State estimation method and state estimation device

Abstract

According to the present invention, when a degradation state has been estimated from usage information about the usage history of a vehicle using a machine learning model that has been trained to estimate the degradation state from the usage information and the accuracy of the degradation state as estimated using the machine learning model is below a prescribed accuracy, the degradation state is estimated using a physical model that estimates the degradation state from information about the state of a battery.

Description

State Estimation Method and State Estimation Device

[0001] The present invention relates to a state estimation method and a state estimation device for estimating the degradation state of a battery.

[0002] There is known a battery state detection device that generates a regression model learned so that information regarding the degradation state of a battery is output when information having a high correlation with the degradation state of the battery is input, and estimates the degradation state of the battery from information regarding the state of the battery using the generated regression model (Patent Document 1).

[0003] International Publication No. 2023 / 149011

[0004] In the above prior art, when the battery information used for learning the regression model is biased within a specific range, there is a problem that the degradation state of the battery cannot be accurately estimated in a range outside the specific range where the information is biased.

[0005] The problem to be solved by the present invention is to provide a state estimation method and a state estimation device that can suppress a decrease in the estimation accuracy of the degradation state of a battery when estimating the degradation state of the battery using a machine learning model.

[0006] According to the present invention, when estimating the degradation state of a battery using a machine learning model, if the accuracy of the degradation state estimated using the machine learning model is less than a predetermined accuracy, the above problem is solved by estimating the degradation state using a physical model that estimates the degradation state from information regarding the state of the battery.

[0007] According to the present invention, when estimating the degradation state of a battery using a machine learning model, it is possible to suppress a decrease in the estimation accuracy of the degradation state of the battery.

[0008] This is a block diagram showing one embodiment of the state estimation system according to the present invention. This is a block diagram showing an example of the first estimation unit in Figure 1. This is a block diagram showing an example of the second estimation unit in Figure 1. This is a diagram showing an example of the range in which the battery degradation state can be estimated by the first estimation unit in Figure 1. This is a diagram showing an example of training information for training the machine learning model used in the first calculation unit in Figure 2A. This is a flowchart showing an example of a processing procedure for estimating the battery degradation state in the state estimation system of Figure 1. This is a flowchart showing an example of a processing procedure for training a machine learning model in the state estimation system of Figure 1. This is a flowchart showing another example of a processing procedure for training a machine learning model in the state estimation system of Figure 1. This is a flowchart showing yet another example of a processing procedure for training a machine learning model in the state estimation system of Figure 1. This is a flowchart showing an example of a processing procedure for determining the start of battery degradation state estimation in the state estimation system of Figure 1.

[0009] Embodiments of the present invention will be described below with reference to the drawings.

[0010] [Configuration of the State Estimation System] Figure 1 is a block diagram showing one embodiment of the state estimation system according to the present invention. The state estimation system of this embodiment is a group of devices for estimating the degradation state of a vehicle's battery (hereinafter also simply referred to as the degradation state). The vehicle is a vehicle powered by a motor (or an internal combustion engine and a motor), and includes electric vehicles (EVs), hybrid vehicles (HVs), fuel cell vehicles (FCVs), etc. The vehicle's battery (hereinafter also simply referred to as the battery) is a secondary battery mounted on the vehicle, for example, a lithium-ion secondary battery including an all-solid-state battery. The battery supplies power to the motor, which is the power source.

[0011] As shown in Figure 1, the state estimation system 1 of this embodiment comprises a vehicle 10 and a state estimation device 20 installed outside the vehicle 10. The vehicle 10 and the state estimation device 20 are connected to each other via a telecommunications network such as the Internet or a LAN (local area network). The method of communication between the vehicle 10 and the state estimation device 20 is not particularly limited. Note that the vehicle 10 constituting the state estimation system 1 does not necessarily have to be a single vehicle; multiple vehicles 10 may be included in the state estimation system 1, and multiple vehicles 10 may be connected to the state estimation device 20 in a communication manner.

[0012] Vehicle 10 is equipped with a vehicle control device 11, a detection device 12, and a display device 13. These devices are connected to each other via known means such as wired or wireless LAN, and exchange information with each other.

