In-vehicle processing device for learning data

Abstract

The present application relates to a learning data on-vehicle processing device in which learning of weights of a neural network is performed on a vehicle (1) or at a server (2) outside the vehicle (1), the learning data on-vehicle processing device being provided with a data acquisition unit (40) that acquires data related to operation of the vehicle (1), a neural network storage unit (41) that stores a neural network that outputs an output value related to operation control of the vehicle (1) if data acquired at the data acquisition unit (40) is input, and a learning data storage unit (42) that stores learning data of the weights of the neural network. If the frequency of learning of the weights of the neural network on the vehicle (1) or the frequency of transmission of the learning data to the server (2) becomes lower, the amount of storage per unit time of the learning data successively stored in the learning data storage unit (42) or the amount of the learning data completed storage in the learning data storage unit (42) is reduced.

Description

Technical Field

[0001] This invention relates to an in-vehicle data processing device for learning purposes. Background Technology

[0002] A data collection system is known in the art that can collect data related to the vehicle's surrounding environment or the vehicle's state on a server at an appropriate acquisition frequency (see, for example, Japanese Patent Application Publication No. 2018-55191). Summary of the Invention

[0003] However, in this scenario, if the data acquisition frequency is low, the data needs to be stored in the vehicle's data storage unit until it is collected. Therefore, if the storage capacity of the data storage unit is insufficient, the required data may not be stored.

[0004] According to the present invention, an in-vehicle processing apparatus for learning data is provided, wherein the learning of weights of a neural network is performed at a server located on or outside the vehicle, the in-vehicle processing apparatus for learning data includes:

[0005] The data acquisition unit acquires data related to the vehicle's operation.

[0006] The neural network storage unit stores the neural network. If the input is data acquired at the data acquisition unit, the neural network outputs values ​​related to the vehicle's operation control.

[0007] The learning data storage unit stores the learning data for the weights of the neural network.

[0008] A frequency acquisition unit acquires the learning frequency of the weights of the neural network on the vehicle or the frequency at which the learning data is transmitted to the server.

[0009] If the learning frequency of the neural network weights on the vehicle or the frequency of transmitting learning data to the server becomes lower, the learning data modification unit reduces the amount of learning data stored per unit time in the learning data storage unit or the amount of learning data that has been stored in the learning data storage unit.

[0010] When the learning frequency of the neural network weights on the vehicle or the frequency of transmitting learning data to the server becomes lower, the necessary learning data can be stored in the learning data storage unit by reducing the amount of learning data stored per unit time in the learning data storage unit or reducing the amount of learning data stored in the learning data storage unit. Attached Figure Description

[0011] Figure 1 This is an overall view of the onboard processing unit for learning data.

[0012] Figure 2 This is a view of the feature configuration of the first example of machine learning.

[0013] Figure 3 This is a view of the feature configuration for the second example of machine learning.

[0014] Figure 4 This is a view that shows an example of a neural network.

[0015] Figure 5 This shows a view of the dataset used for learning.

[0016] Figure 6A , Figure 6B and Figure 6C These are views showing the amount of data stored for learning purposes.

[0017] Figure 7A , Figure 7B , Figure 7C and Figure 7D These are views showing the relationship between the amount of learning data stored per unit time and the learning frequency or transmission frequency, the relationship between the cycle of acquiring learning data and the learning frequency or transmission frequency, the relationship between the type of learning data and the learning frequency or transmission frequency, and the relationship between the amount of learning data stored and the learning frequency or transmission frequency.

[0018] Figure 8 This is a view that shows an example of a dataset used for learning.

[0019] Figure 9 This is a view of the functional configuration of an example of an in-vehicle processing device according to the present invention.

[0020] Figure 10 yes Figure 9 The diagram shows a functional configuration of an example of an onboard processing unit.

[0021] Figure 11 It is a flowchart used to store learning data.

[0022] Figure 12 It is a flowchart used for learning and processing.

[0023] Figure 13 It is a flowchart used to store learning data.

[0024] Figure 14 It is a flowchart used to store learning data.

[0025] Figure 15 It shows Figure 9 A view of the functional configuration diagram of another example of an on-board processing device.

[0026] Figure 16 It is a flowchart used to process learning data.

[0027] Figure 17A and Figure 17B These are, respectively, a functional configuration diagram of another embodiment of the vehicle-mounted processing device according to the present invention and a view showing a functional configuration diagram of the server.

[0028] Figure 18 This is a flowchart used to handle communication between vehicles and servers.

[0029] Figure 19 It shows Figure 17A A view of the functional configuration diagram of an example of an onboard processing device.

[0030] Figure 20 It is a flowchart used to store learning data.

[0031] Figure 21 It is a flowchart used to store learning data.

[0032] Figure 22 It is a flowchart used to store learning data.

[0033] Figure 23 It shows Figure 17A A view of the functional configuration diagram of another example of an on-board processing device.

[0034] Figure 24 It is a flowchart used to process learning data. Detailed Implementation

[0035] If reference Figure 1 1 represents a vehicle and 2 represents a server. For example... Figure 1 As shown, the electronic control unit 3 is installed inside the vehicle 1. The electronic control unit 3 is composed of a digital computer and is equipped with a CPU (microprocessor) 5 and a memory 6 composed of ROM and RAM, which are interconnected via a bidirectional bus 4.

[0036] Various types of sensors 7 are connected to the electronic control unit 4. Furthermore, a communication unit 8 for communicating with the server 2 is connected to the electronic control unit 4. On the other hand, the electronic control unit 10 is arranged within the server 2. This electronic control unit 10 is composed of a digital computer and is equipped with a CPU (microprocessor) 12 interconnected via a bidirectional bus 11 and a memory 13 composed of ROM and RAM. A communication unit 14 for communicating with the vehicle 1 is connected to the electronic control unit 10.

[0037] In an embodiment of the invention, learning data is collected at vehicle 1, and learning is performed based on the collected learning data to prepare a learning model. In this case, in an embodiment of the invention, sometimes the learning model is prepared at vehicle 1, i.e., in-vehicle learning is performed, and sometimes the learning model is prepared outside vehicle 1, i.e., at server 2. Therefore, firstly, two examples will be briefly explained in which the machine learning used to prepare the learning model can be performed not only on vehicle 1 but also outside vehicle 1.

