Machine learning device

Abstract

The present disclosure relates to a machine learning device. A machine learning device that learns a model unique to a machine learning vehicle (2) includes a learning unit (146, 333) that performs machine learning of the model using a training data set including values of state parameters detected by a detector provided to the vehicle, and a parameter value acquisition unit (145, 332) that acquires values of the state parameters detected by other vehicles in a case where the values of the state parameters detected by the detector have been abnormal. The learning unit performs machine learning of the model using a training data set including values acquired by the parameter value acquisition unit from other vehicles instead of the values of the state parameters detected by the detector in a case where the values of the state parameters detected by the detector have been abnormal. Thus, the model can be properly learned even if the values of the state parameters constituting the training data set have been abnormal.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to a machine learning device that machine learns a learning model unique to a vehicle. BACKGROUND

[0002] It is known to acquire values of state parameters associated with a vehicle and to machine learn a learning model using the acquired values of the state parameters as a training data set (for example, Patent Literature 1). In particular, in Patent Literature 1, machine learning of a learning model that outputs a temperature of an exhaust purification catalyst is performed based on a training data set that contains values of an engine rotation speed, an engine load factor, and the temperature of the exhaust purification catalyst.

[0003] PRIOR ART DOCUMENTS

[0004] Patent Literature 1: Japanese Patent Application Publication No. 2019-183698 SUMMARY

[0005] PROBLEMS TO BE SOLVED BY THE INVENTION

[0006] For example, in a case where a value of a state parameter that constitutes a training data set is detected by a sensor, if the sensor is abnormal, the value of such a state parameter cannot be accurately detected. In a case where a value of a state parameter that constitutes a training data set is abnormal, if a learning model is machine learned using such a training data set, the learning model cannot be properly learned.

[0007] In view of the above problems, an object of the present disclosure is to provide a machine learning device that can properly machine learn a learning model even in a case where a value of a state parameter that constitutes a training data set for machine learning of the learning model is abnormal.

[0008] TECHNICAL SOLUTION TO THE PROBLEM

[0009] The gist of the present disclosure is as follows.

[0010] (1) A machine learning device that machine learns a learning model unique to a vehicle, comprising: a learning unit that performs learning of the learning model using a training data set containing values of state parameters detected by a detector provided to the vehicle; and a parameter value acquisition unit that acquires values of the state parameters detected by other vehicles under conditions that match detection conditions at the time when the values of the state parameters contained in the training data set are detected by the detector, in a case where the values of the state parameters detected by the detector are abnormal, the learning unit performing learning of the learning model using a training data set that contains values acquired by the parameter value acquisition unit from other vehicles instead of the values of the state parameters detected by the detector, in a case where the values of the state parameters detected by the detector are abnormal.

[0011] (2) According to the machine learning device described in (1) above, the machine learning device is installed in the vehicle, and the parameter value acquisition unit acquires the value of the state parameter detected by the other vehicle through vehicle-to-vehicle communication or through communication with a server that can communicate with the other vehicle.

[0012] (3) According to the machine learning apparatus described in (1) above, the machine learning apparatus is provided on a server capable of communicating with the vehicle and the other vehicles, and further includes a model sending unit that sends information of the learning model learned by the learning unit to the vehicle, the parameter value acquisition unit obtains the value of the state parameter detected by the other vehicle from the other vehicle through communication between the other vehicle and the server, and the learning unit uses a training dataset containing the value of the state parameter obtained from the other vehicle and the values ​​of other state parameters associated with the vehicle other than the state parameter sent from the vehicle through communication between the vehicle and the server to learn the learning model.

[0013] (4) The machine learning apparatus according to any one of (1) to (3) above, wherein the state parameter is a physical quantity that has a small difference between vehicles under the same detection conditions compared with other state parameters associated with the vehicle.

[0014] (5) The machine learning apparatus according to any one of (1) to (4) above, wherein the state parameter is a state parameter associated with the external environment of the vehicle.

[0015] (6) According to the machine learning apparatus described in (5) above, the state parameters include at least one of the temperature, humidity and atmospheric pressure of the atmosphere surrounding the vehicle.

[0016] (7) The machine learning apparatus according to any one of (1) to (6) above, wherein the detection conditions include the time and location at which the detector detects the values ​​of the state parameters constituting the training dataset of the vehicle.

[0017] Invention Effects

[0018] According to this disclosure, a machine learning apparatus is provided that can properly machine learn a learning model even when the values ​​of the state parameters constituting the training dataset for the machine learning learning model are abnormal. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the machine learning system involved in the first embodiment.

[0020] Figure 2is a diagram schematically showing a hardware structure of a vehicle.

[0021] Figure 3 is a functional block diagram of a processor of the host vehicle (self vehicle).

[0022] Figure 4 is a diagram schematically showing a hardware structure of a server.

[0023] Figure 5 is a functional block diagram of a processor of the server.

[0024] Figure 6 represents an example of a NN model having a simple structure (configuration).

[0025] Figure 7 is an action sequence diagram of a learning process by the machine learning system.

[0026] Figure 8 is a functional block diagram of a processor of the host vehicle.

[0027] Figure 9 is a functional block diagram of a processor of the server.

[0028] Figure 10 is an action sequence diagram of a learning process by the machine learning system.

[0029] Explanation of Reference Signs

[0030] 1 machine learning system; 2 vehicle; 3 server; 11 ECU (Electronic Control Unit); 14 processor; 33 processor. DETAILED DESCRIPTION

[0031] Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. In addition, in the following description, the same components are given the same reference numerals.

[0032] <First Embodiment>

[0033] Configuration of Machine Learning System

[0034] First, with reference to Figures 1-7 , the machine learning system 1 involved in the first embodiment will be described. Figure 1 is a diagram schematically showing a hardware structure of a vehicle.