[0013] The vehicle control device 11 controls and coordinates the devices that make up the vehicle 10, acquires information regarding the vehicle 10's driving state and battery state from the detection device 12, and transmits it to the state estimation device 20. The vehicle control device 11 is, for example, a computer and includes a CPU (Central Processing Unit) which is a processor, a ROM (Read Only Memory) where programs are stored, and a RAM (Random Access Memory) which functions as an accessible storage device. The CPU of the vehicle control device 11 is an operating circuit that executes programs stored in the ROM of the vehicle control device 11 and realizes the functions of the vehicle control device 11.

[0014] The detection device 12 detects the driving state of the vehicle 10 (hereinafter also simply referred to as the driving state). Examples of the detection device 12 for detecting the driving state include a vehicle speed sensor, an acceleration sensor, a yaw rate sensor, and a steering angle sensor. The detection device 12 also detects the state of the battery. Examples of the detection device 12 for detecting the state of the battery include a thermometer, an ammeter, and a voltmeter. Known sensors can be used without particular limitation, and their arrangement and number can be appropriately set within a range that allows for proper detection of the driving state (or battery state) of the vehicle 10. The vehicle control device 11 acquires the detection results of the detection device 12 at predetermined time intervals (for example, every 0.1 to 1 millisecond).

[0015] The display device 13 provides information to the user of the vehicle 10. The display device 13 is, for example, a projector such as a liquid crystal display or head-up display provided on the instrument panel, and may include an input device for the user to input instructions to the vehicle control device 11 and a speaker as an output device. The user is not particularly limited to anyone who uses the vehicle 10, and includes not only the occupants of the vehicle 10 but also users of services provided through the vehicle 10.

[0016] The state estimation device 20 estimates the degradation state. The degradation state is also called the health level or capacity retention rate, and is also called SOH (State of Health). The capacity retention rate is the ratio of the current battery capacity to the initial battery capacity. The state estimation device 20 is, for example, a server and, like the vehicle control device 11, is equipped with a CPU, ROM, and RAM. The CPU of the state estimation device 20 is an operating circuit that executes the program stored in the ROM of the state estimation device 20 and realizes the functions of the state estimation device 20.

[0017] [Function of the State Estimation Device] The ROM of the state estimation device 20 stores a program for estimating the degradation state, and the CPU of the state estimation device 20 executes this program to perform the degradation state estimation. Figure 1 shows, for convenience, the control unit 21, the first estimation unit 22, and the second estimation unit 23 as functional blocks for performing the degradation state estimation.

[0018] The control unit 21 outputs the detection results from the detection device 12 received from the vehicle control device 11 to the database 24. The detection results output to the database 24 are organized according to certain rules and stored as usage information (hereinafter also referred to as probe information or simply usage information) regarding the usage history of the vehicle 10. If multiple vehicles 10 are connected to the state estimation device 20 in a communicative manner, usage information for each vehicle 10 is stored in the database 24.

[0019] The usage history of vehicle 10 includes the driving history of vehicle 10 (hereinafter also simply referred to as driving history) and the usage history of vehicle 10's battery. In other words, the usage information includes information related to the driving history and information related to the battery usage history. Information related to the driving history includes information on the vehicle 10's driving time, driving distance, location, driving speed, acceleration, etc. Information related to the battery usage history includes information on the battery's charge level (hereinafter also simply referred to as charge level) and its changes, the battery's usage state, the battery's charging time, the battery's charging start time, and the charging end time. The battery's usage state includes the charged state and the discharged state.

[0020] When the control unit 21 receives a request from the vehicle control device 11 for information regarding the deterioration state, it outputs an execution instruction to the first estimation unit 22 and the second estimation unit 23 to estimate the deterioration state. When the first estimation unit 22 receives an execution instruction from the control unit 21, it uses a machine learning model (hereinafter also simply referred to as the machine learning model) that has been trained to estimate the deterioration state from usage information to estimate the deterioration state from usage information. Similarly, when the second estimation unit 23 receives an execution instruction from the control unit 21, it uses a physical model (hereinafter also simply referred to as the physical model) that estimates the deterioration state from information regarding the battery state to estimate the deterioration state from usage information.

[0021] Figure 2A is a block diagram showing an example of the first estimation unit 22. As shown in Figure 2A, the first estimation unit 22 comprises an input unit 221 and a first calculation unit 222. When an execution instruction is input from the control unit 21, the input unit 221 acquires usage information from the database 24 and outputs the acquired usage information as input information to the first calculation unit 222. The first calculation unit 222 has a trained machine learning model and uses the machine learning model to estimate the degradation state from the usage information. The estimated degradation state information is output to the control unit 21 as an estimation result. The first calculation unit 222 may also feed back the estimation result to the input unit 221 and, when the battery usage state changes, use the estimation result from before the battery usage state changed to estimate the degradation state.