[0038] exist Figure 2 The image shows a view of the feature configuration for the first example. (Refer to...) Figure 2 In this first example, the system consists of a target torque calculation unit 20, a control parameter calculation unit 21, a switching unit 22, an engine control unit 23 for controlling the engine 24 of vehicle 1, a feedback correction unit 25, a torque deviation calculation unit 26, and a switching control unit 27. Note that a torque sensor 24a is attached to the engine 24 to detect the actual output torque Tr of the engine. Figure 2 As shown, for example, the target torque calculation unit 20 consists of, as... Figure 4 The neural network NN configuration is shown. If the input values ​​x1 (accelerator opening), x2 (engine speed), x3 (air temperature), and x4 (altitude) are input to the target torque calculation unit 20, then the target torque calculation unit 20 is configured to output the target torque Tt of the engine 24. Note that in Figure 4 In the diagram, L=1 indicates the input layer, L=2, L=3, and L=4 indicate the hidden layers, and L=5 indicates the output layer. x1 to xn represent the input values ​​of the nodes in the input layer (L=1), while "y" represents the output values ​​of the nodes in the output layer (L=5).

[0039] On the other hand, the relationship between fuel injection quantity, air-fuel ratio, ignition timing, intake valve timing, exhaust valve timing, and the target torque control value "y" of engine 24d is determined experimentally in advance, such that when the target torque control value "y" of engine 2 is input to engine control unit 23, the actual output torque Tr of engine 24 becomes the target torque Tt, and this relationship is stored in advance in engine control unit 23. Therefore, generally, if the target torque control value "y" of engine 24 is input to engine control unit 23, the actual output torque Tr of engine 24 becomes the target torque Tt. On the other hand, for example, the control parameter calculation unit 21 is composed of... Figure 4 The neural network NN shown is used. The control parameter calculation unit 21 is configured such that if input values ​​x1 (accelerator opening), x2 (engine speed), x3 (air temperature), and x4 (altitude) are input to the control parameter calculation unit 21, then the control parameter calculation unit 21 outputs the target torque control value "y" for the engine 24. Normally, this target torque control value "y" is directly sent to the engine control unit 23 by the switching unit 22. At this time, the actual output torque Tr of the engine becomes the target torque Tt.

[0040] Now, if vehicle 1 is used for an extended period, a torque deviation will occur between the actual output torque Tr of engine 24 and the target torque Tt due to engine 24 aging. This torque deviation ΔTt (=Tt-Tr) is calculated at torque deviation calculation unit 26 based on the output of target torque calculation unit 20 and the detection value of torque sensor 24a. If the torque deviation ΔTt becomes larger, switching unit 22 is switched by switching control unit 27, causing the output value of control parameter calculation unit 21 to be input to feedback correction unit 25. At this time, at feedback correction unit 25, C·ΔTt (C is a small constant) is added to the target torque control value “y” output from control parameter calculation unit 21, making the torque deviation ΔTt smaller, and the addition result “y” (=y+C·ΔTt) is input to engine control unit 23. Next, if the torque deviation ΔTt becomes an acceptable value or smaller, switching unit 22 is switched so that the output value of control parameter calculation unit 21 is directly input to engine control unit 23.

[0041] In the first example, typically, whenever the torque deviation ΔTt becomes an allowable value or smaller, the input values ​​x1, x2, x3, x4 when the torque deviation ΔTt becomes an allowable value or smaller, and the target torque control value “y” (=y+C·ΔTt) output from the feedback correction unit 25 when the torque deviation ΔTt becomes an allowable value or smaller are stored sequentially. Therefore, as prepared... Figure 5The training dataset is shown. Note that in this case, the target torque control value “y” (=y+C·ΔTt) output from the feedback correction section 25 is stored as training data yt in the training dataset. If you prepare as follows... Figure 5 The learning dataset shown is used in an embodiment of the invention where the learning of the weights of the neural network NN forming the control parameter calculation unit 21 is performed on the vehicle 1 or the server 2, and the learning model for outputting the target torque control value "y" is generated by the learning neural network NN.

[0042] When learning the weights of a neural network (NN), firstly, the weights are... Figure 5 The learning example shown uses the input values ​​x1, x2, x3, and x4 from the first set of the dataset. Figure 4 The neural network NN is shown, and the weights of the neural network NN are learned through the backpropagation method, so that the squared error E(=1 / 2(y-yt)) between the output value "y" of the neural network NN and the corresponding training data yt is equal to that between the output value "y" and the training data yt. 2 The error becomes smaller. If the learning of the weights of the neural network NN based on the first dataset is complete, then the input values ​​x1, x2, x3, and x4 from the second set are input into the neural network NN, and the weights of the neural network NN are learned through backpropagation, so that the squared error E(=1 / 2(y-yt)) between the output value “y” from the neural network NN and the corresponding training data yt becomes smaller. 2 The weights of the neural network NN are then learned sequentially using the same technique, based on the corresponding datasets from the third set to the "m"th set. If the learning of the weights of the neural network NM based on all datasets from the first set to the "m"th set is completed, the learned weights are used to update the weights of the neural network NN that form the control parameter calculation unit 21.

[0043] Figure 3 This shows a view of the feature configuration for the second example. (Refer to...) Figure 3 The second example comprises a catalyst temperature estimation unit 30 for estimating the temperature of the catalyst arranged in the engine exhaust passage, a switching unit 31, an engine control unit 23 for controlling the engine 24 of the vehicle 1, and a switching control unit 32. Note that a temperature sensor 24b for detecting the actual catalyst temperature Td is attached to the engine 24. The detection signal from this temperature sensor 24b is input to the engine control unit 23 via the switching unit 31. Based on the actual catalyst temperature Td detected by the temperature sensor 24b, for example, preheating operation control and other controls of the engine 24 are performed.

[0044] On the other hand, in the second example, when the temperature sensor 24b malfunctions, a catalyst temperature estimation unit 30 is provided. For example, this catalyst temperature estimation unit 30 is composed of... Figure 4 The neural network NN shown is configured as follows. The catalyst temperature estimation unit 30 is configured such that if input values ​​x1 (engine load rate), x2 (engine speed), x3 (air-fuel ratio), x4 (ignition timing), and x5 (HC or CO concentration in exhaust gas) are input to the catalyst temperature estimation unit 30, the catalyst temperature estimation unit 30 outputs an estimated value Te of the catalyst temperature. In the switching control unit 32, based on the detection value of the temperature sensor 24b, it is determined whether the temperature sensor 24b is functioning correctly. When it is determined that the temperature sensor 24b is faulty, the switching unit 31 is switched so that the output value of the catalyst temperature estimation unit 30 is input to the engine control unit 23. At this time, the estimated value Te of the catalyst temperature calculated by the catalyst temperature estimation unit 30 is input to the engine control unit 23, and engine 24 control is performed based on this estimated value Te of the catalyst temperature.

