Vehicle data prediction device and vehicle data prediction method

Abstract

The purpose of the present invention is to improve the accuracy of a future prediction value that is used for vehicle control. This vehicle data prediction device comprises: a future prediction value calculation unit that calculates a future prediction value; a future prediction value error calculation unit that calculates a future prediction value error as a prediction value of an error of the future prediction value; and a prediction value correction unit that corrects the future prediction value using the future prediction value error. The future prediction value calculation unit is constituted by a first machine learning model generated using a control parameter related to a vehicle control device and a correct value as training data. The future prediction value error calculation unit is constituted by a second machine learning model generated using the control parameter, the correct value, and the future prediction value as training data. A prediction period or a prediction execution interval for the future prediction value is variably set in accordance with the control parameter or the future prediction value error for the vehicle control device.

Description

Vehicle Data Prediction Device and Vehicle Data Prediction Method

[0001] The present invention relates to a vehicle data prediction device and a vehicle data prediction method.

[0002] In recent years, higher precision and efficiency of vehicle powertrain control have been demanded, and powertrain control that performs model predictive control using a neural network model has been studied. A neural network model is a mathematical model that mimics the mechanism of the human brain nerve circuit.

[0003] However, in order to improve the accuracy of the estimation of the neural network model, even if the learning of the neural network is repeatedly performed multiple times, the correct value (observed value) and the predicted value do not completely match, the accuracy of the estimation reaches a plateau, and an error remains. In order to improve the precision and efficiency of powertrain control, reduction of the error of the neural network model is required.

[0004] Patent Document 1 discloses a data prediction system and a data prediction method that model the tendency of the error amount of the prediction calculation result and correct the prediction calculation result. The data prediction system disclosed in Patent Document 1 calculates the future error from the difference between the "past observed value" and the "predicted value predicted in the past" and the difference between the "current predicted value" and the "current observed value".

[0005] Japanese Unexamined Patent Application Publication No. 2017-224268

[0006] However, since the data prediction system disclosed in Patent Document 1 calculates the future error using meteorological data such as temperature, humidity, solar radiation amount, wind speed, air pressure, and past direct observation data such as typhoon and event information, it is difficult to calculate the error of a variable indirectly estimated from parameters related to the vehicle such as the torque of an in-vehicle motor and the cooling water temperature, such as the temperature of the rotor of the in-vehicle motor. Also, in the estimation of future predicted values used for vehicle control, the appropriate future prediction period and prediction execution interval (time resolution) differ depending on the estimation target. Therefore, it has been difficult to improve the accuracy of future predicted values used for vehicle control with the technique described in Patent Document 1.

[0007] In view of the above-mentioned problems, the object of the present invention is to provide a vehicle data prediction device and a vehicle data prediction method that can improve the accuracy of future predicted values ​​used in vehicle control.

[0008] To solve the above problems and achieve the present objective, a vehicle data prediction device according to one aspect of the present invention calculates future predicted values ​​related to a vehicle control device that controls a vehicle. The vehicle data prediction device comprises a future predicted value calculation unit that calculates future predicted values, a future predicted value error calculation unit that calculates a future predicted value error as a predicted value of the error of the future predicted values, and a predicted value correction unit that corrects the future predicted values ​​using the future predicted value error. The future predicted value calculation unit is composed of a first machine learning model generated using control parameters and correct values ​​related to the vehicle control device as training data. The future predicted value error calculation unit is composed of a second machine learning model generated using control parameters, correct values, and future predicted values ​​as training data. The prediction period or prediction execution interval for the future predicted values ​​is variably set according to the control parameters or future predicted value error of the vehicle control device.

[0009] One aspect of the present invention is a vehicle data prediction method, which is a method for calculating future predicted values ​​related to a vehicle control device that controls a vehicle. In the vehicle data prediction method, first, a future predicted value calculation unit calculates future predicted values. Next, a future predicted value error calculation unit calculates a future predicted value error as a predicted value of the error in the future predicted values. Then, a predicted value correction unit corrects the future predicted values ​​using the future predicted value error. The future predicted value calculation unit is composed of a first machine learning model generated using control parameters and correct values ​​related to the vehicle control device as training data. The future predicted value error calculation unit is composed of a second machine learning model generated using control parameters, correct values, and future predicted values ​​as training data. The prediction period or prediction execution interval for the future predicted values ​​is variably set according to the control parameters or future predicted value error of the vehicle control device.

[0010] According to one aspect of the present invention, the accuracy of future prediction values ​​used in vehicle control can be improved. Other problems, configurations, and effects not mentioned above will be clarified by the following description of embodiments.

[0011] This is a block diagram showing the configuration of a vehicle control system according to the first embodiment. This is a block diagram showing an example of the hardware configuration of a vehicle data prediction device according to the first embodiment. This is a conceptual diagram showing the weights and biases of each neuron constituting a neural network model. This is a diagram showing an overview of the learning of a neural network model according to the first embodiment. This is a block diagram showing the functional configuration of a control unit (CPU) according to the first embodiment. This is a diagram illustrating the learning of the neural network of the future prediction value calculation unit according to the first embodiment. This is a block diagram showing the functional configuration of the future prediction value calculation unit according to the first embodiment. This is a diagram illustrating the learning of the neural network of the future prediction value error calculation unit according to the first embodiment. This is a graph showing future prediction values ​​and error values ​​according to the first embodiment. This is a diagram showing an example of the execution phase for estimating future prediction values ​​and error values ​​according to the first embodiment. This is a flowchart showing the vehicle data prediction process according to the first embodiment. This is a graph comparing future prediction values ​​according to the first embodiment, conventional future prediction values, and actual values. This is a diagram illustrating the functions of the periodicity determination unit and prediction execution interval setting unit according to the first embodiment. This is a diagram illustrating the functions of the periodicity verification unit and prediction period setting unit according to the first embodiment. This is a diagram illustrating the functions of the communication speed determination unit and communication setting unit according to the first embodiment. This is a block diagram showing the configuration of a vehicle control system according to the second embodiment.