[0022] The machine learning model is, for example, a neural network comprising an input layer, at least one hidden layer, and an output layer. When usage information is input to the input layer, information on the degradation state corresponding to the input usage information is output from the output layer. The neural network is pre-trained using training data, for example, in which usage information and information on the degradation state corresponding to the usage information are associated. Information input to the input layer includes the charge level at the start of estimation, the battery's usage period, the vehicle's mileage, the amount of power consumed since the last charge was completed, the charge level at the last charge was completed, the amount of power supplied to an external load of the vehicle 10, and the ambient temperature while the vehicle 10 is parked. Information output from the output layer includes the degradation state, as well as the amount of SEI (Solid Electrolyte Interphase) loss, the battery cell temperature, the depth of cracks in the active material of the lithium-ion secondary battery, the electrolyte concentration, and the specific surface area of ​​the electrodes. SEI is a film that is mainly formed at the interface between the negative electrode and the electrolyte in a lithium-ion secondary battery, mainly during charging.

[0023] For example, if the input layer of a neural network receives information such as the current charge level, battery usage period, vehicle 10's mileage, charge level at the end of the last charge, and battery ambient temperature, the hidden layer converts this information into a one-dimensional vector array, and extracts the first feature from this vector array. If the second feature has already been extracted from the training data using a similar process, the hidden layer compares the first and second features. Since the training data (i.e., the second feature) is associated with the degradation state, the hidden layer outputs the degradation state corresponding to the second feature that is identical or similar to the first feature to the output layer. The output layer outputs the information about the degradation state input from the hidden layer as an estimation result.

[0024] Whether the first feature is similar to the second feature is determined by the similarity between the first and second features (e.g., Euclidean distance). For example, if the similarity between the first and second features is greater than or equal to a predetermined similarity, the first and second features are determined to be similar in the intermediate layer. If the similarity between the first and second features is less than the predetermined similarity, the first and second features are determined to be dissimilar in the intermediate layer. The predetermined similarity can be set to an appropriate value within a range that allows for accurate estimation of the degradation state, for example, between 0.8 and 1.

[0025] On the other hand, Figure 2B is a block diagram showing an example of the second estimation unit 23. As shown in Figure 2B, the second estimation unit 23 comprises an input information generation unit 231 and a second calculation unit 232. When an execution instruction is input from the control unit 21, the input information generation unit 231 acquires usage information from the database 24, generates input information from the acquired usage information, and outputs it to the second calculation unit 232. The second calculation unit 232 has a physical model and uses the physical model to estimate the degradation state from the usage information. The estimated degradation state information is output to the control unit 21 as an estimation result. The second calculation unit 232 may also feed back the estimation result to the input information generation unit 231 and reflect the current estimation result in the next input information.

[0026] The physical model is a combination of theoretical equations representing physical phenomena, such as a prediction formula for SEI generation, the Butler-Volmer equation for determining the battery current, and the diffusion equation and Stefan-Maxwell equations for predicting the electrolyte concentration in the electrolyte solution. It estimates the degradation state from the time change of the battery state. Alternatively, the physical model may be a model that performs map calculations using a map that shows the relationship between the battery state and the degradation state. The map used in the map calculation is created in advance from information such as the results of experiments using batteries and the relationship between the battery usage history and the degradation state.

[0027] The input information generation unit 231 generates time-series information such as the time change of battery power and the time change of the battery's ambient temperature from the acquired usage information and outputs it to the second calculation unit 232. The input information generation unit 231 also outputs information regarding the battery state to the second calculation unit 232, such as the battery usage state, the battery charging method (i.e., whether it is rapid charging or normal charging), and the battery characteristics.

[0028] The second calculation unit 232 substitutes the information input from the input information generation unit 231 into the physical model to determine the current value, voltage value, battery temperature, battery charge level, battery DC resistance, degradation state, etc., and outputs them as estimated results. In addition, the second calculation unit 232 outputs information such as the amount of SEI loss, the depth of cracks in the active material of the lithium-ion secondary battery, the electrolyte concentration, and the specific surface area of ​​the electrodes, similar to the first calculation unit 222.