[0045] In the second example, when the switching control unit 32 determines that the temperature sensor 24b is functioning correctly based on its detection value, for example, it periodically stores the input values ​​x1, x2, x3, x4, x5 and the actual catalyst temperature Td detected by the temperature sensor 24b at that time. Therefore, preparation is as follows... Figure 5 The training dataset is shown. Note that in this case, the actual catalyst temperature Td detected by temperature sensor 24b is stored as training data yt in the training dataset. If prepared as follows... Figure 5 The training dataset shown indicates that, in an embodiment according to the invention, the learning of the weights of the neural network NN forming the catalyst temperature estimation unit 30 is performed on vehicle 1 or server 2. A learning model is generated by learning the neural network NN to produce an estimated value Te for the output catalyst temperature.

[0046] When learning the weights of a neural network (NN), in the same situation, firstly, as follows... Figure 5 The learning example shown uses the input values ​​x1, x2, x3, x4, and x5 from the first set of the dataset. Figure 4 The neural network NN is shown. At this point, the weights of the neural network NN are learned using the backpropagation method, such that the squared error E(=1 / 2(y-yt)) between the output value "y" from the neural network NN and the corresponding training data yt is equal to the error between the output value "y" and the training data yt. 2The value becomes smaller. If the learning of the weights of the neural network NN based on the first dataset is complete, then the values ​​x1, x2, x3, x4, and x5 from the second set are input into the neural network NN. At this point, the weights of the neural network NN are learned through backpropagation, such that the squared error E(=1 / 2(y-yt)) between the output value "y" from the neural network NN and the corresponding training data yt becomes smaller. 2 The weights of the neural network NN are then learned sequentially based on the corresponding datasets from the third set to the "m"th set, using a similar method. If the learning of the weights of the neural network NN based on all datasets from the third set to the "m"th set is completed, the learned weights are used to update the weights of the neural network NN configuring the catalyst temperature estimation unit 30.

[0047] Now, in an embodiment of the invention, the learning process for preparing the learning model is repeatedly performed on vehicle 1 or at server 2 during the operation of vehicle 1. In this case, in order to accurately learn the weights of the neural network NN, a sufficient amount of learning data is needed to prepare the model. Figure 5 The training dataset is shown, and the training data must be continuously acquired between the execution of the previous learning process and the execution of the current learning process. Therefore, in an embodiment according to the invention, training data is continuously acquired between repeatedly executed learning processes until a sufficient amount of training data is acquired to prepare the training dataset. Figure 5 The training dataset is shown. In this case, in an embodiment according to the invention, the training data required for learning the weights of the neural network NN is stored in the memory 6 of the electronic control unit 3 of the vehicle 1. Therefore, in an embodiment according to the invention, during the operation of the vehicle 1, the amount of training data stored in the memory 6 increases slowly.

[0048] In this regard, electricity is required to learn the weights of the neural network NN. Therefore, if the learning of the weights of the neural network NN is performed on vehicle 1, the learning of the weights of the neural network NN is performed, for example, when predetermined learning conditions such as an operating state with low power consumption are met. Therefore, the learning frequency when the learning of the weights of the neural network NN is performed on vehicle 1—that is, the on-board learning frequency—is not constant, but fluctuates according to the operating state of vehicle 1, etc. Figure 6A The diagram shows the variation in the amount of learning data M stored in memory 6 when on-board learning is performed at a relatively high frequency. Note that in... Figure 6AIn this context, time "t" indicates when the on-board learning process is executed. If on-board learning processing is executed, the learning data stored in memory 6 is erased. Therefore, if on-board learning processing is executed, then after that, as... Figure 6A As shown, the amount of data stored for learning, M, increases slowly over time. Then, if on-board learning is performed, the amount of data stored for learning, M, becomes zero.

[0049] On the other hand, Figure 6A In this context, MM represents the storage capacity of memory 6 that can be used to store data for learning purposes. If onboard learning is performed, this storage capacity MM becomes the capacity to store data for preparation, such as... Figure 5 The dataset shown represents the required amount of learning data. Now, when performing onboard learning at a relatively high frequency, such as... Figure 6A As shown, the time interval for performing onboard learning processing is relatively short; therefore, in this case, such as Figure 6A As shown, the storage amount M of the learning data stored in memory 6 becomes storage capacity MM or less, and there is no problem with insufficient storage capacity in memory 6. Conversely, if the on-board learning frequency decreases, the time interval for performing on-board learning processing becomes longer. As a result, if on-board learning processing is performed under these circumstances, such as Figure 6B As shown, the amount of learning data M stored in memory 6 before performing onboard learning processing will eventually reach the storage capacity MM.

[0050] If, in this manner, the amount of learning data M reaches the storage capacity MM, then subsequently acquired learning data will be discarded instead of being stored in memory 6. That is, in this case, the storage capacity of memory 6 will become insufficient relative to the learning data. If, in this manner, the storage capacity of memory 6 becomes insufficient relative to the learning data, as if from... Figure 6B As understood in the text, learning data acquired after the storage volume M of the learning data reaches the storage capacity M and before the execution of in-vehicle learning processing cannot be used for in-vehicle learning. In this regard, there are many situations in learning processing where data acquired close to the time of execution of the learning process has a particularly significant impact on the learning results. Therefore, it is essential to avoid the situation where learning data acquired after the storage volume M of the learning data reaches the storage capacity M and before the execution of in-vehicle learning processing cannot be used for in-vehicle learning.

[0051] Therefore, in an embodiment according to the present invention, when learning the weights of a neural network NN is performed on vehicle 1, i.e., when performing in-vehicle learning, if the learning frequency of the weights of the neural network NN in the vehicle becomes lower, the amount of learning data stored per unit time in the memory 6—i.e., the learning data storage unit—or the amount of learning data stored in the learning data storage unit is reduced. If the amount of learning data stored per unit time in the learning data storage portion or the amount of learning data stored in the learning data storage unit is reduced in this way, such as... Figure 6C As shown, the amount of learning data stored, M, increases slowly. Therefore, the amount of learning data stored, M, no longer exceeds the storage capacity, MM, and all the learning data acquired before performing the learning process becomes usable for learning.

