the parameter storage unit 110 of the vehicle 100, and each has a unique name (identifier). This allows existing control parameters to be updated (upgraded) to new control parameters based on the unique names of the control parameters. The vehicle control device 120 controls the equipment of the vehicle 100 using the control parameters stored in the parameter storage unit 110. Therefore, when updating the control parameters, the user does not need to change the control program itself of the vehicle control device 120, and the server 200 can automatically update or maintain the control parameters to appropriate values through automatic learning.

The measurement data collected from the vehicle 100 is stored in the data storage unit 310 in binary format. Therefore, the data capacity of the data storage unit 310 can be reduced.

Furthermore, the automatic learning system 1 of this embodiment applies the new second control parameters to a portion of the train formations, and verifies the suitability of the second control parameters (i.e., the second model) from the update data collected from the portion of the train formations. The automatic learning system 1 then applies the second control parameters to all train formations that are to be updated. This makes it possible to minimize the impact and risk on the operation and maintenance of the train formations, even if the second control parameters are not actually appropriate for the vehicle 100.

Second Embodiment
Fig. 13 is a block diagram showing an example of the configuration of an automatic learning system 1 according to the second embodiment. In the first embodiment, the server 200 is installed on the ground. In contrast, in the second embodiment, the server 200 is mounted on the vehicle 100. Data generated by devices and sensors in the vehicle 100 is not transmitted to the outside, but is processed within the vehicle 100. The other configurations and operations of the second embodiment are similar to those of the first embodiment. Therefore, the second embodiment can obtain the same effects as the first embodiment.

At least a part of the data processing method in the automatic learning system 1 according to this embodiment may be configured with hardware or software. When configured with software, a program that realizes at least a part of the functions of the data processing method may be stored in a recording medium such as a flexible disk or CD-ROM, and may be read and executed by a computer. The recording medium is not limited to removable ones such as magnetic disks or optical disks, but may be fixed recording media such as hard disk drives or memories. In addition, a program that realizes at least a part of the functions of the data processing method may be distributed via a communication line (including wireless communication) such as the Internet. Furthermore, the program may be encrypted, modulated, or compressed and distributed via a wired or wireless line such as the Internet, or stored in a recording medium.

Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the gist of the invention. These embodiments and their modifications are within the scope of the invention and its equivalents as set forth in the claims, as well as the scope and gist of the invention.

1 Automatic learning system, 100 Vehicle, 200 Server, 109 Parameter control device, 110 Parameter storage unit, 120 Vehicle control device 120 Air conditioner, 140 Temperature and humidity sensor, 150 Load sensor, 300 Automatic learning device, 400 Control parameter generation device, 500 Control parameter output device, 600 Display unit

Claims (12)

1. An auto-learning system having a first control parameter for use in controlling a vehicle, the system comprising:
a first automatic learning device including: a data storage unit that accumulates data on the first control parameters collected from the vehicle during operation; a learning execution unit that analyzes the data to generate a regression model or a classification model; a parameter generation unit that generates new second control parameters using the regression model or the classification model; and an output unit that outputs the second control parameters ;
a second automatic learning device mounted on the vehicle, the second automatic learning device including: a parameter storage unit that stores the second control parameters from the first automatic learning device; a parameter control unit that updates the first control parameters with the second control parameters; and a vehicle control unit that controls the vehicle using the second control parameters; and
a learning item storage unit for storing learning method information including variables and algorithms of the regression model or the classification model, model performance information indicating a degree of compatibility between the regression model or the classification model and the data, learning start trigger information indicating a start condition for analysis of the data, and the regression model or the classification model ;
Equipped with an automatic learning system. the data storage unit stores the data in a compressed binary format;
2. The automatic learning system according to claim 1, wherein the learning execution unit restores only learning data related to the first and second control parameters from the binary format of the data, generates the regression model, and then discards the restored learning data. 3. The automatic learning system according to claim 2, wherein the learning execution unit restores the learning data from a binary format to a text format, and discards the learning data in the text format after generating the regression model. The automatic learning system according to claim 1 , wherein the learning start trigger information is a date and time when data analysis is to start, or a threshold value of the degree of suitability. 5. The automatic learning system according to claim 1, wherein the learning execution unit compares first model performance information indicating the goodness of fit of a first regression model or a first classification model used to generate the first control parameter with second model performance information indicating the goodness of fit of a second regression model or a second classification model used to generate the second control parameter. When the second model performance information is higher in the goodness of fit than the first model performance information, the parameter control unit updates the first control parameter with the second control parameter;
6. The automatic learning system of claim 5, wherein when the second model performance information is lower in the degree of fit than the first model performance information, the parameter control unit does not update the first control parameter , and the vehicle control unit controls the vehicle using the first control parameter. The automatic learning system according to claim 1, wherein the output unit transmits the second control parameter to the second automatic learning device, or outputs the second control parameter to an electronic recording medium communicably connected to the output unit. The automatic learning system according to claim 1 , further comprising a display unit that displays whether the output unit is permitted or prohibited from outputting the second control parameter. the first automatic learning device is communicatively connected to a plurality of train configurations made up of the vehicles,
9. The automatic learning system according to claim 5 , wherein the output unit outputs the second control parameters to a preset portion of the plurality of train configurations. the data storage unit further collects and accumulates update data from the part of the train configurations controlled using the second control parameters;
10. The automatic learning system according to claim 9, wherein when a degree of fit between a second regression model or a second classification model used to generate the second control parameter and the update data is higher than a threshold, the output unit outputs the second control parameter to a vehicle other than an update vehicle updated with the second control parameter. the data storage unit further collects and accumulates update data from the part of the train configurations controlled using the second control parameters;
The automatic learning system according to claim 10, wherein when a degree of fit between the second regression model or the second classification model and the update data is lower than a threshold value, the output unit does not output the second control parameter to vehicles other than the update vehicle, and the parameter control unit returns the second control parameter of the update vehicle to the first control parameter. An automatic learning system as described in any one of claims 1 to 9, further comprising a display unit that displays the first or second control parameters used in each of the vehicles in comparison with data collected from the vehicles related to the first or second control parameters.

Images