[0035] As Figure 1As shown, the machine learning system 1 has a plurality of vehicles 2 and a server 3 that are capable of communication. Each of the plurality of vehicles 2 and the server 3 are configured to be capable of communication with each other via a communication network 4 and a wireless base station 5, the communication network 4 being constituted by an optical communication line or the like, and the wireless base station 5 being connected to the communication network 4 via a gateway (not shown). The communication between the vehicle 2 and the wireless base station 5 is communication in compliance with an arbitrary communication standard. Further, in the following description, one of the vehicles 2 in which a learning model that has been subjected to machine learning by the machine learning system 1 is used will be referred to as a host vehicle 2a, and the vehicles other than the host vehicle 2a will be referred to as other vehicles 2b.

[0036] Figure 2 is a diagram schematically showing the hardware structure of the vehicle 2. As shown, the vehicle 2 has an electronic control unit (ECU) 11. The ECU 11 has an in-vehicle communication interface 12, a memory 13, and a processor 14. The in-vehicle communication interface 12 and the memory 13 are connected to the processor 14 via signal lines. Further, in the present embodiment, one ECU 11 is provided in the vehicle 2, but a plurality of ECUs divided by function can also be provided. Figure 2

[0037] The in-vehicle communication interface 12 has an interface circuit for connecting the ECU 11 to an in-vehicle network 15 in compliance with a standard such as CAN (Controller Area Network). The ECU 11 communicates with other in-vehicle devices via the in-vehicle communication interface 12.

[0038] The memory 13 is an example of a storage section that stores data. The memory 13 has, for example, a volatile semiconductor memory (e.g., a RAM) and a non-volatile semiconductor memory (e.g., a ROM). The memory 13 stores a computer program for executing various processes in the processor 14, and various data used when the processor 14 executes various processes, and the like. Thus, the memory 13 stores a learning model.

[0039] The processor 14 has one or a plurality of CPUs (Central Processing Units) and peripheral circuits thereof. The processor 14 can also have a GPU (Graphics Processing Unit) or an arithmetic circuit such as a logical arithmetic unit or a numerical arithmetic unit. The processor 14 executes various processes based on the computer program stored in the memory 13. Thus, when a value of an input parameter of a learning model is input, the processor 14 performs an arithmetic process in accordance with the learning model and outputs a value of an output parameter.

[0040] Figure 3 is a functional block diagram of the processor 14 of the host vehicle 2a. As shown, Figure 3 ​As shown, the processor 14 is provided with a control section 141 that controls the control devices 22 of the host vehicle 2a using a learning model, a data transmission section 142 that transmits a training data set for machine learning of the learning model to the server 3, and a model update section 143 that updates the learning model used in the control section 141. These functional blocks possessed by the processor 14 are, for example, functional modules realized by a computer program operating on the processor 14. Alternatively, these functional blocks possessed by the processor 14 can be dedicated arithmetic circuits provided to the processor 14. Details of each functional block of the processor 14 of the host vehicle 2a will be described later.

[0041] In addition, as shown in FIG. 1, the vehicle 2 is provided with a plurality of sensors 23. The sensors 23 are connected to the ECU 11 via the in-vehicle network 15. Figure 2 As shown, the vehicle 2 is further provided with an off-vehicle communication module 21, a plurality of control devices 22, and a plurality of sensors 23. The off-vehicle communication module 21, the control devices 22, and the sensors 23 are connected to the ECU 11 via the in-vehicle network 15.

[0042] The off-vehicle communication module 21 is an example of a communication section that communicates with devices outside the vehicle. The off-vehicle communication module 21 is, for example, a device for communicating with the server 3 and other vehicles 2b. The off-vehicle communication module 21 includes, for example, a data communication module (DCM: Data communication module). The data communication module communicates with the server 3 via the wireless base station 5 and the communication network 4.

[0043] The control devices 22 are devices that perform various controls related to the vehicle 2. Specifically, the control devices 22 include, for example, a drive actuator of a throttle valve that adjusts an opening degree of a throttle valve provided in an intake passage of an internal combustion engine, an injector that supplies fuel to a combustion chamber of the internal combustion engine, a drive actuator of an EGR valve that controls an EGR rate of the internal combustion engine, a blower of an air conditioning device, a drive actuator of an air mixer door that controls air flow of the air conditioning device, and the like. These control devices 22 are connected to the ECU 11 via the in-vehicle network 15 and operate according to a drive signal from the ECU 11.

[0044] The sensor 23 is an example of a detector that detects values (state quantities) of various state parameters relating to the vehicle 2. The sensor 23 includes, for example, an air amount sensor that detects an intake air amount supplied to an internal combustion engine, an injection pressure sensor that detects a fuel injection pressure of an injector from the internal combustion engine, an exhaust gas temperature sensor that detects a temperature of exhaust gas (exhaust), an input detection sensor that detects an input of a driver on a touch panel or the like, a self-position sensor (e.g., GPS) that detects a self-position of the vehicle 2, and the like. Further, the sensor 23 includes, for example, an outside air temperature sensor that detects a temperature of air around the vehicle 2 (outside air temperature), an outside air humidity sensor that detects a humidity of air around the vehicle 2 (outside air humidity), an atmospheric pressure sensor that detects an atmospheric pressure around the vehicle 2, an indoor temperature sensor that detects a temperature of an indoor of the vehicle 2 (indoor temperature), an indoor humidity sensor that detects a humidity of the indoor of the vehicle 2 (indoor humidity), a sunlight sensor that detects a sunlight amount, and the like. These sensors 23 are connected to the ECU 11 via the in-vehicle network 15, and transmit output signals to the ECU 11.