[0029] The first estimation unit 22 does not need to generate new input information for the first calculation unit 222 and uses a machine learning model that requires less computation than a physical model, so it can estimate the degradation state with less computation than the second estimation unit 23. However, for batteries that have deteriorated, the training data that can be used to train the machine learning model may be limited, and in this case, the machine learning model may not be sufficiently trained in the range where the battery has deteriorated. Thus, in the range where training is insufficient, the machine learning model may not be able to accurately estimate the degradation state.

[0030] Therefore, when the first estimation unit 22 estimates the deterioration state, the control unit 21 calculates the accuracy of the deterioration state estimated by the first estimation unit 22 (hereinafter also referred to as estimation accuracy) and determines whether the estimation accuracy is less than a predetermined accuracy. If it is determined that the estimation accuracy is less than the predetermined accuracy, the control unit 21 has the second estimation unit 23 estimate the deterioration state. Specifically, if it is determined that the estimation accuracy is less than the predetermined accuracy, the control unit 21 has the second estimation unit 23 estimate the deterioration state instead of the first estimation unit 22. On the other hand, if it is determined that the estimation accuracy is equal to or greater than the predetermined accuracy, the control unit 21 has the first estimation unit 22 continue estimating the deterioration state.

[0031] The estimation accuracy is, for example, the accuracy corresponding to the shortest distance between the degradation state estimated by the first estimation unit 22 and the training data used to train the machine learning model. The distance between the two data points may be the Euclidean distance, the Manhattan distance, or the max distance. Alternatively, the estimation accuracy may correspond to the value of the variance of the Gaussian process regression model, which was determined in advance when training the machine learning model. The predetermined accuracy can be set to an appropriate value within a range that allows for proper estimation of the degradation state; for example, if it is the variance of the Gaussian process regression model, it may be between 0.5 and 1.

[0032] Figure 3 shows an example of an approximation curve for the degradation state. In the graph shown in Figure 3, the horizontal axis represents the battery usage period (in hours), and the vertical axis represents the degradation state (in percent). The approximation curve B1 shown in Figure 3 is an approximation curve obtained using a Gaussian process regression model from the training data A1 to A8. The variance (uncertainty) of the approximation curve B1 corresponds to the range B2 shown in Figure 3. The variance (uncertainty) shown by range B2 corresponds to the accuracy of the estimated degradation state; the larger the variance at a given time (i.e., the wider the range B2), the lower the accuracy of the estimated degradation state.

[0033] For example, in the example shown in Figure 3, if the battery usage period is t1 hours, the degradation state is estimated to be X1, and if the battery usage period is t2 hours, the degradation state is estimated to be X2. In this case, the variance in the estimated value X2 is smaller than the variance in the estimated value X1, so the estimated value X2 is a more accurate estimate than the estimated value X1.

[0034] If either the first estimation unit 22 or the second estimation unit 23 is estimating the deterioration state, the control unit 21 may switch the main estimator of the deterioration state from the first estimation unit 22 to the second estimation unit 23 if the accuracy of the deterioration state estimated using the machine learning model (i.e., estimation accuracy) is less than a predetermined accuracy, stop the estimation of the deterioration state by the first estimation unit 22, and start the estimation of the deterioration state by the second estimation unit 23. Furthermore, if the first estimation unit 22 and the second estimation unit 23 are estimating the deterioration state simultaneously (in parallel), the control unit 21 may transmit the deterioration state estimated by the second estimation unit 23 (i.e., the deterioration state estimated using the physical model) to the vehicle control device 11 when the estimation accuracy is less than a predetermined accuracy, and transmit the deterioration state estimated by the first estimation unit 22 (i.e., the deterioration state estimated using the machine learning model) to the vehicle control device 11 when the estimation accuracy is equal to or greater than the predetermined accuracy.

[0035] The control unit 21 may predict the remaining battery life based on the estimated degradation state. The control unit 21 may also transmit information regarding at least one of the degradation state and the remaining battery life to the vehicle control device 11. When the vehicle control device 11 receives information regarding the degradation state from the control unit 21, it may display the received information on the display device 13 to present it to the user. The vehicle control device 11 may also set or change the battery state (e.g., output) based on the information received from the control unit 21.

[0036] The first estimation unit 22 may train a machine learning model using information (hereinafter also referred to as second information) regarding the deterioration state estimated using a physical model (i.e., the deterioration state estimated by the second estimation unit 23). If multiple vehicles 10 are connected to the state estimation device 20 in a communicable manner, the first estimation unit 22 may train a machine learning model using usage information of other vehicles 10 (or usage information of vehicle 10 and usage information of other vehicles 10).