[0052] On the other hand, when the learning model is prepared at server 2, the learning data stored in memory 6 is transferred to server 2. In this case, the prepared model can be... Figure 5 The learning dataset shown can be transmitted to server 2 in one go, or it can be divided into smaller parts and transmitted to server 2 little by little. In this case, according to an embodiment of the invention, the preparation is as follows: Figure 5 The amount of learning data required for the training dataset shown is divided into small portions and transmitted to server 2 little by little. In this case, the storage capacity MM relative to the amount of learning data is made smaller compared to the storage capacity MM used in the case of performing on-board learning. Even though the storage capacity MM relative to the amount of learning data is made smaller in this way, the amount of learning data M will change in the same way as in the case of performing on-board learning when the learning model is prepared at server 2, such as... Figure 6A and Figure 6B As shown. However, in this case, Figure 6A and Figure 6B The time "t" in the figure represents the time when the learning data will be transmitted to server 2.

[0053] That is, the amount and speed of wireless communication to server 2 are limited. Learning data is transmitted to server 2 when predetermined transmission conditions are met. If learning data stored in memory 6 is transmitted to server 2, the learning data stored in memory 6 is erased. Therefore, in the same situation, if the frequency of transmitting learning data to server 2 is very high, such as... Figure 6AAs shown, the intervals between repeated transmission processes are very short. Therefore, even if learning data is continuously acquired during this period, the amount of learning data stored in memory 6 will not increase, and the storage capacity MM of memory 6 will not be insufficient. Conversely, if the frequency of transmitting learning data to server 2 becomes lower, such as... Figure 6B As shown, the intervals between repeated transmission processes become longer. Therefore, if learning data is continuously acquired during this period, the amount of learning data to be stored in memory 6 becomes larger, and the storage capacity MM of memory 6 may become insufficient.

[0054] Therefore, in an embodiment according to the present invention, when transferring learning data stored in memory 6 for preparing the learning model at server 2 to server 2, if the frequency of transferring learning data to server 2 becomes lower, the amount of learning data stored per unit time in memory 6—that is, the learning data storage unit—or the amount of learning data stored in the learning data storage unit is reduced. In this way, if the amount of learning data stored per unit time in the learning data storage unit or the amount of learning data stored in the learning data storage unit is reduced, such as... Figure 6C As shown, the amount of learning data stored, M, increases slowly, and eventually no longer exceeds the storage capacity. Therefore, in the same scenario, all learning data acquired before transmission can be used for learning.

[0055] Next, in reference Figures 7A to 7D At the same time, a more specific explanation will be given. It is important to note that... Figures 7A to 7D The horizontal axis represents the frequency of learning the weights of the neural network NN performed on vehicle 1—that is, the on-board learning frequency, or the frequency at which the learning data stored in memory 6 is transmitted to server 2 to perform the learning of the weights of the neural network NN at server 2. Note that in the following text, these on-board learning frequencies or transmission frequencies to server 2 will be simply referred to as learning frequencies or transmission frequencies.

[0056] As described above, in an embodiment of the present invention, when learning the weights of a neural network (NN) is performed on vehicle 1, if the frequency of learning the weights of the NN becomes lower, the amount of learning data stored per unit time in the learning data storage section decreases. Furthermore, when learning data stored in memory 6 is sent to server 2 for learning the weights of the NN at server 2, if the frequency of transmitting learning data to server 2 becomes lower, the amount of learning data stored per unit time in the learning data storage section decreases. In this case, in an embodiment of the present invention, as... Figure 7A As shown, the lower the learning frequency or transmission frequency, the less learning data is stored per unit time in the learning data storage section.

[0057] In this case, according to one embodiment of the invention, by changing the acquired learning data, the amount of learning data stored per unit time in the learning data storage section is reduced. In this embodiment, sometimes the cycle period for acquiring learning data is changed to change the acquired learning data, and sometimes the type of acquired learning data is changed to change the acquired learning data. When the cycle period for acquiring learning data is changed to change the acquired learning data, by increasing the cycle period for acquiring learning data, the amount of learning data stored per unit time in the learning data storage section is reduced. In this case, for example, as... Figure 7B As shown, the lower the learning frequency or transmission frequency, the longer the cycle of acquiring learning data becomes. If the cycle of acquiring learning data is increased, for example, in... Figure 5 In the training dataset shown, the time interval between obtaining the k-th dataset and obtaining the (k+1)-th dataset becomes longer, thus the storage size M of the training data increases slowly.

[0058] On the other hand, when the type of acquired learning data is changed, the amount of learning data stored per unit time in the learning data storage unit is reduced by decreasing the type of acquired learning data. In this case, for example, as... Figure 7C As shown, the lower the learning frequency or transmission frequency, the more the types of data used for learning are reduced. If the types of data used for learning are reduced, for example, by... Figure 5 The learning example shown uses input values ​​x1, x2, ..., x from the dataset. n-1 x n become Figure 8 The learning example shown uses input values ​​x1, x2, ..., x from the dataset. s-1x s That is, the number of input values ​​is reduced from "n" to "s". Therefore, the amount of learning data acquired each time becomes smaller, and thus the storage capacity M of the learning data increases slowly. However, in this case, as... Figure 8 As shown, the number of datasets prepared before performing the learning process increases from "m" to "r". If a specific example is given, then... Figure 2 In the first example shown, for instance, the type of learning data obtained is reduced by removing altitude from the input values. Figure 3 In the second example shown, for example, the type of learning data obtained is reduced by removing the HC or CO concentration in the exhaust gas from the input values.

[0059] On the other hand, as described above, in the embodiment according to the present invention, when learning the weights of a neural network NN is performed on vehicle 1, if the frequency of learning the weights of the neural network NN becomes lower, the amount of learning data stored in the learning data storage unit decreases. Furthermore, when the learning data stored in memory 6 is transferred to server 2 to perform learning the weights of the neural network NN at server 2, if the frequency of transferring the learning data to server 2 becomes lower, the amount of learning data stored in the learning data storage unit decreases. In this case, in the embodiment according to the present invention, as... Figure 7D As shown, the lower the learning frequency or transmission frequency, the less learning data is stored in the learning data storage unit.

[0060] In this case, according to an embodiment of the invention, the amount of learning data stored in the learning data storage unit is reduced by processing the learning data stored in the learning data storage unit. In this case, the amount of learning data stored in the learning data storage unit is reduced by making a portion of the dataset stored in the learning data storage unit sparse, or alternatively, by making data related to a portion of the input values ​​from the dataset stored in the learning data storage unit sparse.