[0045] The server 3 is provided outside the vehicle 2, and communicates with the vehicles 2 that are traveling in a specific region via the communication network 4 and the wireless base station 5. The server 3 receives various information from the vehicles 2 that are traveling in the specific region.

[0046] Figure 4 is a diagram that schematically shows a hardware structure of the server 3. As shown in Figure 4 , the server 3 is provided with an external communication module 31, a storage device 32, and a processor 33. In addition, the server 3 can have an input device such as a keyboard and a mouse, and an output device such as a display.

[0047] The external communication module 31 is an example of a communication section that communicates with devices outside the server 3. The external communication module 31 is provided with an interface circuit that connects the server 3 to the communication network 4. The external communication module 31 is configured to be able to communicate with a plurality of vehicles 2 respectively via the communication network 4 and the wireless base station 5.

[0048] The storage device 32 is an example of a storage section that stores data. The storage device 32 is provided with, for example, a hard disk drive (HDD), a solid state drive (SSD), or an optical recording medium. The storage device 32 stores a computer program for executing various processes by the processor 33, and various data used when the processor 33 executes various processes.

[0049] The processor 33 has one or a plurality of CPUs and peripheral circuits thereof. The processor 33 can also have a GPU or an arithmetic circuit such as a logic operation unit or a numerical operation unit. The processor 33 executes various processes on the basis of a computer program stored in the storage 32. In the present embodiment, the processor 33 of the server 3 functions as a machine learning device that performs machine learning of a learning model.

[0050] Figure 5 is a functional block diagram of the processor 33 of the server 3. As shown in Figure 5 , the processor 33 includes an abnormal parameter determination section 331 that determines an abnormal parameter in which an abnormality has occurred in a state parameter constituting a training data set transmitted from the host vehicle 2a, a parameter value acquisition section 332 that acquires a value of the abnormal parameter from another vehicle 2b different from the host vehicle 2a, a learning section 333 that performs machine learning of a learning model, and a model transmission section 334 that transmits the machine-learned learning model to the host vehicle 2a. These functional blocks included in the processor 33 are, for example, functional modules realized by a computer program operating on the processor 33. Alternatively, these functional blocks included in the processor 33 can be dedicated arithmetic circuits provided to the processor 33. Details of each functional block of the processor 33 of the server 3 will be described later.

[0051] Learning Model

[0052] In the present embodiment, in the control section 141 of the host vehicle 2a, when the control device 22 mounted on the host vehicle 2a is controlled, a learning model learned through machine learning is used. In the present embodiment, as the learning model, a neural network (hereinafter referred to as "NN model") model is used. Hereinafter, an outline of the NN model will be described with reference to Figure 6 . Figure 6 An example of an NN model having a simple structure is shown in

[0053] Figure 6 The circles in Figure 6 indicate artificial neurons. The artificial neurons are generally referred to as nodes or units (referred to as "nodes" in the present specification). In , L = 1 indicates an input layer, L = 2 and L = 3 indicate hidden layers (or intermediate layers), and L = 4 indicates an output layer.

[0054] In Figure 6 , x1 and x2 indicate each node of the input layer (L = 1) and an output value from the node, and y indicates a node of the output layer (L = 4) and an output value thereof. Similarly, z1 (L＝2) , z2 (L＝2) , and z3 (L＝2) indicate each node of the hidden layer (L = 2) and an output value from the node, and z1(L＝3) and z2 (L＝3) This represents each node in the hidden layer (L=3) and the output value from that node.

[0055] At each node of the input layer, the input is output as is. On the other hand, at each node of the hidden layer (L=2), the output values ​​x1 and x2 of the input layer are used as input, and at each node of the hidden layer (L=2), the total input value u is calculated using its corresponding weight w and bias b. For example, in... Figure 6 In the hidden layer (L=2), z k (L＝2) The total input value u calculated at each node shown in (k = 1, 2, 3) k (L＝2) It becomes the following formula (M is the number of nodes in the input layer).

[0056]

[0057] Next, the total input value u k (L＝2) The transformation is achieved through the activation function f, from the hidden layer (L=2) by z. k (L＝2) The node shown is used as the output value z k (L＝2) (=f(u k (L＝2) The output is as follows: On the other hand, the output values ​​z1 of each node in the hidden layer (L=3) are input to each node in the hidden layer (L=2). (L＝2) z2 (L＝2) and z3 (L＝2) At each node of the hidden layer (L=3), the total input value u (=Σz·w+b) is calculated using its corresponding weight w and bias b. This total input value u is also transformed through an activation function, and then used as the output value z1 from each node of the hidden layer (L=3). (L＝3) z2 (L＝3) Output. The activation function is, for example, the ReLU function (linear rectified function) σ.

[0058] Additionally, the output values ​​z1 of each node in the output layer (L=4) are input to the nodes in the hidden layer (L=3). (L＝3) and z2 (L ＝3) At the nodes of the output layer, the total input value u(Σz·w+b) is calculated using its respective weight w and bias b, or simply using its respective weight w. For example, the identity function can be used as the activation function at the nodes of the output layer. In this case, the total input value u calculated at the nodes of the output layer is output as the output value y from the nodes of the output layer.

[0059] Thus, the NN model has an input layer, a hidden layer, and an output layer, and when one or more input parameters are input from the input layer, one or more output parameters corresponding to the input parameters are output from the output layer.