[0012] <First Embodiment> The vehicle data prediction device and vehicle data prediction method according to the first embodiment will be described below. In each figure, common parts are denoted by the same reference numerals.

[0013] [Configuration of a vehicle control system using the cloud] First, the configuration of the vehicle control system according to the first embodiment will be described. Figure 1 is a block diagram showing the configuration of the vehicle control system according to the first embodiment.

[0014] As shown in Figure 1, the vehicle control system according to the first embodiment comprises a vehicle control device 1, various sensors 6 installed on the vehicle, and a cloud area 7. The vehicle control device 1 includes a powertrain control ECU (Electronic Control Unit) 2, a vehicle control ECU 3, a central control unit 4, and a communication unit 5.

[0015] The powertrain control ECU 2 controls the vehicle's motors and other components. The powertrain control ECU 2 also calculates future predicted values ​​related to the control of the powertrain control ECU 2 or the vehicle control ECU 3. The powertrain control ECU 2 corresponds to the vehicle data prediction device according to the present invention.

[0016] The vehicle control ECU 3 controls the vehicle's braking system, steering system, or drive system, etc. The vehicle control ECU 3 transmits information such as accelerator opening, brake pedal depression amount, and vehicle speed to the central control unit 4. The vehicle control ECU 3 corresponds to the vehicle control device according to the present invention.

[0017] Cloud Area 7 will store driving information such as the distance the vehicle will travel and the gradient of the roads it will be driving on. This driving information, including distance and gradient, can be obtained from map data. In addition, the driving information stored in Cloud Area 7 will also include data on temperature and the driver's driving tendencies.

[0018] The communication unit 5 transmits and receives data to and from the cloud area 7. For example, the communication unit 5 receives driving information stored in the cloud area 7. The communication unit 5 also transmits the driving information to the central control unit 4. Various sensors 6 acquire sensor values ​​such as motor cooling water temperature and GPS signals and transmit them to the central control unit 4.

[0019] The central control unit 4 transmits and receives data between the powertrain control ECU 2 and the vehicle control ECU 3. The central control unit 4 transmits driving information and device operation information to the powertrain control ECU 2. The central control unit 4 transmits the estimated motor torque (motor torque), rotational speed, and driving information to the vehicle control ECU 3.

[0020] [Powertrain Control ECU] Next, the configuration of the powertrain control ECU 2 will be described. Figure 2 is a block diagram showing an example of the hardware configuration of the powertrain control ECU 2.

[0021] As shown in Figure 2, the powertrain control ECU 2 includes an input circuit 101, input / output ports 102, RAM 103, ROM 104, CPU 105, output circuit 106, and storage unit 107.

[0022] Signals output by the vehicle control ECU 3, communication unit 5, and various sensors 6 are input to the input circuit 101 via the central control unit 4. The signals input to the input circuit 101 are sent to the RAM 103 via the input / output port 102. The RAM 103 stores the signals sent via the input / output port 102. ROM 104 is used as an example of a computer-readable, non-transient storage medium that stores a control program executed by the CPU 105.

[0023] The CPU 105 processes the output signals stored in the RAM 103 according to the control program by executing a control program (not shown) stored in the ROM 104. The CPU 105 calculates future predicted values ​​related to the control of the vehicle control ECU 3 according to the control program. The CPU 105 temporarily stores the calculation results in the RAM 103. The calculation results stored in the RAM 103 are output to the output circuit 106 via the input / output port 102.

[0024] The memory unit 107 is electrically connected to the RAM 103, ROM 104, and CPU 105 via the input / output port 102. The memory unit 107 can employ storage devices such as hard disk drives and solid-state drives, internal computer storage areas, and network storage areas. For example, the memory unit 107 stores future prediction values ​​est1 and future prediction values ​​est2, which will be described later.

[0025] [Neural Network Model] Next, the neural network model according to this embodiment will be described with reference to Figures 3 and 4. Figure 3 is a conceptual diagram showing the weights and biases of each neuron constituting the neural network model. Figure 4 is a conceptual diagram showing the neural network model.

[0026] A neural network model is a mathematical model that mimics the structure of the human brain's neural circuits. As shown in Figure 3, each neuron in a neural network model is assigned a weight w and a bias b.

[0027] In a neural network model, n neurons each receive inputs a1 to an, and multiply them by weights w1 to wn set for each neuron. Then, in the next layer of neurons, the inputs a1 to an, multiplied by weights w1 to wn, are added (combined), and a bias b is applied to the sum to obtain output z. The next layer of neurons outputs a, which is represented by the function f(z).