[0037] The second estimation unit 23 may store the second information in the database 24 if it has estimated the deterioration state. The first estimation unit 22 may also determine whether a predetermined amount or more of the second information has been stored in the database 24. If it is determined that a predetermined amount or more of the second information has been stored, the first estimation unit 22 uses the second information to train the machine learning model. On the other hand, if it is determined that less than a predetermined amount of the second information has been stored, the first estimation unit 22 does not train the machine learning model using the second information. The predetermined amount can be set to an appropriate value within a range that allows the machine learning model to be properly trained. For example, the first estimation unit 22 uses the second information to train the machine learning model when the ratio of the second information to the information regarding the deterioration state estimated using the machine learning model (i.e., the deterioration state estimated by the first estimation unit 22) (hereinafter also referred to as the first information) becomes 1 / 3 or more.

[0038] The first estimation unit 22 may determine whether the degradation state estimated using the physical model falls within the range of degradation states that can be estimated using the machine learning model (hereinafter also referred to as the estimable range). For example, in the example shown in Figure 3, ranges C1 and C2 are the estimable range, and within ranges C1 and C2, the variance is relatively small, and the estimated values ​​of the degradation state within ranges C1 and C2 have an accuracy of a predetermined precision or higher. If it is determined that the degradation state estimated using the physical model does not fall within the estimable range, the first estimation unit 22 trains the machine learning model using the second information. On the other hand, if it is determined that the degradation state estimated using the physical model falls within the estimable range, the first estimation unit 22 does not train the machine learning model using the second information.

[0039] When the first estimation unit 22 trains a machine learning model using the second information, it may extract training information from the second information by clustering and train the machine learning model using the training information. Figure 4 shows an example of training information extracted by clustering. In the graph shown in Figure 4, the horizontal axis represents the average driving speed of the vehicle 10 (in km / h), and the vertical axis represents the energy consumption (in kW·h). As shown in Figure 4, when training a machine learning model using training data D1 to D17, the first estimation unit 22 aggregates the training data D1 to D17 into four groups E1 to E4 and trains the machine learning model using training data that represents each group (hereinafter also referred to as representative data). Specifically, training data D1 to D5 are grouped into group E1 and training data D5 is used as representative data; training data D6 to D8 are grouped into group E2 and training data D8 is used as representative data; training data D9 to D10 are grouped into group E3 and training data D9 is used as representative data; and training data D11 to D17 are grouped into group E4 and training data D13 and D15 are used as representative data.

[0040] The first estimation unit 22 may estimate future usage information from the usage information stored in the database 24. For example, the first estimation unit 22 calculates the probability that the vehicle 10's driving state (or battery state) will change for each user and for each time period, and generates a virtual behavior pattern for the user based on the calculated probability. Then, it generates usage information corresponding to the virtual behavior pattern. The first estimation unit 22 may also train a machine learning model using the usage information estimated to be acquired in the future.

[0041] Further, the first estimation unit 22 may estimate future acquired usage information from the usage information before the ratio of the electrolyte of the battery to the active material contained in the battery becomes less than a predetermined ratio. For example, in the case of a lithium-ion secondary battery, the main factors of battery deterioration are the formation of SEI at the interface between the negative electrode and the electrolyte and the deactivation of the active material accompanying the consumption of the electrolyte. Among these, the deactivation of the active material appears only when the consumption of the electrolyte has progressed to a certain extent (that is, when the ratio of the electrolyte to the active material becomes less than a predetermined ratio). After the deactivation of the active material appears, the deterioration of the battery progresses faster than before the deactivation of the active material appears. By estimating the usage information to be acquired after the deactivation of the active material appears before the deactivation of the active material appears and using the estimated usage information to train the machine learning model, it is possible to suppress a decrease in estimation accuracy after the deactivation of the active material appears. The predetermined ratio is, for example, 0.8 to 1.2.

[0042] Note that at least a part of the processes executed by the above-described state estimation device 20 (control unit 21, first estimation unit 22, and second estimation unit 23) may be executed by the vehicle control device 11.

[0043] [Processing in the State Estimation System] Referring to FIGS. 5 to 9, the procedure when the state estimation device 20 processes information will be described. The processes described below are executed by a processor (CPU) included in the state estimation device 20 at predetermined time intervals (for example, every 0.1 to 1 millisecond).