[0060] In the present embodiment, as such a learning model, for example, a model that outputs the temperature of the exhaust gas as the value of the output parameter when the values of the outside air temperature, the intake air amount, the fuel injection amount, the fuel injection timing, the fuel injection pressure, and the EGR rate are input as the values of the input parameters is used. In the control portion 141 of the present vehicle 2a, the temperature of the exhaust gas is output by inputting the values of the respective input parameters detected by the sensors 23 to such a learning model. The control portion 141 controls the control devices 22 related to the internal combustion engine on the basis of the output temperature of the exhaust gas. Here, since the exhaust gas temperature sensor that detects the temperature of the exhaust gas has a response delay, if the control of the internal combustion engine is performed on the basis of the output of the exhaust gas temperature sensor, the internal combustion engine cannot necessarily be controlled appropriately. In contrast, since the temperature of the exhaust gas is calculated using the learning model without delay, the internal combustion engine can be controlled more appropriately by using the temperature of the exhaust gas calculated by the learning model to control the internal combustion engine.

[0061] Alternatively, as such a learning model, for example, a model that outputs the target temperature of the air conditioning device as the value of the output parameter when the values of the outside air temperature, the inside air temperature, the inside air humidity, and the amount of sunlight are input as the values of the input parameters can also be used. In this case, in the control portion 141 of the present vehicle 2a, the target temperature of the air conditioning device is output by inputting the values of the respective input parameters detected by the sensors 23 to such a learning model. The ECU 11 controls the control devices 22 related to the air conditioning device so that the inside air temperature becomes the target temperature output from the learning model.

[0062] Further, as the learning model, various models can be used. Thus, as the input parameters, various state parameters that indicate the state of the vehicle, such as the outside air temperature, the outside air humidity, the atmospheric pressure, the inside air temperature, the inside air humidity, the amount of sunlight, the intake air amount, the intake air temperature, the fuel injection pressure, the fuel injection timing, the fuel injection amount, the air-fuel ratio, the ignition timing, the engine cooling water temperature, and the supercharging pressure can be used. In addition, as the output parameters, various state parameters that indicate the state of the vehicle, such as the temperature of the exhaust gas purification catalyst, the concentration of NOx in the exhaust gas, the output torque of the internal combustion engine, and the inside air humidity can be used.

[0063] Basic Learning of Learning Model

[0064] Next, the machine learning of the learning model (NN model) described above will be described. In order to improve the accuracy of such an NN model, machine learning of the NN model needs to be performed. Therefore, in the present embodiment, the learning unit 333 of the server 3 performs machine learning of the NN model. First, the learning method of the NN model performed in the learning unit 333 will be described simply.

[0065] In the machine learning of the NN model, a training data set including values of the state parameters detected by the sensors 23 provided to the vehicle 2 is used. The training data set is composed of a combination of a plurality of measured values of a plurality of input parameters and a plurality of measured values (correct answer data) of at least one output parameter corresponding to these measured values. In the present embodiment, the measured values of the input parameters and the measured values of the output parameters are values detected by the sensors 23 of the host vehicle 2a or control command values from the ECU 11 to the control device 22. In addition, in the present embodiment, when the machine learning of the NN model is performed in the server 3, the measured values used as the training data set are transmitted from the host vehicle 2a to the server 3.

[0066] The learning unit 333 of the server 3 performs machine learning of the NN model after pre-processing (normalization, standardization, etc.) of the training data set transmitted from the host vehicle 2a. In performing the machine learning of the NN model, the learning unit 333 repeatedly updates the weights w and the bias b in the NN model so that the difference between the output value of the NN model and the measured value of the output parameter included in the training data set becomes small, for example, by the well-known error back propagation method. As a result, the NN model is learned, and a learned NN model is generated. The information of the learned NN model (the configuration of the model, the weights w, the bias b, etc.) is stored in the storage device 32 of the server, and is transmitted from the server 3 to the host vehicle 2a.

[0067] "Learning in the case where the value of the state parameter is abnormal"

[0068] When the sensor 23 that detects the value of the state parameter constituting the training data set, for example, has a failure or the like, the output of the sensor 23 can have an error, or the sensor 23 can become not to output a signal. If the machine learning of the learning model is performed using the output of the sensor 23 that has such an abnormality, the machine learning of the learning model cannot be performed properly.

[0069] Therefore, in the present embodiment, in the case where the value of the state parameter detected by the sensor 23 is abnormal, the value of the state parameter is acquired from another vehicle 2b in the same region as the host vehicle 2a, and the machine learning of the learning model is performed using a training data set including the value acquired from the other vehicle 2b instead of the value of the state parameter detected by the sensor 23. Hereinafter, the case where the value of the state parameter detected by the sensor 23 is abnormal will be described with reference to FIG. 6. Figure 7, specifically, machine learning of the learning model in a case where the value of the state parameter detected by the sensor 23 has become abnormal is described.

[0070] Figure 7 is an action timing chart of the learning process by the machine learning system 1. Specifically, Figure 7 is an action timing chart when machine learning of the learning model used in controlling the control device 22 of the host vehicle 2a is performed in a case where the outside air temperature sensor of the host vehicle 2a has become abnormal.

[0071] As shown in Figure 7 , the other vehicle 2b periodically detects the value of the state parameter including the outside air temperature and the detection condition at that time, respectively (step S11). The value of the state parameter is detected by the sensor 23 of the other vehicle 2b. In particular, the outside air temperature among the state parameters is detected by the outside air temperature sensor. In addition, the detection condition indicates the condition at the time when the value of each state parameter is detected by the sensor 23, and in the present embodiment, includes the time and the place at which the value of each state parameter is detected. The time at which the value of each state parameter is detected is acquired from the ECU 11 of the other vehicle 2b which functions as a clock. In addition, the place at which the value of each state parameter is detected is detected by the own position sensor. Furthermore, the detection condition can also include, for example, the amount of precipitation detected by a rain sensor, the style of the vehicle, and various other conditions.