[0061] Figure 9 This diagram shows a view of the functional configuration of an on-board processing device according to an embodiment of the present invention, in the case of learning the weights of a neural network NN on vehicle 1, i.e., in the case of performing on-board learning. Note that... Figure 9 The functions shown are performed in the electronic control unit 3 installed in vehicle 1. Furthermore, in Figure 9 The image shows the memory 6 of the electronic control unit 3. (Refer to...) Figure 9In this embodiment, the vehicle-mounted processing device includes: a data acquisition unit 40, which acquires data related to the operation of the vehicle 1; a neural network storage unit 41, which stores a neural network NN that outputs output values ​​related to the operation control of the vehicle if the data acquired at the data acquisition unit 40 is input; a learning data storage unit 42, which stores learning data of the weights of the neural network NN; a frequency acquisition unit 43, which acquires the learning frequency of the weights of the neural network NN in the vehicle; and a learning data modification unit 44.

[0062] In this embodiment, based on the learning frequency of the weights of the neural network NN in vehicle 1 acquired by the frequency acquisition unit 43, if the learning frequency of the weights of the neural network NN in vehicle 1 becomes lower, the learning data acquisition unit 44 reduces the amount of learning data stored per unit time in the learning data storage unit 42 or the amount of learning data stored in the learning data storage unit 42. Furthermore, in this embodiment, the on-board processing device includes: a learning unit 45 that performs the learning of the weights of the neural network NN; and a learning history storage device 46 that stores the learning history of the weights of the neural network NN in vehicle 1. In the frequency acquisition unit 43, the learning frequency is found from the learning history stored in the learning history storage unit 46.

[0063] Figure 10 It shows Figure 9 This is a view showing an example of the functional configuration of an onboard processing unit. Figure 10 In the example shown, as Figure 9 The learning data changing unit 44 shown uses a learning data control unit 44a, which controls the learning data acquired by the data acquisition unit 40. Figures 11 to 14 In the middle, it is shown that according to Figure 10 The functional configuration shown executes various processing routines. Therefore, these processing routines will be explained sequentially below.

[0064] Figure 11 This illustrates a storage routine for learning data when the learning frequency decreases during on-board learning, leading to a reduction in the acquisition frequency of learning data at data acquisition unit 40. This routine is suitable for... Figure 2The first example is shown. That is, as mentioned earlier, in the first example, typically, whenever the torque deviation ΔTt becomes an allowable value or smaller, the input values ​​x1, x2, x3, x4 when the torque deviation ΔTt becomes an allowable value or smaller are successively stored, and the target torque control value "y" (=y+C·ΔTt) output from the feedback correction section 25 is completed when the torque deviation ΔTt becomes an allowable value or smaller. That is, in this first example, typically, whenever the torque deviation ΔTt becomes an allowable value or smaller, the learning data is successively stored. In this case, in Figure 11 In the example shown, when the learning frequency decreases, learning data is not necessarily stored even if the torque deviation ΔTt becomes an acceptable value or smaller. Learning data is stored by determining the acquisition frequency of the learning data.

[0065] If reference Figure 11 In step 100, the learning history of the vehicle-mounted learning stored in the learning history storage unit 46 is read. Next, in step 101, the acquisition frequency of the learning data is determined from the learning history of the vehicle-mounted learning. In this case, the lower the learning frequency of the vehicle-mounted learning, the lower the acquisition frequency of the learning data acquired at the data acquisition unit 40. Next, in step 102, the learning data is acquired using the acquisition frequency determined in step 101. Next, in step 103, the acquired learning data is stored in the learning data storage unit 42. Next, in step 104, it is determined whether the learning conditions are met. If the learning conditions are not met, the routine returns to step 102, where the acquisition of learning data continues. On the other hand, if the learning conditions are met in step 104, the routine proceeds to step 105, where a learning start command is issued.

[0066] If the learning start instruction is issued, then at learning unit 45, execution will take place. Figure 12 The learning processing routine is shown below. If you refer to... Figure 12 In step 110, the number of nodes and weights of the input, hidden, and output layers of the neural network NN are read. Based on these node numbers, the following is prepared: Figure 4 The neural network NN is shown. Next, in step 111, the learning data stored in the learning data storage unit 42 is read, i.e., as shown... Figure 5The training dataset is shown. Next, in step 112, the weights of the neural network NN are learned using the aforementioned error backpropagation method based on this training dataset. Next, in step 113, it is determined whether the learning of the weights of the neural network NN has been completed. If it is determined that the learning of the weights of the neural network NN has not been completed, the routine returns to step 112, where the learning of the neural network weights continues. On the other hand, if it is determined in step 113 that the learning of the weights of the neural network NN has been completed, the routine proceeds to step 114, where the weights of the neural network NN are updated. Next, in step 115, the training data stored in the training data storage unit 42 is erased. If the training data is erased, the process is repeated. Figure 11 The example shown is a storage routine for learning data.

[0067] Figure 13 This illustrates the storage routine for the learning data when the learning frequency is lower during on-board learning, which increases the cycle length of the learning data acquisition loop acquired at the data acquisition unit 40. Specifically, it is used for... Figure 7B The illustrated embodiment demonstrates a routine for storing learning data. This routine is suitable for... Figure 3 The second example is shown. That is, as previously described, in the second example, the input values ​​x1, x2, x3, x4, x5 and the actual catalyst temperature Td detected by the temperature sensor 24b at that time are periodically and sequentially stored. That is, in the second example, learning data is periodically and sequentially stored. In this case, in Figure 13 In the example shown, as the learning frequency becomes lower, the cycle for acquiring learning data increases, and the cycle for storing learning data becomes longer.

[0068] If reference Figure 13 In step 120, the learning history of the vehicle-mounted learning stored in the learning history storage unit 46 is read. Next, in step 121, the cycle period for acquiring learning data is determined from the learning history of the vehicle-mounted learning. In this case, the lower the learning frequency of the vehicle-mounted learning, the longer the cycle period for acquiring learning data acquired in the data acquisition unit 40. Next, in step 122, learning data is acquired according to the cycle period determined in step 121. Next, in step 123, the acquired learning data is stored in the learning data storage unit 42. Next, in step 124, it is determined whether the learning conditions are met. If the learning conditions are not met, the routine returns to step 122, where the acquisition of learning data continues. On the other hand, if the learning conditions are met in step 124, the routine proceeds to step 125, where a learning start command is issued. If a learning start command is issued, execution is performed at the learning unit 45. Figure 12The learning processing routine shown.

[0069] Figure 14 This illustrates the storage routine for learning data when the learning frequency is lower during on-board learning, resulting in a decrease in the type of learning data acquired at the data acquisition unit 40; that is, the routine used for... Figure 7C The illustrated embodiment demonstrates a routine for storing learning data. In this case, in Figure 2 In the first example shown, if the learning frequency becomes lower, one or more of the input values ​​x1, x2, x3, and x4 are no longer acquired and stored, but... Figure 3 In the second example shown, if the learning frequency becomes lower, one or more of the input values ​​x1, x2, x3, x4, and x5 are no longer acquired and stored.