[0028] Furthermore, each neuron has a function called an activation function defined for it. The activation function can be a logistic function (sigmoid function) or a ramp function (ReLU (Rectified Linear Unit) function), etc., as appropriate. Figure 11 shows an example where the greater the input x is than 0, the more the neuron is activated and the output y (= f(x)) approaches 1, and the less the input x is than 0, the more the neuron is deactivated and approaches 0. For example, when the input x is "5", the output y of the activation function becomes "1", and this neuron outputs "1" to the next layer of neurons.

[0029] As shown in Figure 4, a neural network model is formed by multiple neurons in a single layer. A neural network model has an input layer into which explanatory variables are input, an output layer into which the target variable is input or output, and an intermediate layer located between the input and output layers.

[0030] The hidden layers are constructed by stacking multiple layers, each made up of multiple neurons. Neural network models can approximate more complex input-output relationships by increasing the number of neurons and hidden layers. However, there is a trade-off between approximation accuracy and model size, so a balance point that satisfies both requirements is selected.

[0031] As shown in the upper part of Figure 4, the neural network model of this embodiment uses machine learning (supervised learning) to learn the weights and biases of each neuron in the hidden layer by inputting explanatory variables into the input layer and the correct values ​​corresponding to the target variable into the output layer. This allows for approximation of the input-output relationship. For the machine learning algorithm, known algorithms such as backpropagation can be applied. Furthermore, as shown in the lower part of Figure 4, the neural network model, whose input-output relationship between the target variable and the correct values ​​has been approximated by machine learning, outputs (estimates) a target variable that approximates the learned correct values ​​by inputting explanatory variables into the input layer.

[0032] [Functional Configuration of CPU 105] Next, the functional configuration of the CPU 105 in the powertrain control ECU 2 will be described. Figure 5 is a block diagram showing the functional configuration of the CPU 105.

[0033] As shown in Figure 5, the CPU 105 includes a future prediction value calculation unit 11, a previous value output unit 12, a future prediction value error calculation unit 13, and a prediction value correction unit 14.

[0034] The future prediction value calculation unit 11 is composed of a neural network. The neural network constituting the future prediction value calculation unit 11 corresponds to the first machine learning model according to the present invention. Driving information and device operation information are input to the future prediction value calculation unit 11. Based on this, the future prediction value calculation unit 11 calculates a future prediction value (target variable) for a predetermined step ahead, using the driving information and device operation information as explanatory variables. The future prediction value calculation unit 11 outputs the calculated future prediction value to the previous value output unit 12 and the prediction value correction unit 14.

[0035] The predetermined step ahead is, for example, n (n > 0) seconds later, and is set according to the control parameters of the vehicle control ECU 3 or the future prediction error described later. The predetermined step ahead corresponds to the prediction period described later.

[0036] Driving information includes information supplied from the cloud area 7 and information obtained as sensor values. Device operation information includes information obtained as sensor values, command values, current values, etc. Driving information and device operation information correspond to control parameters according to the present invention.

[0037] Sensor values ​​are information obtained from various sensors 6. Examples of sensor values ​​include motor coolant temperature, vehicle speed, motor drive current, accelerator opening, and brake pedal depression. Command values ​​are values ​​related to control commands output by the powertrain control ECU 2 and vehicle control ECU 3. An example of a command value is the duty cycle of the coolant pump.

[0038] The current value is an estimated value obtained by calculation using a physical formula. For example, the motor torque value can be detected using a torque measuring instrument. However, some vehicles do not have a torque measuring instrument installed from a cost reduction perspective. In that case, the control unit (e.g., CPU 105) estimates the current motor torque value based on the accelerator opening and the motor rotation speed. In addition to the motor torque value mentioned above, the motor rotation speed can also be used as a current value. The motor rotation speed is calculated, for example, from the number of rotation pulses detected by a motor resolver.

[0039] The driving information supplied from the cloud area 7 includes, for example, the distance the vehicle is expected to travel, the gradient of the road the vehicle will travel on, the elevation of the road the vehicle will travel on, and the curvature of the road the vehicle will travel on. The driving information obtained as sensor values ​​includes the accelerator pedal position and the amount the brake pedal is pressed. The device operation information obtained as sensor values ​​includes the motor coolant temperature, vehicle speed, and motor drive current.

[0040] The previous value output unit 12 outputs the previous future predicted value to the future predicted value error calculation unit 13. The future predicted value error calculation unit 13 is constituted by a neural network. The neural network constituting the future predicted value error calculation unit 13 corresponds to the second machine learning model according to the present invention. The previous future predicted value, market information, and device operation information are input to the future predicted value error calculation unit 13. The device operation information includes at least one of the sensor value and the current value that are the correct values.

[0041] The future predicted value error calculation unit 13 calculates a future predicted value error (objective variable) using the previous future predicted value, market information, and device operation information including the correct value as explanatory variables. The future predicted value error is a prediction of the error of the future predicted value. The future predicted value error calculation unit 13 outputs the calculated future predicted value error to the predicted value correction unit 14.

[0042] The predicted value correction unit 14 corrects the future predicted value using the future predicted value error. Specifically, the future predicted value error is added to the future predicted value to calculate the corrected future predicted value. The predicted value correction unit 14 outputs the calculated corrected future predicted value to the output circuit 106 (see FIG. 2). Thereafter, the corrected future predicted value is transmitted from the output circuit 106 to the vehicle control ECU 3 via the central control device 4 (see FIG. 1).