[0072] The processor 14 of the other vehicle 2b causes the memory 13 to store the detected value of each state parameter and the detection condition of each state parameter. Furthermore, the data transmission part 142 of the processor 14 of the other vehicle 2b transmits the detection condition and the kind of the detected state parameter stored in the memory 13 to the server 3 via the external communication module 21 at a certain time interval. The server 3 stores the received detection condition and the kind of the state parameter in the storage device 32.

[0073] On the other hand, the host vehicle 2a periodically detects the value of the state parameter for machine learning of the learning model and the detection condition at that time (step S12). The value of the state parameter is detected by the sensor 23 of the host vehicle 2a, and the detection condition is acquired from the ECU 11 or detected by the sensor 23. Furthermore, in the example shown in Figure 7 , since the outside air temperature sensor of the host vehicle 2a has become abnormal, a large error occurs in the outside air temperature detected by the outside air temperature sensor.

[0074] The processor 14 of the host vehicle 2a causes the memory 13 to store the values of the time series of the plurality of detected state parameters and the detection conditions of the respective values of the state parameters. Further, when the required number of values of the state parameters for machine learning are detected and stored in the memory 13, the data transmission section 142 of the host vehicle 2a transmits the values of the time series of the plurality of state parameters stored in the memory 13 as a training data set to the server 3, and transmits the detection conditions of the respective values of the state parameters to the server 3. The server 3 stores the received training data set and the detection conditions in the storage device 32. Further, the data transmission section 142 of the host vehicle 2a can also transmit the values of the state parameters and the detection conditions to the server 3 each time they are detected.

[0075] When the required number of values of the state parameters for machine learning are stored in the storage device 32 as a training data set, the abnormal parameter determination section 331 of the server 3 determines whether there is a state parameter for which an abnormality has occurred in the training data set, and determines the state parameter in the case where there is a state parameter for which an abnormality has occurred (step S13). The abnormal parameter determination section 331 determines that an abnormality has occurred in the value of the state parameter in the case where the value of the state parameter is an abnormal value such as an impossible value (for example, in the case where the outside air temperature is detected to be more than 50°C in winter), in the case where the value of the state parameter is fixed to a certain value (for example, in the case where the outside air temperature is detected to be fixed to 25°C during the day), or in the case where the value of the state parameter is missing because no signal is output from the sensor. In the case where it is determined in the abnormal parameter determination section 331 that there is no state parameter for which an abnormality has occurred, the server 3 performs machine learning of the learning model in step S16 described later.

[0076] Further, in the present embodiment, the abnormal parameter determination section 331 of the server 3 determines the parameter for which an abnormality has occurred. However, the parameter for which an abnormality has occurred can also be determined using another method. For example, in the case where the processor 14 of the host vehicle 2a is able to perform failure diagnosis of the sensors 23 mounted on the host vehicle 2a, the state parameter detected by the sensor 23 diagnosed as having failed can also be determined as the state parameter for which an abnormality has occurred. In this case, the processor 14 of the host vehicle 2a functions as the abnormal parameter determination section 331, and the data transmission section 142 transmits the type of the parameter for which an abnormality has occurred to the server 3 in addition to the values of the state parameters and the detection conditions.

[0077] When the state parameter for which an abnormality has occurred is determined by the abnormal parameter determination section 331 (hereinafter referred to as an "abnormal occurrence state parameter"), the server 3 determines the type of the abnormal occurrence state parameter (step S14). The type of the abnormal occurrence state parameter is determined by the abnormal parameter determination section 331 based on the detection conditions of the respective values of the state parameters. For example, in the case where the value of the outside air temperature is detected to be more than 50°C in winter, the abnormal parameter determination section 331 determines that the abnormal occurrence state parameter is the outside air temperature. Further, in the case where the value of the outside air temperature is detected to be fixed to 25°C during the day, the abnormal parameter determination section 331 determines that the abnormal occurrence state parameter is the outside air temperature. Further, in the case where the value of the outside air temperature is missing because no signal is output from the sensor, the abnormal parameter determination section 331 determines that the abnormal occurrence state parameter is the outside air temperature. Figure 7In the example shown, the parameter value acquisition section 332 of the server 3 acquires the value of the abnormality occurrence state parameter from the other vehicle 2b when the value of the state parameter detected by the sensor 23 is abnormal. In particular, in the present embodiment, the parameter value acquisition section 332 acquires the value of the abnormality occurrence state parameter detected by the other vehicle 2b under conditions that coincide with the detection conditions when the value of the state parameter constituting the training data set is detected in the host vehicle 2a.

[0078] Therefore, in a case where it is determined that the value of the state parameter detected by the sensor 23 is abnormal, the parameter value acquisition section 332 of the server 3 determines the other vehicle 2b from which the value of the abnormality occurrence state parameter is to be acquired (step S14). The parameter value acquisition section 332 searches for the other vehicle 2b that has transmitted the detection conditions coinciding with the detection conditions when the value of the abnormality occurrence state parameter is detected in the host vehicle 2a, from among the many other vehicles 2b that have transmitted the detection conditions. Specifically, for example, the other vehicle 2b that has detected the state parameter at the time and place that most closely coincide with the time and place when the value of the state parameter included in the training data set is detected in the host vehicle 2a is determined as the other vehicle 2b from which the value of the abnormality occurrence state parameter is to be acquired.

[0079] When the other vehicle 2b from which the value of the abnormality occurrence state parameter is to be acquired is determined, the parameter value acquisition section 332 requests the other vehicle 2b to transmit the value of the abnormality occurrence state parameter (i.e., the outside air temperature). In particular, the parameter value acquisition section 332 transmits a signal requesting the determined other vehicle 2b to transmit the value of the abnormality occurrence state parameter detected by the other vehicle 2b under the detection conditions coinciding with the detection conditions of the training data set, to the other vehicle 2b via the external communication module 31.