[0070] If reference Figure 14 In step 130, the learning history of the vehicle-mounted learning stored in the learning history storage unit 46 is read. Next, in step 131, the type of data to be acquired is determined from the learning history of the vehicle-mounted learning. In this case, the lower the learning frequency of the vehicle-mounted learning, the fewer types of learning data are acquired at the data acquisition unit 40. Next, in step 132, the learning data to be acquired as determined in step 131 is acquired. Next, in step 133, the acquired learning data is stored in the learning data storage unit 42. Next, in step 134, it is determined whether the learning conditions are met. If the learning conditions are not met, the routine returns to step 132, where the acquisition of learning data continues. On the other hand, if the learning conditions are met in step 134, the routine proceeds to step 135, where a learning start command is issued. If a learning start command is issued, execution is performed at the learning unit 45. Figure 12 The learning processing routine shown.

[0071] Figure 15 It shows Figure 9 A view showing the functional configuration of another example of an onboard processing unit. Figure 15 In the example shown, as Figure 9 The learning data modification unit 44 shown uses a learning data processing unit 44b for processing learning data stored in the learning data storage unit 42. In this example, the learning data stored in the learning data storage unit 42 is processed, such as... Figure 7DAs shown, the lower the learning frequency, the greater the reduction in the amount of learning data stored in the learning data storage unit 42. Note that in this case, as previously described, the amount of learning data stored in the learning data storage unit 42 is reduced by making a portion of the dataset stored in the learning data storage unit 42 sparse, or alternatively, by making the data related to a portion of the input values ​​from the dataset stored in the learning data storage unit 42 sparse.

[0072] exist Figure 16 In, it is shown that in Figure 15 The view showing the functional configuration demonstrates the data processing routine for learning, specifically, the data processing routine for learning data in the learning data storage unit 42 when the learning frequency is lower during on-board learning, thus reducing the amount of learning data stored. (Refer to...) Figure 16 In step 140, the learning history of the vehicle-mounted learning stored in the learning history storage unit 46 is read. Next, in step 141, learning data is acquired. Next, in step 142, the acquired learning data is stored in the learning data storage unit 42. Next, in step 143, it is determined whether the amount of data stored at the beginning of storing the learning data in the learning data storage unit 42 reaches a preset reference amount MX (< storage capacity MM), or whether the amount of data stored after processing the learning data reaches the preset reference amount MX. If it is determined that the amount of data stored at the beginning of storing the learning data in the learning data storage unit 42 or the amount of data stored after processing the learning data does not reach the reference amount MX, the routine jumps to step 145.

[0073] Conversely, when it is determined that the amount of data stored when the learning data is first stored in the learning data storage unit 42, or the amount of data stored after processing the learning data, reaches a reference value MX, the routine proceeds to step 144, where the learning data that has been stored in the learning data storage unit 42 is processed. At this time, the lower the learning frequency, the more learning data is stored in the learning data storage unit 42. In this case, as described above, the amount of learning data stored in the learning data storage unit 42 is reduced by making a portion of the dataset from the dataset that has been stored in the learning data storage unit 42 sparse, or alternatively, by making data related to a portion of the input values ​​from the dataset that has been stored in the learning data storage unit 42 sparse. Next, the routine proceeds to step 145. In step 145, it is determined whether the learning condition is met. When it is determined that the learning condition is not met, the routine returns to step 141, where the acquisition of learning data continues. On the other hand, when the learning condition is determined to be met in step 145, the routine proceeds to step 146, where a learning start command is issued. If a learning start command is issued, then at learning unit 45, execution... Figure 12 The learning processing routine shown.

[0074] Figure 17A This diagram shows a view of an example functional configuration of an in-vehicle processing device according to the invention, in which learning data stored in memory 6 of vehicle 1 is transferred to server 2 to perform weight learning of a neural network (NN) at server 2. Figure 17B This shows a view of the functional configuration of server 2. Note that in... Figure 17A The diagram shows the electronic control unit 3, memory 6, and communication unit 8 installed in vehicle 1. Figure 17B The diagram shows the electronic control unit 10, memory 13, and communication unit 14 located in server 2. First, if referring to... Figure 17A In this embodiment, the on-board processing device includes: a data acquisition unit 50, which acquires data related to the operation of the vehicle 1; a neural network storage unit 51, which stores a neural network NN that outputs output values ​​related to the operation control of the vehicle if the data acquired by the data acquisition unit 50 is input; a learning data storage unit 52, which stores learning data of the weights of the neural network NN; a frequency acquisition unit 53, which acquires the frequency at which the learning data is transmitted to the server 2; and a learning data modification unit 54. The learning data stored in the learning data storage unit 52 is transmitted to the server 2 by the communication unit 8.

[0075] In this embodiment, based on the frequency of transmitting learning data to server 2 obtained by frequency acquisition unit 53, if the frequency of transmitting learning data to server 2 becomes lower, the learning data modification unit 54 reduces the amount of learning data stored per unit time in learning data storage unit 52, or the amount of learning data stored in learning data storage unit 52. Furthermore, in this embodiment, the vehicle-mounted processing device is provided with a transmission history storage unit 55, which stores the transmission history of learning data to server 2. At frequency acquisition unit 53, the transmission frequency is found from the transmission history stored in transmission history storage unit 55.

[0076] On the other hand, if return Figure 17B In this embodiment, server 2 is provided with: a node count and weight storage unit 60, which stores the node count and weights of the neural network NN; a learning data storage unit 61, which stores learning data for the weights of the neural network NN; and a learning unit 62, which performs learning of the weights of the neural network NN. The learning data transmitted from vehicle 1 is received by communication unit 14.

[0077] Figure 18 Various processing routines are illustrated, including the transmission and reception processes performed at vehicle 1 and the learning processes performed at server 2. First, referring to the transmission processing routine performed at vehicle 1, in step 70, it is determined whether the conditions for transmitting learning data from vehicle 1 to server 2 are met. When the transmission conditions are met, the routine proceeds to step 71, where the number of nodes and weights of the neural network NN, along with the learning data stored in the learning data storage unit 52, are transmitted to server 2. Next, in step 72, the learning data stored in the learning data storage unit 52 is erased.

