Learning device and model learning system

Abstract

The present invention provides a learning device and a model learning system, which suppress the reduction in the accuracy of the learning model when the device equipped with the learning model is used in an area different from the normal use area. The learning device (1) is configured to communicate with multiple devices (2) equipped with the learning model, wherein, among the multiple devices (2), a first device (2a) equipped with a learning model learned using a training data set obtained in a specified first area (110) for control is used in a specified second area (120), and the training data set obtained in the second area (120) is used to implement re-learning of the learning model equipped in the first device (2a).

Description

Technical Field

[0001] The invention relates to a learning device and a model learning system. Background Art

[0002] Patent Document 1 discloses a technique for estimating the temperature of an exhaust gas purification catalyst of an internal combustion engine mounted on a vehicle using a learning model such as a neural network model learned in a server or a vehicle.

[0003] Prior art literature

[0004] Patent Literature

[0005] Patent Document 1: Japanese Patent Application Publication No. 2019-183698 Summary of the invention

[0006] Problems to be solved by the invention

[0007] It is envisioned that various mobile or portable devices (for example, transportation devices such as vehicles or portable devices such as smartphones) are equipped with learning models to control the devices. Even for mobile or portable devices, the area where each device is used is usually limited to a certain extent. For example, in the case of a vehicle as a representative example of transportation equipment, if it is a private car, it is basically used within the living circle of the owner of each private car. If it is a commercial vehicle such as a taxi, bus, or multi-purpose vehicle for mobile services, it is basically used within the service provision area of ​​the operator who owns each commercial vehicle. In addition, if it is a smartphone as a representative example of a portable device, it is basically used within the living circle of the owner.

[0008] Therefore, when learning the learning model used in each device, by using an appropriate training data set corresponding to the characteristics of the usage area where each device is used on a daily basis, a high-precision learning model optimized according to the characteristics of the usage area of ​​each device can be generated.

[0009] However, in the case of mobile or portable devices, each device may be used in an area different from the usual use area (for example, a travel destination or a relocation destination, a temporary dispatch destination of a multi-purpose vehicle, etc.). In this case, the learning model of each device is a learning model optimized based on the characteristics of the usual use area, and since learning that reflects the characteristics of other areas is not performed, if each device is used in an area different from the usual use area, the accuracy of the learning model may decrease.

[0010] The present invention has been made with attention paid to such a problem, and an object of the present invention is to suppress a decrease in the accuracy of a learning model when a device equipped with a learning model is used in an area different from a normal use area.

[0011] Technical solutions to solve problems

[0012] In order to solve the above-mentioned problems, a learning device according to one embodiment of the present invention is configured to communicate with a plurality of devices equipped with learning models. Furthermore, the learning device is configured so that when a first device among the plurality of devices equipped with a learning model learned using a training data set obtained in a predetermined first area is used in a predetermined second area, the learning model equipped in the first device is re-learned using the training data set obtained in the second area.

[0013] In addition, a model learning system according to one embodiment of the present invention includes a server and a plurality of devices configured to communicate with the server. The server is configured such that, among the plurality of devices, a first device equipped with a learning model learned using a training data set obtained in a predetermined first area for control is used in a predetermined second area, and relearns the learning model using the training data set obtained in the second area.

[0014] Effects of the Invention

[0015] According to these aspects of the present invention, even when a device equipped with a learning model is used in an area different from a normal usage area, it is possible to suppress a decrease in the accuracy of the learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a schematic diagram of the structure of a model learning system according to one embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram showing a part of the hardware configuration of a vehicle according to one embodiment of the present invention.

[0018] Figure 3 This is a diagram showing an example of a learning model according to an embodiment of the present invention.

[0019] Figure 4 This is a flowchart showing an example of processing performed between a server and each vehicle in order to obtain an appropriate training data set corresponding to the characteristics of each smart city.

[0020] Figure 5 This is a flowchart showing an example of processing performed between a server and each multi-purpose vehicle in a first smart city and a second smart city in order to generate an appropriate learning model corresponding to the characteristics of each smart city.

[0021] Figure 6This is a flowchart showing an example of processing performed between a server and each multi-purpose vehicle in a first smart city when there is a shortage or potential shortage of multi-purpose vehicles in a second smart city. DETAILED DESCRIPTION

[0022] Hereinafter, one embodiment of the present invention will be described in detail with reference to the accompanying drawings. In the following description, the same reference numerals are given to the same components.