[0044] FIG. 5 is a flowchart showing an example of a processing procedure executed when estimating the deterioration state in the state estimation system 1.

[0045] First, in step S1, the state estimation device 20 estimates the degradation state using a machine learning model. Subsequently, in step S2, it is determined whether the accuracy (estimation accuracy) of the degradation state estimated using the machine learning model is less than a predetermined accuracy. If it is determined that the estimation accuracy is less than the predetermined accuracy, the process proceeds to step S3, where the state estimation device 20 estimates the degradation state using a physical model. Subsequently, in step S4, information regarding the degradation state estimated using the physical model is output to the vehicle 10. On the other hand, if it is determined in step S2 that the estimation accuracy is equal to or greater than the predetermined accuracy, the process proceeds to step S4, and the state estimation device 20 outputs information regarding the degradation state estimated using the machine learning model to the vehicle 10.

[0046] Next, FIG. 6 is a flowchart showing an example of a processing procedure executed when training a machine learning model in the state estimation system 1.

[0047] First, in step S11, the state estimation device 20 determines whether information (second information) regarding the degradation state estimated using the physical model has been accumulated by a predetermined amount or more. If it is determined that the second information has been accumulated by a predetermined amount or more, the process proceeds to step S12, and the state estimation device 20 trains the machine learning model using the second information. On the other hand, if it is determined that the second information has been accumulated by less than the predetermined amount, the state estimation device 20 ends the process.

[0048] Next, FIG. 7 is a flowchart showing another example of a processing procedure executed when training a machine learning model in the state estimation system 1.

[0049] First, in step S21, the state estimation device 20 determines whether the degradation state estimated using the physical model falls within the range of degradation states that can be estimated using the machine learning model (estimateable range). If it is determined that the degradation state estimated using the physical model is within the estimateable range, the state estimation device 20 terminates the process. On the other hand, if it is determined that the degradation state estimated using the physical model is not within the estimateable range, the process proceeds to step S22, where the state estimation device 20 extracts learning information from the second information, and in the subsequent step S23, trains the machine learning model using the extracted learning information.

[0050] Next, Figure 8 is a flowchart showing yet another example of the processing steps performed when training a machine learning model in the state estimation system 1.

[0051] First, in step S31, the state estimation device 20 obtains past usage information from the database 24, and in the following step S32, it estimates future usage information from the obtained past usage information. Then, in step S33, the state estimation device 20 trains a machine learning model using the estimated usage information.

[0052] Next, Figure 9 is a flowchart showing an example of the processing procedure performed when determining whether or not to start estimating the deterioration state in the state estimation system 1.

[0053] First, in step S41, the state estimation device 20 calculates the ratio of the battery electrolyte to the active material contained in the battery, and in the following step S42, it determines whether the difference between the calculated ratio and a predetermined ratio is less than a predetermined threshold. If it is determined that the difference between the calculated ratio and the predetermined ratio is less than the predetermined threshold, the process proceeds to step S43, where the state estimation device 20 estimates usage information to be acquired in the future (for example, usage information acquired after the deactivation of the active material due to a decrease in electrolyte). On the other hand, if it is determined that the difference between the calculated ratio and the predetermined ratio is greater than or equal to the predetermined threshold, the state estimation device 20 terminates the process. The predetermined threshold can be set to an appropriate value within the range in which the state estimation device 20 can start estimating the deterioration state before the ratio of electrolyte to active material falls below the predetermined ratio, for example, between 0.1 and 0.5.

[0054] [Embodiment of the Invention] According to this embodiment, a state estimation method for estimating the degradation state of a vehicle 10 is provided, which involves estimating the degradation state from usage information relating to the usage history of the vehicle 10 using a machine learning model trained to estimate the degradation state from the usage information, and if the accuracy of the degradation state estimated using the machine learning model is less than a predetermined accuracy, estimating the degradation state using a physical model that estimates the degradation state from information relating to the state of the battery, and a state estimation device 20 that executes the state estimation method is provided. This makes it possible to suppress a decrease in the accuracy of estimating the degradation state while suppressing a decrease in the processing speed for estimating the degradation state when estimating the degradation state using a machine learning model.

[0055] In the state estimation method and state estimation device 20 of this embodiment, the machine learning model is trained using the information on the degradation state estimated using the physical model. This expands the range in which the machine learning model can estimate the degradation state.