[0080] When requested to transmit the value of the abnormality occurrence state parameter, the processor 14 of the other vehicle 2b acquires the value of the abnormality occurrence state parameter detected under the detection conditions coinciding with the detection conditions of the training data set, from the memory 13 (step S15). Furthermore, the processor 14 of the other vehicle 2b transmits the acquired value of the abnormality occurrence state parameter to the server 3 via the vehicle external communication module 21. The parameter value acquisition section 332 of the server 3 acquires the value of the abnormality occurrence state parameter detected by the other vehicle 2b in this way via the communication between the other vehicle 2b and the server 3. Therefore, in a case where the value of a certain state parameter (abnormality occurrence state parameter) detected by the sensor 23 is abnormal, the parameter value acquisition section 332 of the server 3 acquires the value of the abnormality occurrence state parameter detected by the other vehicle 2b under conditions that coincide with the detection conditions when the value of the abnormality occurrence state parameter included in the training data set of the host vehicle 2a is detected by the sensor 23.

[0081] The learning unit 333 of the server 3 performs machine learning of the learning model using the training data set made by the host vehicle 2a and / or the values of the abnormality occurrence state parameters transmitted from the other vehicle 2b (step S16). When it is determined in the abnormality parameter determination unit 331 that there is no abnormality occurrence state parameter, the learning unit 333 performs machine learning of the learning model using only the training data set transmitted from the host vehicle 2a by the above-described learning method. On the other hand, when it is determined in the abnormality parameter determination unit 331 that there is an abnormality occurrence state parameter, the learning unit 333 performs machine learning of the learning model using the training data set that contains the values of the abnormality occurrence state parameters acquired by the parameter value acquisition unit 332 from the other vehicle 2b instead of the values of the abnormality occurrence state parameters detected by the sensor 23 of the host vehicle 2a, by the above-described learning method. That is, the learning unit 333 performs machine learning of the learning model using, in this case, the training data set that contains the values of the abnormality occurrence state parameters acquired by the parameter value acquisition unit 332 from the other vehicle 2b and the values of the other state parameters than the abnormality occurrence state parameters detected by the sensor 23 of the host vehicle 2a and transmitted from the host vehicle 2a. Specifically, in the present embodiment, the training data set contains data of the atmospheric temperature acquired by the parameter value acquisition unit 332 from the other vehicle 2b instead of data of the atmospheric temperature detected by the atmospheric temperature sensor of the host vehicle 2a, and contains data of the other state parameters (e.g., the amount of intake air, the fuel injection pressure) detected by the other sensors (e.g., the air flow sensor, the injection pressure sensor) of the host vehicle 2a and / or data of the other state parameters (e.g., the fuel injection amount, the fuel injection timing) as the control command values from the ECU 11 to the control device 22.

[0082] The learning unit 333 causes the storage device 32 of the server 3 to store information of the learned learning model when the machine learning of the learning model is completed. In addition, the model transmission unit 334 of the server 3 transmits information (e.g., the updated weight w and / or the bias b) of the learned learning model to the host vehicle 2a when the machine learning of the learning model is completed.

[0083] The model update unit 143 of the host vehicle 2a updates the information of the learning model stored in the memory 13 of the host vehicle 2a when it receives the information of the learned learning model (step S17). Thereby, the control unit 141 performs control of the control device 22 based on the updated learning model when it performs control of the control device 22 using the learning model thereafter.

[0084] <Effects and Modified Examples>

[0085] According to the above-described embodiment, for example, when the value of the state parameter constituting the training data set is abnormal due to an abnormality such as a failure of the sensor 23 of the host vehicle 2a, the machine learning of the learning model is performed using the training data set in which the value of the abnormality-occurrence state parameter detected by the other vehicle 2b is included instead of the abnormal value of the abnormality-occurrence state parameter. As a result, according to the above-described embodiment, even when the value of the state parameter detected by the sensor 23 of the host vehicle 2a is abnormal, the machine learning of the learning model can be performed using the training data set in which the value of the state parameter is relatively accurate. Therefore, even when the value of the state parameter constituting the training data set for the machine learning of the learning model is abnormal, the learning model can be appropriately machine learned.

[0086] In addition, in the present embodiment, the machine learning of the learning model is performed in the server 3. Therefore, it is not necessary to perform the learning processing with a high calculation load in the host vehicle 2a, and thus it is not necessary to make the processing capacity of the processor 14 of the host vehicle 2a too high, thereby enabling the reduction of the manufacturing cost of the host vehicle 2a.

[0087] Further, in the above-described embodiment, an example in which the machine learning of the learning model is performed when the outside air temperature detected by the outside air temperature sensor is abnormal due to a failure of the outside air temperature sensor is described. However, the above-described learning method of the learning model can also be applied to a case in which the value of the state parameter other than the outside air temperature is abnormal due to a failure of the sensor 23 other than the outside air temperature sensor.

[0088] In particular, the above-described learning method of the learning model can be applied to a case in which the value of the state parameter that indicates the same value even when the vehicles 2 are different under the same detection condition is abnormal. Therefore, the above-described learning method of the learning model can be applied when the value detected by the sensor 23 of the host vehicle 2a is abnormal for the state parameter of the physical quantity that indicates a smaller difference between the vehicles 2 under the same detection condition than other state parameters among the state parameters associated with the host vehicle 2a. In particular, the outside environment around the host vehicle 2a is almost the same between the host vehicle 2a and the other vehicle 2b around the host vehicle 2a, and thus the above-described learning method of the learning model can be applied when the value of the state parameter associated with the outside environment around the host vehicle 2a detected by the sensor 23 of the host vehicle 2a is abnormal. As such a state parameter, in addition to the atmospheric temperature around the host vehicle 2a, for example, the humidity and the atmospheric pressure of the atmosphere around the host vehicle 2a can be listed.