[0023] Figure 1 It is a schematic configuration diagram of a model learning system 100 according to one embodiment of the present invention.

[0024] The model learning system 100 of the present embodiment includes a server 1 and a plurality of vehicles 2 as an example of movable or portable devices.

[0025] The server 1 includes a server communication unit 11 , a server storage unit 12 , and a server processing unit 13 .

[0026] The server communication unit 11 includes a communication interface circuit for connecting the server 1 to the network 3 via a gateway or the like, and is configured to be able to communicate with each vehicle 2 .

[0027] The server storage unit 12 includes a storage medium such as a HDD (Hard Disk Drive), an optical recording medium, or a semiconductor memory, and stores various computer programs and data used in the processing of the server processing unit 13 .

[0028] The server processing unit 13 includes one or more processors and peripheral circuits thereof. The server processing unit 13 executes various computer programs stored in the server storage unit 12 and generally controls the overall operation of the server 1. For example, the server processing unit 13 is a CPU (Central Processing Unit).

[0029] The vehicle 2 in this embodiment is a multi-purpose vehicle for mobile services for various purposes such as transportation, logistics, and commodity sales, and is configured to automatically perform driving operations related to acceleration, steering, and braking. Figure 1 As shown, each vehicle 2 is associated with a specific smart city (in this embodiment, the first smart city 110 or the second smart city 120) and is basically used within the associated smart city.

[0030] In the following, for the sake of convenience, the vehicle 2 associated with the first smart city 110 and basically used in the first smart city 110 is referred to as the first multi-purpose vehicle 2a as needed, and the vehicle 2 associated with the second smart city 120 and basically used in the second smart city 120 is referred to as the second multi-purpose vehicle 2b as needed. In addition, the vehicle 2 may be a vehicle having only an internal combustion engine as a power source, or may be a hybrid vehicle, a plug-in hybrid vehicle, an electric vehicle (electric vehicle, fuel cell vehicle, etc.).

[0031] Figure 2 It is a schematic diagram showing a part of the hardware configuration of the vehicle 2 .

[0032] The vehicle 2 includes an electronic control unit 20, an off-vehicle communication device 24, various control components 25 mounted on the vehicle 2, such as an internal combustion engine, an electric motor, and an air conditioner, and various sensors 26 required for controlling the various control components 25 or detecting actual values ​​of input parameters and output parameters of a learning model described later. The electronic control unit 20, the off-vehicle communication device 24, and the various control components 25 and sensors 26 are connected to each other via an in-vehicle network 27 conforming to a standard such as CAN (Controller Area Network).

[0033] The electronic control unit 20 includes an in-vehicle communication interface 21, a vehicle storage unit 22, and a vehicle processing unit 23. The in-vehicle communication interface 21, the vehicle storage unit 22, and the vehicle processing unit 23 are connected to each other via a signal line.

[0034] The in-vehicle communication interface 21 is a communication interface circuit for connecting the electronic control unit 20 to an in-vehicle network 27 conforming to a standard such as CAN (Controller Area Network).

[0035] The vehicle storage unit 22 includes a storage medium such as a HDD (Hard Disk Drive), an optical recording medium, or a semiconductor memory, and stores various computer programs and data used in the processing of the vehicle processing unit 23 .

[0036] The vehicle processing unit 23 includes one or more processors and peripheral circuits thereof. The vehicle processing unit 23 executes various computer programs stored in the vehicle storage unit 22 and generally controls various control components mounted on the vehicle 2. For example, the vehicle processing unit 23 is a CPU.

[0037] The vehicle external communication device 24 is a vehicle-mounted terminal having a wireless communication function. The vehicle external communication device 24 communicates with the network 3 (see Figure 1 ) connected wireless base station 4 (refer to Figure 1), thereby connecting to the network 3 via the wireless base station 4. Thus, mutual communication with the server 1 is performed.

[0038] In each vehicle 2, when controlling various control components 25 mounted on each vehicle 2, a learning model (artificial intelligence model) obtained by learning, such as machine learning, is used as needed. In the present embodiment, as a learning model, a model obtained by deep learning of a neural network model (hereinafter referred to as "NN model") using a deep neural network (DNN: Deep Neural Network) or a convolutional neural network (CNN: Convolutional Neural Network) is used. Therefore, the learning model of the present embodiment can also be referred to as a learned NN model that has been implemented with deep learning. Deep learning is one of the machine learning methods that represents artificial intelligence (AI: Artificial Intelligence).