[0043] [Neural Network of Future Predicted Value Calculation Unit] Next, the learning of the neural network in the future predicted value calculation unit 11 will be described. FIG. 6 is a diagram for explaining the learning of the neural network in the future predicted value calculation unit 11.

[0044] As shown in FIG. 6, the neural network of the future predicted value calculation unit 11 performs machine learning in advance by inputting the market information and device operation information of the vehicle as explanatory variables into the input layer and inputting the correct value of the future predicted value corresponding to the objective variable into the output layer. That is, the neural network of the future predicted value calculation unit 11 learns the variables necessary for calculating the future predicted value.

[0045] For example, when the target variable is the current motor torque value, the current motor torque value is input to the output layer, and as explanatory variables, the current travel distance, current gradient, current elevation, current curvature, current accelerator opening, and current brake depression amount are input to the input layer, thereby performing machine learning of the neural network. As a result, the neural network of the future prediction value calculation unit 11 is trained to be able to calculate the current motor torque value as the target variable.

[0046] Further, the neural network of the future prediction value calculation unit 11 trained to be able to calculate the current motor torque value calculates the motor torque value (future prediction value) after 5 seconds by inputting the travel distance after 5 seconds, gradient after 5 seconds, elevation after 5 seconds, curvature after 5 seconds, accelerator opening after 5 seconds, and brake depression amount after 5 seconds to the input layer. Note that the combination of explanatory variables when calculating the motor torque value is an example, and it is also possible to use other variables with a large influence on the motor torque as explanatory variables.

[0047] When the target variable is the motor temperature, as explanatory variables, for example, the current motor rotation speed, current motor torque value, current motor cooling water temperature, current vehicle speed, current motor drive current, and current duty of the cooling water pump are input to the input layer, and the correct value of the motor temperature is input to the output layer, thereby performing machine learning in advance. As a result, the neural network of the future prediction value calculation unit 11 is trained to be able to calculate the current motor temperature as the target variable.

[0048] Further, the neural network of the future prediction value calculation unit 11 trained to be able to calculate the current motor temperature value calculates the motor temperature (future prediction value) after 5 seconds when the motor rotation speed after 5 seconds, motor torque value after 5 seconds, motor cooling water temperature after 5 seconds, vehicle speed after 5 seconds, motor drive current after 5 seconds, and duty of the cooling water pump after 5 seconds are input to the input layer. Note that the combination of explanatory variables when calculating the motor temperature is an example, and it is also possible to use other variables with a large influence on the motor temperature as explanatory variables.

[0049] [Functional Configuration of the Future Prediction Calculation Unit] Next, the functional configuration of the future prediction value calculation unit 11 will be described. Figure 7 is a block diagram showing the functional configuration of the future prediction value calculation unit 11.

[0050] As shown in Figure 7, the future prediction value calculation unit 11 includes a current value estimation unit 111 and a future prediction unit 112.

[0051] The current value estimation unit 111 is composed of a neural network. The current value estimation unit 111 estimates the current value of the target variable (for example, motor torque) corresponding to predetermined explanatory variables. The future prediction value calculation unit 11 calculates the current motor torque value using, for example, the current mileage, current gradient, current elevation, current curvature, current accelerator opening, and current brake depression as explanatory variables. The current value estimation unit 111 outputs the calculated current motor torque value to the future prediction unit 112 and the future prediction value error calculation unit 13.

[0052] The future prediction unit 112 is composed of a neural network. The future prediction unit 112 calculates a target variable (for example, motor torque after 5 seconds) corresponding to predetermined explanatory variables as a future prediction value. For example, the future prediction unit 112 calculates the motor torque value after 5 seconds using the current values ​​of the distance traveled after 5 seconds, the gradient after 5 seconds, the elevation after 5 seconds, the curvature after 5 seconds, the accelerator opening after 5 seconds, the brake depression amount after 5 seconds, and the motor torque as explanatory variables. The future prediction unit 112 outputs the calculated motor torque value after 5 seconds to the previous value output unit 12 and the prediction value correction unit 14 (Figure 5).

[0053] [Neural Network in Future Prediction Error Estimation Unit] Next, we will explain the training of the neural network in the future prediction value error calculation unit 13. Figure 8 is a diagram illustrating the training of the neural network in the future prediction value error calculation unit 13.

[0054] The future prediction error calculation unit 13 is composed of a neural network separate from the current value estimation unit 111 and the future prediction unit 112. As shown in the middle of Figure 8, the neural network of the future prediction error calculation unit 13 performs machine learning in advance by inputting predetermined explanatory variables into the input layer and inputting the prediction error (hereinafter referred to as "prediction error"), which is the difference between previously obtained prediction values ​​and the current correct value, into the output layer. As a result, the neural network of the future prediction error calculation unit 13 is trained to calculate the prediction error as the target variable.

[0055] For example, if the target variable of the neural network in the future prediction value calculation unit 11 is the motor torque value, the error value of the current motor torque value, which is the difference between the current motor torque value (correct value) calculated by the current value estimation unit 111 and the motor torque value 5 seconds later (est1) calculated by the future prediction unit 112 5 seconds ago, is input to the output layer of the neural network in the future prediction value error calculation unit 13. In addition, as explanatory variables, for example, the motor torque value 5 seconds later (est1) calculated by the future prediction unit 112 5 seconds ago, the current motor torque value (correct value) calculated by the current value estimation unit 111, the current mileage, the current gradient, the current altitude, the current curvature, the current accelerator opening, and the current brake depression amount are input to the input layer of the neural network in the future prediction value error calculation unit 13, thereby performing machine learning of the neural network. As a result, the neural network in the future prediction value error calculation unit 13 is learned to calculate the prediction error of the current motor torque as the target variable.