[0089] In addition, in the above-described embodiment, the other vehicle 2b transmits the kind of the state parameter detected in step Sll and the detection condition to the server 3. However, the other vehicle 2b can also transmit the value of the state parameter detected in step Sll and the detection condition to the server 3. In this case, the value of the state parameter detected in the other vehicle 2b is stored in the storage device 32 of the server 3. Therefore, when the server 3 determines the other vehicle 2b in step S14, the server 3 does not request the other vehicle 2b to transmit the value of the abnormality occurrence state parameter, but acquires the value of the abnormality occurrence state parameter detected by the other vehicle 2b from the storage device 32.

[0090] Further, the learning model subjected to machine learning in the server 3 can also be a machine learning model other than a neural network, using a random forest, a k-nearest neighbor method, a support vector machine, or the like.

[0091] <Second Embodiment>

[0092] Next, the machine learning system 1 related to the second embodiment will be described with reference to Figures 8-10 The machine learning system 1 related to the second embodiment will be described. Hereinafter, mainly the differences from the machine learning system related to the first embodiment will be described.

[0093] In the present embodiment, the processor 14 of the host vehicle 2a functions as a machine learning device that performs machine learning on a learning model. Figure 8 is a functional block diagram of the processor 14 of the host vehicle 2a related to the second embodiment. As Figure 8 indicated, the processor 14 includes a control section 141 that controls the control device 22 of the host vehicle 2a using a learning model, an abnormal parameter determination section 144 that determines an abnormal state parameter in which an abnormality has occurred among the state parameters constituting the training data set created by the host vehicle 2a, a parameter value acquisition section 145 that acquires the value of the abnormal state parameter from the other vehicle 2b different from the host vehicle 2a, a learning section 146 that performs learning of the learning model, and a model update section 143 that updates the learning model used in the control section 141.

[0094] Figure 9 is a functional block diagram of the processor 33 of the server 3 related to the second embodiment. As Figure 9 indicated, the processor 33 includes a vehicle determination section 335 that determines the other vehicle 2b from which the value of the abnormality occurrence parameter is to be acquired, and a parameter value transmission section 336 that transmits the value of the abnormality occurrence state parameter acquired from the other vehicle 2b to the host vehicle 2a.

[0095] Figure 10 is an action timing chart of the learning processing performed by the machine learning system 1 related to the second embodiment. In particular, Figure 10This is a timing diagram of the actions performed when the external air temperature sensor of vehicle 2a malfunctions, and the learning model of the control device 22 used to control vehicle 2a is subjected to machine learning.

[0096] like Figure 10 As shown, with Figure 7 Similarly, in step S11, other vehicles 2b periodically detect the values ​​of status parameters including the outside air temperature and the current detection conditions, and store them in memory 13 (step S21). Moreover, the data transmission unit 142 of the processor 14 of other vehicles 2b sends the detection conditions stored in memory 13 and the types of detected status parameters to the server 3 via the vehicle external communication module 21 at certain time intervals.

[0097] On the other hand, with Figure 7 Similarly, in step S12, the vehicle 2a periodically detects the values ​​of the state parameters used for machine learning of the learning model and the detection conditions at that time, and stores them in memory 13 (step S22). When the processor 14 of the vehicle 2a detects the values ​​of the number of state parameters required for learning and stores them in memory 13, it summarizes the time series values ​​of the multiple state parameters stored in memory 13 to create a training dataset.

[0098] When creating the training dataset, and Figure 7 Similarly, in step S13, the abnormal parameter determination unit 144 of the processor 14 of the vehicle 2a determines whether there is an abnormal state parameter (abnormal occurrence state parameter) in the training dataset. If there is an abnormal occurrence state parameter (outdoor temperature in this example), the state parameter is determined (step S23).

[0099] When the anomaly occurrence state parameter is determined by the anomaly parameter determination unit 144, the parameter value acquisition unit 145 of the vehicle 2a acquires the value of the anomaly occurrence state parameter from the other vehicle 2b. In particular, in this embodiment, the parameter value acquisition unit 145 acquires the value of the anomaly occurrence state parameter detected by the other vehicle 2b under the conditions that match the detection conditions when the value of the state parameter constituting the training dataset is detected in the vehicle 2a.

[0100] Therefore, firstly, when the parameter value acquisition unit 145 determines that the state parameter detected by the sensor 23 has become abnormal, it sends the type of abnormal state parameter and the detection conditions to the server 3.

[0101] When the vehicle determination unit 335 of server 3 receives the type of abnormality status parameters and detection conditions from vehicle 2a, it... Figure 7Step S14 of the same as the above, the other vehicle 2b to be taken to determine the value of the abnormal occurrence state parameter (step S24). Vehicle determining section 335 from the sending of many other vehicles 2b to detect conditions search sent in the vehicle 2a in which the abnormal occurrence state parameter is detected in accordance with the detection conditions of the detection conditions of the other vehicle 2b that sent the detection conditions. Vehicle determining section 335 in the determination of the other vehicle 2b to be taken to determine the value of the abnormal occurrence state parameter, the other vehicle 2b to request the transmission of the value of the abnormal occurrence state parameter (i.e. the outside air temperature).