[0039] Figure 3 This is a diagram showing an example of a learning model (NN model) according to this embodiment.

[0040] Figure 3 The circle mark in represents an artificial neuron. An artificial neuron is usually called a node or a unit (in this specification, referred to as a "node"). Figure 3 In , L = 1 represents the input layer, L = 2 and L = 3 represent the hidden layer, and L = 4 represents the output layer. The hidden layer is also called the intermediate layer. Figure 3 Although a NN model with two hidden layers is illustrated in FIG, the number of hidden layers is not particularly limited, and the number of nodes in each of the input layer, the hidden layer, and the output layer is not particularly limited.

[0041] exist Figure 3 In , x1 and x2 represent the nodes of the input layer (L=1) and the output values ​​from the nodes, and y represents the nodes of the output layer (L=4) and their output values. (L＝2) 、z2 (L＝2) and z3 (L＝2) Represents each node of the hidden layer (L=2) and the output value from the node, z1 (L＝3) and z2 (L＝3) It represents each node of the hidden layer (L=3) and the output value from the node.

[0042] At each node of the input layer, the input is directly output. On the other hand, the output values ​​x1 and x2 of each node of the input layer are input to each node of the hidden layer (L=2), and the total input value u is calculated using the corresponding weights w and biases b at each node of the hidden layer (L=2). For example, Figure 3 In the hidden layer (L = 2), zk (L＝2) The total input value u calculated at each node represented by (k=1, 2, 3) k (L＝2) As shown below (M is the number of nodes in the input layer).

[0043]

Mathematical formula 1

[0044]

[0045] Next, the total input value u k (L＝2) Transformed by activation function f, from z in hidden layer (L=2) k (L＝2) The node shown as the output value z k (L＝2) (=f(u k (L＝2) On the other hand, the output value z1 of each node of the hidden layer (L=2) is input to each node of the hidden layer (L=3). (L＝2) 、z2 (L＝2) and z3 (L＝2) , at each node of the hidden layer (L = 3), the corresponding weight w and bias b are used to calculate the total input value u (= ∑z·w+b). The total input value u is also transformed by the activation function and output from each node of the hidden layer (L = 3) as the output value z1 (L＝3) 、z2 (L＝3) The activation function is, for example, a sigmoid function σ.

[0046] In addition, the output value z1 of each node of the hidden layer (L=3) is input to the node of the output layer (L=4). (L＝3) and z2 (L ＝3) , at the nodes of the output layer, the total input value u(∑z·w+b) is calculated using the weights w and biases b that correspond to each other, or the total input value u(∑z·w) is calculated using only the weights w that correspond to each other. For example, at the nodes of the output layer, the identity function is used as the activation function. In this case, the total input value u calculated at the nodes of the output layer is directly output from the nodes of the output layer as the output value y.

[0047] In this way, the learning model of this embodiment includes an input layer, a hidden layer, and an output layer. 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.

[0048] As examples of input parameters, for example, when the learning model is used to control the air conditioner mounted on the vehicle 2, various parameters that affect the temperature inside the vehicle, such as the outside air temperature, the location (latitude and longitude) of the vehicle 2, the date and time, and the parking time immediately before the vehicle travels, can be cited. Furthermore, as an example of an output parameter corresponding to such an input parameter, the set temperature of the air conditioner can be cited. Thus, by controlling the air conditioner to reach the set temperature obtained as the output parameter, the temperature inside the vehicle can be maintained at an appropriate temperature.

[0049] In addition, as an example of input parameters, for example, when the learning model is used to control the internal combustion engine mounted on the vehicle 2, the current values ​​of various parameters indicating the operating state of the internal combustion engine such as the internal combustion engine speed, the internal combustion engine cooling water temperature, the fuel injection amount, the fuel injection timing, the fuel pressure, the intake air amount, the intake air temperature, the EGR rate, and the supercharging pressure can be cited. In addition, as an example of output parameters corresponding to such input parameters, the estimated values ​​of various parameters indicating the performance of the internal combustion engine such as the CO2 concentration in the exhaust gas, the NOx concentration or the concentration of other substances, and the internal combustion engine output torque can be cited. Thus, by inputting the current values ​​of various parameters indicating the operating state of the internal combustion engine as input parameters into the NN model, the estimated values ​​of various parameters indicating the performance of the internal combustion engine (current estimated values ​​or future estimated values) can be obtained as output parameters, so that, for example, the internal combustion engine can be controlled based on the output parameters so that the performance of the internal combustion engine is close to the desired performance. In addition, in the case of having a sensor for measuring the output parameter, etc., it is also possible to judge the failure of the internal combustion engine or the sensor based on the difference between the measured value and the estimated value.