[0056] Furthermore, the neural network of the future prediction error calculation unit 13, which has been trained to calculate the prediction error of the current motor torque, calculates the prediction error of the motor torque 5 seconds later (future prediction error value) by inputting the current motor torque value (correct value) calculated by the current value estimation unit 111, the latest motor torque value 5 seconds later (est2) currently calculated by the future prediction unit 112, the distance traveled 5 seconds later, the gradient 5 seconds later, the elevation 5 seconds later, the curvature 5 seconds later, the accelerator opening 5 seconds later, and the brake depression amount 5 seconds later into the input layer. Note that the combination of explanatory variables used to calculate the prediction error of the motor torque is just one example, and it is also possible to use other variables that have a large influence on the prediction error of the motor torque as explanatory variables.

[0057] When the target variable of the neural network in the future prediction value calculation unit 11 is motor temperature, the error value of the current motor temperature, which is the difference between the current motor temperature value (correct value) calculated by the current value estimation unit 111 and the motor temperature value 5 seconds later (est1) calculated by the future prediction unit 112 5 seconds ago, is input to the output layer of the neural network in the future prediction value error calculation unit 13. In addition, as explanatory variables, for example, the motor temperature value 5 seconds later (est1) calculated by the future prediction unit 112 5 seconds ago, the current motor temperature value calculated by the current value estimation unit 111, the current motor rotation speed, the current motor torque value, the current motor coolant temperature, the current vehicle speed, the current motor drive current, and the current coolant pump duty cycle are input to the input layer, thereby executing machine learning of the neural network. As a result, the neural network in the future prediction value error calculation unit 13 is learned to calculate the prediction error of the current motor temperature value as the target variable.

[0058] Furthermore, the neural network of the future prediction error calculation unit 13, which has been trained to calculate the prediction error of the current motor temperature, inputs the current motor temperature value (correct value) calculated by the current value estimation unit 111, the latest motor temperature value 5 seconds later (est2) currently calculated by the future prediction unit 112, the current motor rotation speed, the current motor torque value, the current motor coolant temperature, the current vehicle speed, the current motor drive current, and the current coolant pump duty cycle into its input layer to calculate the prediction error of the motor temperature 5 seconds later (future prediction error value). Note that the combination of explanatory variables used to calculate the prediction error of the motor temperature is just one example, and it is also possible to use other variables that have a large influence on the prediction error of the motor temperature as explanatory variables.

[0059] [Execution Phase] Next, the execution phase of this embodiment will be described. Figure 9 is a graph showing future predicted values ​​and error values ​​according to the first embodiment. Figure 10 is a diagram showing an example of the execution phase for estimating future predicted values ​​and error values ​​according to the first embodiment.

[0060] As shown in Figures 9 and 10, in the execution phase of this embodiment, for example, at the 5-second mark, the future prediction unit 112 calculates (estimates) the future predicted value est1 for the 10-second mark, which is 5 seconds later. Next, when the actual time reaches the 10-second mark, the current value estimation unit 111 calculates (estimates) the current value (correct value) for the 10-second mark. Furthermore, at the 10-second mark, the future prediction unit 112 calculates (estimates) the future predicted value est2 for the 15-second mark, which is 5 seconds later. Then, the future predicted value est1 calculated at the 5-second mark (5 seconds ago), the correct value calculated by the current value estimation unit 111 at the 10-second mark, and the future predicted value est2 calculated by the future prediction unit 112 at the 10-second mark for the 15-second mark (5 seconds later) are input to the neural network of the future prediction value error calculation unit 13.

[0061] Subsequently, at the 10-second mark, the future prediction error calculation unit 13 calculates (estimates) the future prediction error esterr at the 15-second mark, which is 5 seconds after the 10-second mark. Then, at the 10-second mark, the prediction value correction unit 14 corrects the future prediction value est2 at the 15-second mark using the future prediction error esterr at the 15-second mark. This improves the accuracy of the future prediction value est2.

[0062] [Vehicle Data Prediction Processing] Next, the vehicle data prediction processing according to the first embodiment will be described. Figure 11 is a flowchart of the vehicle data prediction processing according to the first embodiment.

[0063] First, the future prediction value calculation unit 11 and the future prediction value error calculation unit 13 receive input data (explanatory variables) (S1). Next, the future prediction value calculation unit 11 calculates the future prediction value for a predetermined step ahead (S2). Next, the future prediction value error calculation unit 13 calculates the future prediction value error for a predetermined step ahead (S3).

[0064] Next, the prediction value correction unit 14 corrects the future prediction value using the future prediction value error (S4). Then, the CPU 105 outputs the corrected future prediction value to the vehicle control ECU 3 (S5).

[0065] [Predicted and Actual Values] Next, we compare the predicted values ​​according to the first embodiment, the conventional predicted values, and the actual values. Figure 12 is a graph comparing the predicted values ​​according to the first embodiment, the conventional predicted values, and the actual values.