[0078] Next, referring to the learning processing routine performed at server 2, in step 80, it is determined whether the number of nodes and weights of the neural network NN, as well as the learning data, have been received from vehicle 1. When it is determined that the number of nodes and weights of the neural network NN, as well as the learning data, have been received, the routine proceeds to step 81, where the number of nodes and weights of the input layer, hidden layer, and output layer of the neural network NN are stored in the node number and weight storage unit 60. Based on these node numbers, preparation is made as follows... Figure 4 The neural network NN is shown. Next, in step 82, the received learning data is stored in the learning data storage unit 61. Next, in step 83, it is determined whether the amount of learning data stored in the learning data storage unit 61 exceeds a preset amount MS sufficient to perform learning. When it is determined that the amount of learning data stored in the learning data storage unit 61 exceeds the preset amount MS, the routine proceeds to step 84.

[0079] In step 84, the weights of the neural network NN are learned using the aforementioned backpropagation method based on the learning data stored in the learning data storage unit 42. Next, in step 85, it is determined whether the learning of the neural network NN's weights has been completed. If it is determined that the learning of the neural network NN's weights is not yet complete, the routine returns to step 84, where the learning of the neural network's weights continues. On the other hand, if it is determined in step 85 that the learning of the neural network NN's weights has been completed, the routine proceeds to step 86, where the weights of the neural network NN are updated, and a learning model is prepared using the neural network NN with the updated weights. Next, in step 87, the prepared learning model is transmitted to the vehicle 1 via the communication unit 14.

[0080] Next, referring to the receiving process performed at vehicle 1, in step 90, it is determined whether the learning model has been received from server 2. If it is determined that the learning model has been received from server 2, the routine proceeds to step 91, where the learning model is stored in neural network storage unit 51. If the learning model is stored in neural network storage unit 51, the operation control of vehicle 1 is performed using the learning module. For example, in the first example above, the target torque control value "y" is found using the learning model, while in the second example above, the estimated value Te of the catalyst temperature is found using the learning model.

[0081] Figure 19 It shows Figure 17A This is a view showing an example of the functional configuration of an onboard processing unit. Figure 19 In the example shown, as Figure 17A The learning data changing unit 54 shown uses a learning data control unit 54a, which controls the learning data acquired by the data acquisition unit 50. Figures 20 to 22 In the middle, it is shown that according to Figure 19 The functional configuration shown executes various processing routines. Therefore, these processing routines will be explained in turn.

[0082] Figure 20 This illustrates a storage routine for the learning data when the transmission frequency is lower, thus reducing the acquisition frequency of the learning data acquired at the data acquisition unit 50, in the case where the learning data is transmitted to server 2 for learning the weights of the neural network NN at server 2. This routine is suitable for... Figure 2The first example is shown. That is, as previously described, in the first example, typically, whenever the torque deviation ΔTt becomes an allowable value or smaller, the input values ​​x1, x2, x3, x4 when the torque deviation ΔTt becomes an allowable value or smaller, and the target torque control value "y" (=y+C·ΔTt) output from the feedback correction section 25 when the torque deviation ΔTt becomes an allowable value or smaller are stored successively. That is, in this first example, typically, whenever the torque deviation ΔTt becomes an allowable value or smaller, learning data is stored successively. In this case, in Figure 20 In the example shown, when the transmission frequency decreases, even if the torque deviation ΔTt becomes an acceptable value or smaller, the learning data may not necessarily be stored. The learning data is stored by determining the acquisition frequency of the learning data.

[0083] If reference Figure 20 In step 200, the transmission history of the learning data to server 2 stored in the transmission history storage unit 55 is read. Next, in step 201, the acquisition frequency of the learning data is determined from the transmission history of the learning data to server 2. In this case, the lower the transmission frequency of the learning data to server 2, the lower the acquisition frequency of the learning data acquired at the data acquisition unit 50. Next, in step 202, the learning data is acquired using the acquisition frequency determined in step 201. Next, in step 203, the acquired learning data is stored in the learning data storage unit 52. Next, in step 204, it is determined whether the transmission condition is met. If the transmission condition is not met, the routine returns to step 202, where the acquisition of the learning data continues. On the other hand, if the transmission condition is met in step 204, the processing loop ends. At this time, as shown in the previous step... Figure 18 As understood in the transmission processing routine executed at vehicle 1, the number of nodes and weights of the neural network NN, as well as the learning data stored at the learning data storage unit 52, are transmitted to server 2.

[0084] Figure 21 This illustrates a storage routine for the learning data when, in the case of transmitting learning data to server 2 for learning the weights of a neural network NN at server 2, a lower transmission frequency results in an increased cycle period for the learning data acquired at data acquisition unit 50. Specifically, this routine is used for... Figure 7B The illustrated embodiment demonstrates a routine for storing learning data. This routine is suitable for... Figure 3 The second example is shown. That is, as previously described, in the second example, the input values ​​x1, x2, x3, x4, x5 and the actual catalyst temperature Td detected by the temperature sensor 24b at that time are periodically and sequentially stored. In other words, in this second example, learning data is periodically and sequentially stored. In this case, in Figure 21 In the example shown, as the learning frequency becomes lower, the cycle for acquiring learning data increases, and the cycle for storing learning data becomes longer.

[0085] If reference Figure 21 In step 210, the transmission history of the learning data to server 2 stored in transmission history storage unit 55 is read. Next, in step 211, the period of the learning data acquisition cycle is determined from the transmission history of the learning data to server 2. In this case, the lower the transmission frequency of the learning data to server 2, the longer the period of the learning data acquisition cycle acquired at data acquisition unit 40. Next, in step 212, the learning data is acquired according to the acquisition cycle period determined in step 211. Next, in step 213, the acquired learning data is stored in learning data storage unit 52. Next, in step 214, it is determined whether the transmission condition is met. When the transmission condition is determined not to be met, the routine returns to step 212, where the acquisition of learning data continues. On the other hand, when the transmission condition is determined to be met in step 214, the processing loop ends. At this time, if the data is obtained from... Figure 18 As understood in the transmission processing routine executed at vehicle 1, the number of nodes and weights of the neural network NN, as well as the learning data stored in the learning data storage unit 52, are transmitted to server 2.

[0086] Figure 22 This illustrates a storage routine for the learning data when, in the case of transmitting learning data to server 2 to perform weight learning of a neural network NN at server 2, the lower the transmission frequency, the fewer the types of learning data acquired at data acquisition unit 50. Specifically, it describes a routine for storing learning data. Figure 7C The illustrated embodiment demonstrates a routine for storing learning data. In this case, in Figure 2 In the first example shown, if the learning frequency becomes lower, one or more of the input values ​​x1, x2, x3, and x4 are no longer acquired and stored. Figure 3 In the second example shown, if the learning frequency becomes lower, one or more of the input values ​​x1, x2, x3, x4, and x5 are no longer acquired and stored.