[0050] In order to improve the accuracy of the learning model, it is necessary to make the learning model learn. In the learning of the learning model, a large number of training data sets containing the measured values ​​of the input parameters and the measured values ​​of the output parameters corresponding to the measured values ​​of the input parameters (correct answer data) are used. By using a large number of training data sets, the values ​​of the weights w and the deviations b in the neural network are repeatedly updated using the known error back propagation method, thereby learning the values ​​of the weights w and the deviations b and improving the accuracy of the learning model.

[0051] Here, in the case of a vehicle 2 whose usage area is basically determined, such as the first multi-purpose vehicle 2a or the second multi-purpose vehicle 2b, by using an appropriate training data set corresponding to the characteristics of the usage area of ​​the vehicle 2 (for example, the climate or terrain in the area, environmental standards, age groups of residents, etc.), the learning model used in the vehicle 2 is learned, and a learning model optimized according to the characteristics of the usage area of ​​the vehicle 2 can be generated.

[0052] Thus, by inputting input parameters to the learning model, appropriate output parameters corresponding to the characteristics of the use area can be obtained. Therefore, for example, when controlling the air conditioner, the temperature inside the vehicle can be maintained at an appropriate temperature corresponding to the characteristics of the use area. In addition, when controlling the internal combustion engine, for example, the CO2 concentration, NOx concentration, etc. in the exhaust gas can be controlled to appropriate concentrations corresponding to the characteristics of the use area.

[0053] Therefore, in the present embodiment, each vehicle 2 is equipped with a learning model obtained by learning using an appropriate training data set corresponding to the use area of ​​each vehicle 2. Specifically, the first multi-purpose vehicle 2a is equipped with a learning model obtained by learning using a training data set obtained in the first smart city 110, and the second multi-purpose vehicle 2b is equipped with a learning model obtained by learning using a training data set obtained in the second smart city 120.

[0054] However, for some reason, such as an event scheduled to be held in the second smart city 120 and an envisioned situation where the demand for multi-purpose vehicles in the second smart city 120 temporarily increases, it is also possible to consider temporarily moving the first multi-purpose vehicle 2a to the second smart city 120 and using it in the second smart city 120.

[0055] However, the learning model of the first multi-purpose vehicle 2a is a learning model optimized according to the characteristics of the first smart city 110. Since no learning is performed to reflect the characteristics of the second smart city 120, if the first multi-purpose vehicle 2a is used in the second smart city 120, even if input parameters are input into the learning model, it may not be possible to obtain appropriate output parameter values ​​corresponding to the input parameters.

[0056] Therefore, in the present embodiment, when the first multi-purpose vehicle 2a is used in the second smart city 120 for some reason, the learning model mounted on the first multi-purpose vehicle 2a is re-learned using an appropriate training data set corresponding to the characteristics of the second smart city 120. Thus, the learning model of the first multi-purpose vehicle 2a can be updated to a learning model also corresponding to the second smart city 120, so when the first multi-purpose vehicle 2a is used in the second smart city 120, the value of the appropriate output parameter corresponding to the input parameter can be obtained through the learning model.

[0057] Figure 4 It is a flowchart showing an example of processing performed between a server 1 and a vehicle 2 (a multi-purpose vehicle 2a and a multi-purpose vehicle 2b) in order to obtain an appropriate training data set corresponding to the characteristics of each smart city 110, 120 in the model learning system 100 of the present embodiment.

[0058] In step S11, the electronic control unit 20 of the vehicle 2 determines whether the data volume of the training data set stored in the vehicle storage unit 22 is greater than the specified transmission volume. If the data volume of the training data set is greater than the transmission volume, the electronic control unit 20 proceeds to the process of step S12. On the other hand, if the data volume of the training data set is less than the transmission volume, the electronic control unit 20 ends the process of this time.

[0059] In addition, in the present embodiment, the electronic control unit 20 of the vehicle 2 obtains a training data set (actual measured values ​​of the input parameters and the output parameters of the learning model) at any time during the driving of the vehicle, and stores the obtained training data set in association with the acquisition location information (in the present embodiment, the first smart city 110 or the second smart city 120) in the vehicle storage unit 22.