[0066] The upper graph in Figure 12 compares the uncorrected future predicted value, the corrected future predicted value, and the actual value. The future predicted value is, for example, the motor torque value after 5 seconds. The horizontal axis of the upper graph in Figure 12 represents time, and the vertical axis represents the motor torque value. The horizontal axis of the lower graph in Figure 12 also represents time, and the vertical axis represents the error between the future predicted value and the actual value.

[0067] As shown in the upper graph of Figure 12, at 6 seconds, the uncorrected and corrected future predictions are significantly different. At this point, the corrected future predictions show an improvement of approximately 20 Nm in error compared to the uncorrected future predictions (see the lower graph of Figure 12).

[0068] The actual value at 6 seconds is approximately 50 Nm. Generally, the allowable range for torque error is 20%. Therefore, the allowable range for torque error at 6 seconds is 10 Nm. The uncorrected predicted future value at 6 seconds is approximately 20 Nm, which exceeds the above allowable range. On the other hand, the corrected predicted future value at 6 seconds is approximately 1.5 Nm, which is within the above allowable range. Therefore, the accuracy of the corrected predicted future value can be improved.

[0069] [Periodicity Determination Unit and Predicted Execution Interval Setting Unit] Next, the periodicity determination unit and the predicted execution interval setting unit of the CPU 105 will be described. Figure 13 is a diagram illustrating the functions of the periodicity determination unit and the predicted execution interval setting unit.

[0070] As shown in Figure 13, the CPU 105 has a periodicity determination unit 131 and an execution interval setting unit 132.

[0071] The periodicity determination unit 131 calculates the difference between the current value of a predetermined control parameter and the predicted value after a specific time has elapsed. The periodicity determination unit 131 then determines that there is periodicity if the calculated difference is less than or equal to an arbitrarily set determination threshold. The periodicity determination unit 131 determines that there is no periodicity if the calculated difference is greater than the determination threshold. If the periodicity determination unit 131 determines that there is periodicity, it outputs "0" to the prediction execution interval setting unit 132, and if it determines that there is no periodicity, it outputs "1" to the prediction execution interval setting unit 132.

[0072] The periodicity determination unit 131 calculates, for example, the difference between the current accelerator opening and the accelerator opening at a predetermined step ahead. If the accelerator opening is constant and the difference is less than or equal to a determination threshold (for example, close to zero), it determines that the vehicle exhibits periodicity, such as during steady-state driving. On the other hand, if the difference is greater than the determination threshold, it determines that the accelerator opening is being opened and closed, resulting in transient operation and lack of periodicity.

[0073] The prediction execution interval setting unit 132 transmits an ON signal for the periodic prediction execution interval to the future prediction value calculation unit 11 and the future prediction value error calculation unit 13 when it is determined that there is periodicity. The periodic prediction execution interval is the set value for the interval at which future prediction values ​​are estimated when periodicity is determined to be present. The prediction execution interval setting unit 132 transmits an ON signal for the non-periodic prediction execution interval to the future prediction value calculation unit 11 and the future prediction value error calculation unit 13 when it is determined that there is no periodicity. The non-periodic prediction execution interval is the set value for the interval at which future prediction values ​​are estimated when periodicity is determined to be absent.

[0074] When there is periodicity, the change in the predicted future value per unit time is small, while when there is no periodicity, the change in the predicted future value per unit time is large. Therefore, the prediction execution interval for non-periodic data is set to a smaller value than the prediction execution interval for periodic data. In other words, when there is no periodicity, the predicted future value is calculated more frequently than when there is periodicity. This increases the time resolution of the predicted future value and improves the effective accuracy. Also, when there is periodicity, the calculation frequency can be reduced compared to when there is no periodicity, thereby reducing the computational load on the powertrain control ECU2 related to the predicted future value.

[0075] [Periodicity Determination Unit and Prediction Period Setting Unit] Next, the periodicity determination unit and prediction period setting unit of the CPU 105 will be described. Figure 14 is a diagram illustrating the functions of the periodicity determination unit and prediction period setting unit.

[0076] As shown in Figure 14, the CPU 105 has a periodicity determination unit 141 and a prediction period setting unit 142.

[0077] The periodicity determination unit 141 is the same as the periodicity determination unit 131 described above, and calculates the difference between the current value of a predetermined control parameter and the predicted value after a specific time has elapsed. The periodicity determination unit 141 determines that there is periodicity if the calculated difference is less than or equal to an arbitrarily set determination threshold. The periodicity determination unit 141 determines that there is no periodicity if the calculated difference is greater than the determination threshold. If the periodicity determination unit 141 determines that there is periodicity, it outputs "0" to the prediction period setting unit 142, and if it determines that there is no periodicity, it outputs "1" to the prediction period setting unit 142.

[0078] If the prediction period setting unit 142 determines that there is periodicity, it transmits a prediction period for periodicity (a value at a predetermined step ahead) to the future prediction value calculation unit 11 and the future prediction value error calculation unit 13. If the prediction period setting unit 142 determines that there is no periodicity, it transmits a prediction period for non-periodicity (a value at a predetermined step ahead) to the future prediction value calculation unit 11 and the future prediction value error calculation unit 13.