[0087] If reference Figure 22In step 220, the transmission history of the learning data to server 2 stored in transmission history storage unit 55 is read. Next, in step 221, the type to be acquired is determined from the transmission history of the learning data to server 2. In this case, the lower the transmission frequency of the learning data to server 2, the fewer types of learning data are acquired at data acquisition unit 50. Next, in step 222, the type of learning data to be acquired determined in step 221 is acquired. Next, in step 223, the acquired learning data is stored in learning data storage unit 52. Next, in step 224, it is determined whether the transmission condition is met. When the transmission condition is determined not to be met, the routine returns to step 222, where the acquisition of learning data continues. On the other hand, when the transmission condition is determined to be met in step 224, the processing loop ends. At this time, if the data is... Figure 18 As understood in the transmission processing routine executed in vehicle 1, the number of nodes and weights of the neural network NN, as well as the learning data stored in the learning data storage unit 52, are transmitted to server 2.

[0088] Figure 23 It shows Figure 17A A view showing the functional configuration of another example of an onboard processing unit. Figure 23 In the example shown, as Figure 17A The learning data modification unit 54 shown uses a learning data processing unit 54b for processing learning data stored in the learning data storage unit 52. In this example, the learning data stored in the learning data storage unit 52 is processed, such as... Figure 7D As shown, the learning frequency becomes lower, resulting in a greater reduction in the amount of learning data stored in the learning data storage unit 52. It should be noted that, in this case, as previously described, the amount of learning data stored in the learning data storage unit 52 is reduced by making a portion of the dataset stored in the learning data storage unit 52 sparse, or alternatively, by making the data related to a portion of the input values ​​from the dataset stored in the learning data storage unit 52 sparse.

[0089] exist Figure 24 In, it is shown that Figure 23 The learning data processing routine executed in the view of the functional configuration shown is, in the case of transmitting learning data to server 2 to perform weight learning of neural network NN in server 2, the lower the transmission frequency, the less learning data is stored in the learning data storage unit 52. Figure 24In step 230, the transmission history of the learning data stored in the transmission history storage unit 55 to the server 2 is read. Next, in step 231, the learning data is acquired. Next, in step 232, the acquired learning data is stored in the learning data storage unit 52. Next, in step 233, it is determined whether the amount of data stored in the learning data storage unit 52 since the beginning of the learning data storage reaches a preset reference amount MY, or after processing the learning data, it is determined whether the amount of data stored after processing the learning data reaches the preset reference amount MY. This reference amount MY is then set to be less than... Figure 16 The value of the reference quantity MX during onboard learning is shown. When it is determined that the amount of data stored when the learning data is first stored in the learning data storage unit 52, or the amount of data stored after processing the learning data, does not reach the preset reference quantity MY, the routine jumps to step 235.

[0090] Conversely, if it is determined that the amount of data stored in the learning data storage unit 52 since the beginning of the learning data storage, or the amount of data stored after processing the learning data, reaches a preset reference amount MY, the routine proceeds to step 234, where the processing of the learning data stored in the learning data storage unit 52 is performed. At this time, the lower the transmission frequency, the greater the reduction in the amount of learning data stored in the learning data storage unit 52. In this case, as described above, the amount of learning data stored in the learning data storage unit 52 is reduced by making a portion of the dataset stored in the learning data storage unit 52 sparse, or alternatively, by making data related to a portion of the input values ​​from the dataset stored in the learning data storage unit 52 sparse. Next, in step 235, it is determined whether the transmission condition is met. If it is determined that the transmission condition is not met, the routine returns to step 231, where the acquisition of learning data continues. On the other hand, when it is determined that the learning condition is met in step 235, the processing loop ends. At this time, as will be Figure 18 As understood in the transmission processing routine executed at vehicle 1, the number of nodes and weights of the neural network NN, as well as the learning data stored at the learning data storage unit 52, are transmitted to server 2.

Claims

1. An onboard processing device for learning data, wherein, The on-board processing device for learning the weights of a neural network, either on-board or at a server located outside the vehicle, includes: A data acquisition unit acquires data related to the operation of the vehicle. A neural network storage unit stores a neural network. If data acquired at the data acquisition unit is input, the neural network outputs an output value related to the operation control of the vehicle. A learning data storage unit stores learning data for the weights of the neural network. A frequency acquisition unit acquires either the learning frequency of the weights of the neural network on the vehicle or the transmission frequency of transmitting the learned data to the server, wherein the learning frequency is the number of times the data is learned per unit time, and the transmission frequency is the number of times the data is transmitted per unit time. The learning data modification unit, if the learning frequency of the weights of the neural network on the vehicle or the transmission frequency becomes lower, reduces the amount of learning data stored per unit time in the learning data storage unit or the amount of learning data completely stored in the learning data storage unit. Specifically, by changing the learning data acquired by the data acquisition unit, the learning data changing unit reduces the amount of learning data stored per unit time in the learning data storage unit. Specifically, by reducing the types of learning data acquired by the data acquisition unit, the learning data modification unit reduces the amount of learning data stored per unit time in the learning data storage unit, so that the lower the learning frequency or transmission frequency, the more types of learning data are reduced.

2. The vehicle-mounted data processing device for learning purposes according to claim 1, wherein, By processing the learning data stored in the learning data storage unit, the learning data modification unit reduces the amount of learning data stored in the learning data storage unit.

3. The vehicle-mounted data processing device for learning purposes according to claim 1, wherein, By changing the cycle of the learning data acquisition loop obtained by the data acquisition unit, the learning data modification unit reduces the amount of learning data stored per unit time in the learning data storage unit.

4. The vehicle-mounted data processing device for learning purposes according to claim 1, wherein, The on-board processing device for learning data includes a learning unit that performs the learning of the weights of the neural network.

5. The vehicle-mounted data processing device for learning purposes according to claim 1, wherein, The learning unit that performs the learning of the weights of the neural network is set up at the server.

6. The vehicle-mounted data processing device for learning purposes according to claim 1, wherein, The frequency acquisition unit finds the learning frequency from the learning history of the weights of the neural network on the vehicle.

7. The vehicle-mounted data processing device for learning purposes according to claim 1, wherein, The frequency acquisition unit finds the transmission frequency from the transmission history of the learning data to the server.

Images