[0060] In step S12 , the electronic control unit 20 of the vehicle 2 transmits the training data set together with the acquisition location information thereof to the server 1 , and after the transmission, deletes the data of the training data set stored in the vehicle storage unit 22 .

[0061] In step S13, when the server 1 receives the training data set from the vehicle 2, the server 1 stores the training data set in a specified database of the server storage unit 12 according to the location where the training data set is obtained. In this embodiment, the server 1 stores the training data set obtained in the first smart city 110 in the first database, and stores the training data set obtained in the second smart city 120 in the second database.

[0062] Figure 5 This is a flowchart showing an example of processing performed between the server 1 and the multi-purpose vehicles 2a and 2b in order to generate appropriate learning models corresponding to the characteristics of each smart city 110 and 120 in the model learning system 100 of the present embodiment.

[0063] In step S21, the server 1 determines whether the amount of data of the training data set stored in the first database has increased since the last time the training data set stored in the first database was used for learning, and whether it has increased by more than a predetermined learning start amount. If the amount of data of the training data set stored in the first database has increased by more than the learning start amount, the server 1 proceeds to the processing of step S22. On the other hand, if the amount of data of the training data set stored in the first database has increased by less than the learning start amount, the server 1 proceeds to the processing of step S26.

[0064] In step S22, the server 1 newly creates a learning model obtained by learning using the training data set of the most recent learning start amount stored in the first database.

[0065] In step S23 , the server 1 sends the newly created learning model to each multi-purpose vehicle 2 a used in the first smart city 110 .

[0066] In step S24, the electronic control unit 20 of the utility vehicle 2a determines whether a new learning model is received from the server 1. If a new learning model is received from the server 1, the electronic control unit 20 of the utility vehicle 2a proceeds to the process of step S25. On the other hand, if a new learning model is not received from the server 1, the electronic control unit 20 of the utility vehicle 2a ends the process.

[0067] In step S25, the electronic control unit 20 of the multi-purpose vehicle 2a updates the learning model used in the vehicle to a new learning model received from the server 1, i.e., learning is performed using a training data set of the most recent specified data volume (learning start volume) obtained within the first smart city 110, thereby optimizing the learning model according to the latest characteristics of the first smart city 110.

[0068] In step S26, the server 1 determines whether the amount of data of the training data set stored in the second database has increased since the last time the training data set stored in the second database was used for learning, and has increased to a learning start amount or more. If the amount of data of the training data set stored in the second database has increased to a learning start amount or more, the server 1 proceeds to step S27. On the other hand, if the amount of data of the training data set stored in the second database has increased to a learning start amount or more, the server 1 ends the current processing.

[0069] In step S27 , the server 1 newly creates a learning model obtained by learning using the training data set of the most recent learning start amount stored in the second database.

[0070] In step S28 , the server 1 sends the newly created learning model to each multi-purpose vehicle 2 b used in the second smart city 120 .

[0071] In step S29, the electronic control unit 20 of the utility vehicle 2b determines whether a new learning model is received from the server 1. If a new learning model is received from the server 1, the electronic control unit 20 of the utility vehicle 2b proceeds to the process of step S30. On the other hand, if a new learning model is not received from the server 1, the electronic control unit 20 of the utility vehicle 2b ends the process.

[0072] In step S30, the electronic control unit 20 of the multi-purpose vehicle 2b updates the learning model used in the vehicle to a new learning model received from the server 1, i.e., learning is performed using a training data set of the most recent specified data volume (learning start volume) obtained within the second smart city 120, thereby optimizing the learning model according to the latest characteristics of the second smart city 120.

[0073] Figure 6 This is a flowchart showing an example of processing performed between the server 1 and the multi-purpose vehicle 2a when the multi-purpose vehicles are insufficient or likely to be insufficient in the second smart city 120 in the model learning system 100 of the present embodiment.

[0074] In step S31, the server 1 determines whether there is a shortage or a possibility of shortage of multi-purpose vehicles in the second smart city 120. The method for determining whether there is a shortage or a possibility of shortage of multi-purpose vehicles in the second smart city 120 is not particularly limited. For example, the determination can be made by understanding the congestion situation in the second smart city 120 based on crowd flow data for understanding human activities. If there is a shortage or a possibility of shortage of multi-purpose vehicles in the second smart city 120, the server 1 proceeds to the process of step S32. On the other hand, if there is no shortage and no possibility of shortage of multi-purpose vehicles in the second smart city 120, the server 1 ends the current process.