[0102] The processor 14 of the other vehicle 2b in the request to send the value of the abnormal occurrence state parameter, with Figure 7 Step S15 of the same as the above, from the memory 13 to take the value of the abnormal occurrence state parameter detected under the detection conditions in accordance with the detection conditions of the training data set (step S25). Moreover, the processor 14 of the other vehicle 2b to send the value of the abnormal occurrence state parameter to the server 3. The parameter value sending section 336 of the server 3 in the value of the abnormal occurrence state parameter received from the other vehicle 2b, the value of the abnormal occurrence state parameter to the vehicle 2a. The parameter value acquisition section 145 of the vehicle 2a as such via the server 3 capable of communication with the other vehicle 2b between the vehicle 2a, the value of the abnormal occurrence state parameter detected by the other vehicle 2b. Therefore, the parameter value acquisition section 145 of the vehicle 2a in the value of a certain state parameter (abnormal occurrence state parameter) detected by the sensor 23, the value of the abnormal occurrence state parameter detected by the other vehicle 2b under the conditions in accordance with the detection conditions of the training data set contained in the vehicle 2a in which the value of the abnormal occurrence state parameter is detected by the sensor 23.

[0103] With Figure 7Similarly to the step S16, the learning unit 146 of the host vehicle 2a performs machine learning of the learning model using the training data set made by the host vehicle 2a and / or the values of the abnormality occurrence state parameters transmitted from the other vehicle 2b (step S26). When it is determined in the abnormality parameter determination unit 144 that there is no abnormality occurrence state parameter, the learning unit 146 performs machine learning of the learning model using only the training data set made by the host vehicle 2a by the above-described learning method. On the other hand, when it is determined in the abnormality parameter determination unit 144 that there is an abnormality occurrence state parameter, the learning unit 146 performs machine learning of the learning model using the training data set containing the value of the abnormality occurrence state parameter acquired by the parameter value acquisition unit 145 from the other vehicle 2b instead of the value of the abnormality occurrence state parameter detected by the sensor 23 of the host vehicle 2a by the above-described learning method. That is, the learning unit 146 performs machine learning of the learning model using the training data set containing the value of the abnormality occurrence state parameter acquired by the parameter value acquisition unit 145 from the other vehicle 2b and the values of the other state parameters than the abnormality occurrence state parameter detected by the sensor 23 of the host vehicle 2a in this case.

[0104] The model update unit 143 of the host vehicle 2a updates the information (for example, the weight w and / or the bias b) of the learning model stored in the memory 13 of the host vehicle 2a when the machine learning of the learning model is completed (step S27). Thereby, the control unit 141 performs control of the control device 22 based on the updated learning model when performing control of the control device 22 using the learning model thereafter.

[0105] <Effects and Modified Examples>

[0106] In the present embodiment, the machine learning of the learning model is performed in the host vehicle 2a. If the machine learning of the learning model of all the vehicles 2 is performed by the server 3, the computational load in the server 3 becomes extremely high, but in the present embodiment, the machine learning of the learning model is performed by the host vehicle 2a, so the computational load in the server 3 can be reduced.

[0107] Further, in the above-described embodiment, the parameter value acquisition unit 145 of the host vehicle 2a acquires the value of the abnormality occurrence state parameter detected by the other vehicle 2b via communication between the server 3 capable of communicating with the other vehicle 2b and the host vehicle 2a. However, the parameter value acquisition unit 145 of the host vehicle 2a can acquire the value of the abnormality occurrence state parameter detected by the other vehicle 2b via inter-vehicle communication between the host vehicle 2a and the other vehicle 2b. In this case, the processor 14 of the other vehicle 2b directly transmits the acquired value of the abnormality occurrence state parameter to the host vehicle 2a via the inter-vehicle communication when the value of the abnormality occurrence state parameter is acquired from the memory 13 in the step S25.

[0108] The above describes preferred embodiments of the present application, but the present application is not limited to these embodiments, and various modifications and changes can be made within the scope of the claims.

Claims

1. A machine learning device that is a machine learning device of a learning model peculiar to a vehicle, comprising: a learning unit that performs machine learning of the learning model using a training data set including values of state parameters detected by a detector provided to the vehicle; and a parameter value acquisition unit that acquires, in a case where a value of a state parameter detected by the detector has become abnormal, a value of the abnormal state parameter detected by another vehicle under a condition that a detection condition at the time when the value of the state parameter included in the training data set is detected by the detector is satisfied, wherein the learning unit performs machine learning of the learning model using a training data set that includes, instead of the value of the abnormal state parameter detected by the detector, the value acquired by the parameter value acquisition unit from the other vehicle, in a case where the value of the state parameter detected by the detector has become abnormal, wherein the abnormal state parameter is a physical quantity whose difference between vehicles under the same detection condition is smaller than other state parameters used in the training data set, wherein the abnormal state parameter is a state parameter associated with an external environment of the vehicle, wherein the abnormal state parameter includes at least any one of a temperature, a humidity, and an atmospheric pressure of an atmosphere around the vehicle, wherein the detection condition includes a time and a place at which the value of the state parameter included in the training data set of the vehicle is detected by the detector, and wherein the other vehicle in which detection of the state parameter is performed at a time and a place that are most similar to a time and a place at which the value of the state parameter included in the training data set is detected in the vehicle is determined as the other vehicle from which the value of the abnormal state parameter is to be acquired.

2. The machine learning device according to claim 1, wherein the machine learning device is provided to the vehicle, and wherein the parameter value acquisition unit acquires the value of the abnormal state parameter detected by the other vehicle through inter-vehicle communication with the other vehicle or through communication with a server capable of communicating with the other vehicle.

3. The machine learning device according to claim 1, wherein the machine learning device is provided to a server capable of communicating with the vehicle and the other vehicle, and further comprises a model transmission unit that transmits information of the learning model machine-learned by the learning unit to the vehicle, wherein the parameter value acquisition unit acquires the value of the abnormal state parameter detected by the other vehicle from the other vehicle through communication between the other vehicle and the server, and wherein the learning unit performs machine learning of the learning model using a training data set that includes the value of the abnormal state parameter acquired from the other vehicle and a value of another state parameter associated with the vehicle other than the abnormal state parameter, which is transmitted from the vehicle to the server through communication between the vehicle and the server. ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​

Images