[0079] When periodicity is present, the change in future predicted values ​​over a given period is small, while when periodicity is absent, the change in future predicted values ​​over a given period is large. Therefore, the prediction period when periodicity is absent is set to a shorter value than the prediction period when periodicity is present. When periodicity is absent, future predicted values ​​are calculated for a shorter period (the near future) than when periodicity is present. This allows for improved effective accuracy of future predictions in response to changes in future predicted values.

[0080] Furthermore, if the periodicity determination unit 141 determines that periodicity exists, it may calculate the period based on a value obtained by multiplying the number of times periodicity has been determined to exist continuously by a predetermined periodicity determination execution period of the periodicity determination unit 141. The prediction period setting unit 142 may then set the prediction period for the future predicted value according to the calculated period (the duration of the state in which the future predicted value does not change much). This makes it possible to improve the effective accuracy of the future predicted value.

[0081] [Communication Speed ​​Determination Unit and Communication Setting Unit] Next, the communication speed determination unit and communication setting unit of the CPU 105 will be described. Figure 15 is a diagram illustrating the functions of the communication speed determination unit and communication setting unit.

[0082] As shown in Figure 15, the CPU 105 includes a communication speed determination unit 151 and a communication setting unit 152.

[0083] The communication speed determination unit 151 determines that there is a delay in communication if the communication speed between the cloud area 7 and the communication unit 5 is delayed to a degree that exceeds the prediction period set by the prediction period setting unit 142. If the communication speed determination unit 151 determines that there is a delay in communication, it outputs "0" to the communication setting unit 152, and if it does not determine that there is a delay in communication, it outputs "1" to the communication setting unit 152.

[0084] If the communication speed determination unit 151 determines that there is a delay in communication, the communication setting unit 152 transmits the first prediction period to the future prediction value calculation unit 11 and the future prediction value error calculation unit 13.

[0085] If there is a delay in communication, in this embodiment, there will be a delay in updating the driving information such as driving distance and gradient stored in the cloud area 7, that is, the explanatory variables of the neural networks of the future prediction value calculation unit 11 and the future prediction value error calculation unit 13. If such a communication delay causes the interval between updating (acquiring) the explanatory variables to be longer than the prediction period set by the prediction period setting unit 142, a gap in prediction will occur in the period from the time when the future prediction value is predicted to the time when the explanatory variables are updated. For this reason, the first prediction period is set to a value longer than the prediction period (prediction period for periodic or non-periodic prediction period) set by the prediction period setting unit 142. In addition, the first prediction period is set variably according to the communication speed. This makes it possible to prevent a gap in prediction when there is a delay in communication.

[0086] If there is a delay in communication, the first predicted execution interval output by the communication speed determination unit 151 takes precedence over the predicted period set by the prediction period setting unit 142 (see Figure 14) described above. If there is no delay in communication, the predicted period set by the prediction period setting unit 142 (prediction period for periodic or prediction period for non-periodic use) takes precedence. This ensures that there are no gaps in the prediction even if there is a delay in communication between the cloud area 7 and the communication unit 5.

[0087] Furthermore, if the communication speed determination unit 151 determines that there is a delay in communication, the communication setting unit 152 transmits the first predicted execution interval to the future predicted value calculation unit 11 and the future predicted value error calculation unit 13.

[0088] The first predicted execution interval, which is set when there is a delay in communication, is set to a value greater than the predicted execution interval (predicted execution interval for periodic use or predicted execution interval for non-periodic use) set by the predicted execution interval setting unit 132 (see Figure 13) described above. Furthermore, the first predicted execution interval is set to be variable according to the communication speed.

[0089] If there is a delay in communication, the first predicted execution interval output by the communication setting unit 152 takes precedence over the predicted execution interval set by the predicted execution interval setting unit 132. If there is no delay in communication, the predicted execution interval set by the predicted execution interval setting unit 132 (either a periodic or non-periodic prediction execution interval) is applied. This allows for the calculation of future predicted values ​​even if there is a delay in communication between the cloud area 7 and the communication unit 5.

[0090] The powertrain control ECU 2 (vehicle data prediction device) of this embodiment variably sets the prediction execution interval and prediction period of future predicted values ​​according to predetermined control parameters. However, the vehicle data prediction device according to the present invention may variably set the prediction execution interval and prediction period of future predicted values ​​according to the error in future predicted values.

[0091] When a vehicle operates with periodicity, such as during steady-state driving, the predicted future values ​​do not fluctuate significantly, resulting in a smaller error in the predicted future values. On the other hand, when a vehicle operates with transient, non-periodic conditions, the predicted future values ​​fluctuate significantly, leading to a larger error in the predicted future values. Therefore, when the error in the predicted future values ​​is large, the prediction interval and prediction period are shortened, and when the error is small, the prediction interval and prediction period are lengthened. In other words, the prediction interval and prediction period are variably set to be inversely proportional to the error in the predicted future values. This improves the effective accuracy of the predicted future values.

[0092] <Second Embodiment> [Configuration of a Vehicle Control System that Does Not Utilize the Cloud] Next, a vehicle control system according to the second embodiment will be described. Figure 16 is a block diagram showing the configuration of the vehicle control system according to the second embodiment.

[0093] As shown in Figure 16, the vehicle control system according to the second embodiment includes a vehicle control device 21 and various sensors 26 installed on the vehicle. In other words, the vehicle control system according to the second embodiment does not include a cloud area. The vehicle control device 21 includes a powertrain control ECU 22, a vehicle control ECU 23, and a central control device 24.