[0075] In step S32, the server 1 selects the required number of vehicles (hereinafter referred to as "support vehicles") that can be temporarily moved to and operated in the second smart city 120 from the multi-purpose vehicles 2a in the first smart city 110, and sends a movement request to the selected support vehicles (multi-purpose vehicles 2a).

[0076] In step S33, the electronic control unit 20 of the utility vehicle 2a determines whether a movement request is received. If a movement request is received, the electronic control unit 20 of the utility vehicle 2a proceeds to step S34, and if no movement request is received, the current process ends.

[0077] In step S34 , the electronic control unit 20 of the multi-purpose vehicle 2 a moves the vehicle toward the second smart city 120 with the movement request based on the movement request, and sends the learning model of the vehicle to the server 1 .

[0078] In step S35, the server 1 determines whether the learning model is received from the support vehicle (i.e., the multi-purpose vehicle 2a that sent the movement request). If the learning model is received from the support vehicle, the server 1 proceeds to step S36. On the other hand, if the learning model is not received from the support vehicle, the server 1 ends the current processing.

[0079] In step S36, server 1 uses a training data set of the most recent specified data amount (e.g., the learning start amount) obtained within the second smart city 120 and stored in the second database (i.e., an appropriate training data set corresponding to the characteristics of the second smart city 120) to implement re-learning of the learning model of the supporting vehicle.

[0080] Regarding the re-learning at this time, the entire learning model of the supporting vehicle can be taken as the object of re-learning to re-learn the weight w and the bias b value of each node in each hidden layer of the learning model. Alternatively, transfer learning can be implemented by directly using a part of the learning model of the supporting vehicle and taking the remaining part as the object of re-learning, thereby directly using the weight w and the bias b value of each node in a part of the hidden layers and only re-learning the weight w and the bias b value of each node in the remaining part of the hidden layers.

[0081] If transfer learning is performed during relearning, the amount of computation required for relearning can be reduced and the time required for relearning can be shortened. Therefore, it is also possible to determine whether to perform normal learning in which the entire learning model of the supporting vehicle is the object of relearning or to perform transfer learning, based on the number of supporting vehicles, such as performing transfer learning when there are more supporting vehicles than the specified number.

[0082] In step S37 , the server 1 transmits the relearned learning model to the assisting vehicle that transmitted the learning model.

[0083] In step S38, the electronic control unit 20 of the utility vehicle 2a determines whether the learning model relearned in the server 1 is received. If the learning model relearned in the server 1 is received, the electronic control unit 20 of the utility vehicle 2a proceeds to the process of step S39. On the other hand, if the learning model relearned in the server 1 is not received, the electronic control unit 20 of the utility vehicle 2a ends the current process.

[0084] In step S39, the electronic control unit 20 of the multi-purpose vehicle 2a updates the learning model to the learning model re-learned in the server 1. As a result, the learning model of the multi-purpose vehicle 2a dispatched as a support vehicle to the second smart city 120 can be updated to a learning model also corresponding to the second smart city 120, so when the first multi-purpose vehicle 2a is used in the second smart city 120, the value of the appropriate output parameter corresponding to the input parameter can be obtained through the learning model.

[0085] In addition, when there is a shortage or possible shortage of multi-purpose vehicles in the first smart city 110, as long as the server 1 and the multi-purpose vehicle 2b are connected with each other, Figure 6 The same processing as shown may be performed.

[0086] The model learning system 100 of the present embodiment described above includes a server 1 as a learning device and a plurality of vehicles 2 (devices) configured to communicate with the server 1. Furthermore, the server 1 is configured such that, among the plurality of vehicles 2, a first multi-purpose vehicle 2a (first device) equipped with a learning model learned using a training data set obtained in a first smart city 110 (prescribed first area) for control is re-learned using the training data set obtained in the second smart city 120 when the first multi-purpose vehicle 2a is used in a second smart city 120 (prescribed second area).

[0087] Thus, the learning model of the first multi-purpose vehicle 2a can be updated to a learning model that also corresponds to the second smart city 120, so when the first multi-purpose vehicle 2a is used in the second smart city 120, the value of the appropriate output parameter corresponding to the input parameter can be obtained through the learning model. Therefore, even when the first multi-purpose vehicle 2a is used in the second smart city 120 that is different from the first smart city 110 that is the normal use area, the accuracy of the learning model can be suppressed from decreasing.

[0088] In addition, in the present embodiment, the first multi-purpose vehicle 2a is an automatic driving vehicle that automatically performs driving operations related to acceleration, steering, and braking, and is configured to, upon receiving a movement request from the first smart city 110 to the second smart city 120 from the server 1, move the first multi-purpose vehicle 2a from the first smart city 110 to the second smart city 120 based on the movement request, and send the learning model mounted on the first multi-purpose vehicle 2a to the server 1. In addition, when receiving the learning model from the first multi-purpose vehicle 2a that sent the movement request, the server 1 is configured to re-learn the learning model using the training data set obtained in the second smart city 120, and send the re-learned learning model again to the first multi-purpose vehicle 2a that sent the movement request.

[0089] In addition, the server of the present embodiment is also configured to implement transfer learning in which, when the number of first multi-purpose vehicles 2a making a movement request is greater than a prescribed number, a portion of the learning model is used and the remaining portion is learned using the training data set obtained in the second smart city 120 while re-learning the learning model received from the first multi-purpose vehicle 2a using the training data set obtained in the second smart city 120.

[0090] By performing transfer learning, as described above, compared with the case of performing normal learning in which the entire learning model is relearned, the amount of calculation required for relearning can be reduced and the time required for relearning can be shortened. Therefore, when the number of first multi-purpose vehicles 2a that have made movement requests is large, by performing transfer learning during relearning, the calculation load of the server 1 can be reduced and the time required for relearning can be shortened.

[0091] Although the embodiments of the present invention have been described above, the above embodiments merely illustrate a part of application examples of the present invention and do not limit the technical scope of the present invention to the specific structures of the above embodiments.

[0092] For example, in the above embodiment, a server 1 implements the reference Figures 4 to 6 The various processing described above, but each server and vehicle 2 can also be configured to use multiple servers (for example, a server for managing multi-purpose vehicles 2a in the first smart city 110 and a server for managing multi-purpose vehicles 2b in the second smart city 120) to perform processing equivalent to the various processing described above.

[0093] In addition, in the above embodiment, the case where the vehicle 2 moves between the first smart city 110 and the second smart city 120 is used as an example for explanation, but the present invention is not limited thereto, and the number of smart cities may be three or more. In addition, in the case where the first smart city 110 is divided into more detailed areas for the operation of multi-purpose vehicles, that is, when the multi-purpose vehicles used in each area are fixed, when the multi-purpose vehicles are moved from one area to another area in the first smart city 110, the learning model optimized according to the characteristics of one area may be re-learned using an appropriate training data set corresponding to the characteristics of another area.

[0094] In addition, in the above-mentioned embodiment, the case of using the vehicle 2 as an example of a movable or portable device is described, but the server 1 and each device can also be configured so that when other transportation equipment or portable devices are used in other areas (second areas) different from the normal usage area (first area), the learning model installed in each device (the learning model optimized according to the characteristics of the normal usage area) is re-learned using an appropriate training data set corresponding to the characteristics of the other area.

[0095] Description of symbols

[0096] 1 Server (Learning Device)

[0097] 2 Vehicles (Equipment)

[0098] 100 Model Learning System

[0099] 110 First Smart City (First Region)

[0100] 120 Second Smart City (Second Area)

Claims

1. A model learning system comprising a server and a plurality of vehicles configured to communicate with the server, characterized in that: The plurality of vehicles are equipped with a learning model obtained by learning using a training data set obtained in a specified first area, and when an autonomous driving vehicle controlled by the learning model is used in a specified second area, the server uses the training data set obtained in the second area to re-learn the learning model installed in the autonomous driving vehicle, When the autonomous driving vehicle receives a request to move from the first area to the second area from the server, the autonomous driving vehicle moves from the first area to the second area based on the request, and sends the learning model mounted on the autonomous driving vehicle to the server. When the server receives the learning model from the autonomous driving vehicle that has sent the movement request, the server relearns the received learning model using the training data set obtained in the second area, and sends the relearned learning model to the autonomous driving vehicle that has sent the movement request. When the number of the autonomous driving vehicles making the movement request is greater than a specified number, the server performs transfer learning by using a portion of the learning model and learning the remaining portion using the training data set obtained in the second area while relearning the learning model received from the autonomous driving vehicles using the training data set obtained in the second area.

Images