[0094] The powertrain control ECU 22 and the vehicle control ECU 23 are the same as the powertrain control ECU 2 and the vehicle control ECU 3 in the first embodiment. The powertrain control ECU 22 performs the same processing as the powertrain control ECU 2 in the first embodiment to calculate and correct various future predicted values.

[0095] The central control unit 24 acquires driving information from other devices mounted on the vehicle (for example, a known navigation device installed in the vehicle). As described above, the driving information includes, for example, the distance the vehicle is scheduled to travel, the gradient of the road the vehicle will travel on, the elevation of the road the vehicle will travel on, and the curvature of the road the vehicle will travel on.

[0096] The present invention is not limited to the embodiments described above and shown in the drawings, and various modifications can be made without departing from the spirit of the invention as described in the claims. For example, in the embodiments described above, the system is configured to improve the estimation accuracy of the future predicted values ​​by calculating the predicted future values ​​of motor torque or motor temperature and the error in the predicted values ​​of motor torque or motor temperature, but the present invention is not limited thereto. For example, the system may be configured to improve the estimation accuracy of the target variables by calculating the error in the predicted values ​​of the explanatory variables such as distance traveled, gradient, elevation, curvature, accelerator opening, brake depression amount, motor rotation speed, motor torque value, motor coolant temperature, vehicle speed, motor drive current, and coolant pump duty cycle, which are input to the input layer of the neural network described above. This makes it possible to further improve the estimation accuracy of the future predicted values.

[0097] Furthermore, the embodiments described above are explained in detail for the purpose of clearly illustrating the present invention, and are not necessarily limited to those comprising all the described configurations. It is also possible to replace parts of the configuration of one embodiment with those of another embodiment, and to add configurations from other embodiments to the configuration of one embodiment. Additionally, it is possible to add, delete, or replace parts of the configuration of each embodiment with those of other embodiments.

[0098] 1, 21... Vehicle control device, 2, 22... Powertrain control ECU (vehicle data prediction device), 3, 23... Vehicle control ECU, 4, 24... Central control device, 5... Communication unit, 6, 26... Sensor, 7... Cloud area, 11... Future prediction value calculation unit, 12... Previous value output unit, 13... Future prediction value error calculation unit, 14... Prediction value correction unit, 101... Input circuit, 102... Input / output port, 103... RAM, 104... ROM, 105... CPU, 106... Output circuit, 107... Memory unit, 111... Current value estimation unit, 112... Future prediction unit, 131... Periodicity determination unit, 132... Prediction execution interval setting unit, 141... Periodicity determination unit, 142... Prediction period setting unit, 151... Communication speed determination unit, 152... Communication setting unit

Claims

1. A vehicle data prediction device for calculating future predicted values ​​related to a vehicle control device that controls a vehicle, comprising: a future predicted value calculation unit for calculating future predicted values; a future predicted value error calculation unit for calculating a future predicted value error as a predicted value of the error of the future predicted values; and a predicted value correction unit for correcting the future predicted values ​​using the future predicted value error, wherein the future predicted value calculation unit is composed of a first machine learning model generated using control parameters and correct values ​​related to the vehicle control device as training data; the future predicted value error calculation unit is composed of a second machine learning model generated using the control parameters, correct values, and future predicted values ​​as training data; and the prediction period or prediction execution interval of the future predicted values ​​is variably set according to the control parameters of the vehicle control device or the future predicted value error.

2. The vehicle data prediction device according to claim 1, wherein the prediction period or the prediction execution interval is variably set to be inversely proportional to the error in the future predicted value.

3. The vehicle data prediction device according to claim 1, further comprising a periodicity determination unit that determines whether or not the time series data of the control parameters is periodic, wherein the prediction execution interval when the time series data of the control parameters is aperiodic is variably set to a value smaller than the prediction execution interval when the time series data of the control parameters is periodic.

4. The vehicle data prediction device according to claim 3, wherein the periodicity determination unit calculates the difference between the current value of the control parameter and a predicted value after a specific time has elapsed, and determines that the data has periodicity if the difference is less than or equal to an arbitrarily set threshold.

5. The vehicle data prediction device according to claim 3, wherein the periodicity determination unit detects the length of the period when it determines that the data is periodic, and the prediction period is determined according to the length of the period.

6. The vehicle data prediction device according to claim 1, wherein the control parameters include received information obtained from a database server via a communication network.

7. The vehicle data prediction device according to claim 6, further comprising a communication speed determination unit that determines whether the communication speed of the communication network is below a threshold, wherein the prediction period or the prediction execution interval is variably set according to the determination result of the communication speed determination unit.

8. A vehicle data prediction method for calculating future predicted values ​​related to a vehicle control device that controls a vehicle, comprising: a future predicted value calculation unit that calculates future predicted values; a future predicted value error calculation unit that calculates a future predicted value error as a predicted value of the error of the future predicted value; a predicted value correction unit that corrects the future predicted value using the future predicted value error; the future predicted value calculation unit is composed of a first machine learning model generated using control parameters and correct values ​​related to the vehicle control device as training data; the future predicted value error calculation unit is composed of a second machine learning model generated using the control parameters, the correct values, and the future predicted values ​​as training data; and the prediction period or prediction execution interval of the future predicted value is variably set according to the control parameters of the vehicle control device or the future predicted value error.

Images