[0056] In the state estimation method and state estimation device 20 of this embodiment, information regarding the deterioration state estimated using the physical model is accumulated, and it is determined whether or not a predetermined amount of information regarding the deterioration state estimated using the physical model has been accumulated. If it is determined that a predetermined amount or more of information regarding the deterioration state estimated using the physical model has been accumulated, the machine learning model is trained using the information regarding the deterioration state estimated using the physical model. This expands the range in which the machine learning model can estimate the deterioration state.

[0057] In the state estimation method and state estimation device 20 of this embodiment, it is determined whether the degradation state estimated using the physical model falls within the range of degradation states that can be estimated using the machine learning model. If it is determined that the degradation state estimated using the physical model does not fall within the range, the machine learning model is trained using the information regarding the degradation state estimated using the physical model. This expands the range in which the machine learning model can estimate degradation states.

[0058] In the state estimation method and state estimation apparatus 20 of this embodiment, when training the machine learning model using the information on the degradation state estimated using the physical model, learning information is extracted from the information on the degradation state estimated using the physical model by clustering, and the machine learning model is trained using the learning information. This reduces the load required to train the machine learning model.

[0059] In the state estimation method and state estimation device 20 of this embodiment, future usage information is estimated from the usage information, and the machine learning model is trained using the usage information estimated to be acquired in the future. This makes it possible to generate a machine learning model that corresponds to the future state of the battery.

[0060] In the state estimation method and state estimation device 20 of this embodiment, the usage information to be acquired in the future is estimated from the usage information before the ratio of the electrolyte of the battery to the active material contained in the battery falls below a predetermined ratio. This makes it possible to generate a machine learning model that corresponds to the degradation modes that appear after the degradation of the battery has progressed.

[0061] 1...State estimation system 10...Vehicle 11...Vehicle control device 12...Detection device 13...Display device 20...State estimation device 21...Control unit 22...First estimation unit 221...Input unit 222...First calculation unit 23...Second estimation unit 231...Input information generation unit 232...Second calculation unit 24...Database A1, A2, A3, A4, A5, A6, A7, A8...Training data B1...Approximate curve B2...Range C1, C2...Range D1, D2, D3, D4, D5, D6, D7, D8, D9, D10, D11, D12, D13, D14, D15, D16, D17...Training data E1, E2, E3, E4...Group t1, t2...Time X1, X2...Estimated value

Claims

1. A state estimation method for estimating the degradation state of a vehicle battery, performed by a state estimation device, wherein the state estimation device estimates the degradation state from usage information using a machine learning model trained to estimate the degradation state from usage information relating to the vehicle's usage history, and if the accuracy of the degradation state estimated using the machine learning model is less than a predetermined accuracy, the state estimation method estimates the degradation state using a physical model that estimates the degradation state from information relating to the battery's state.

2. The state estimation method according to claim 1, wherein the state estimation device trains the machine learning model using the information regarding the deterioration state estimated using the physical model.

3. The state estimation method according to claim 1, wherein the state estimation device stores information on the deterioration state estimated using the physical model, determines whether or not a predetermined amount of information on the deterioration state estimated using the physical model has been stored, and if it is determined that a predetermined amount of information on the deterioration state estimated using the physical model has been stored, it trains the machine learning model using the information on the deterioration state estimated using the physical model.

4. The state estimation method according to claim 1 or 3, wherein the state estimation device determines whether the degradation state estimated using the physical model falls within the range of degradation states that can be estimated using the machine learning model, and if it determines that the degradation state estimated using the physical model does not fall within the range, it causes the machine learning model to learn using the information regarding the degradation state estimated using the physical model.

5. The state estimation method according to any one of claims 2 to 4, wherein, when the state estimation device trains the machine learning model using the information on the deterioration state estimated using the physical model, it extracts learning information from the information on the deterioration state estimated using the physical model by clustering, and trains the machine learning model using the learning information.

6. The state estimation method according to any one of claims 1 to 5, wherein the state estimation device estimates the usage information to be acquired in the future from the usage information, and trains the machine learning model using the usage information estimated to be acquired in the future.

7. The state estimation method according to claim 6, wherein the state estimation device estimates future usage information from the usage information before the ratio of the electrolyte of the battery to the active material contained in the battery falls below a predetermined ratio.

8. A state estimation device for estimating the degradation state of a vehicle battery, comprising: a first estimation unit that estimates the degradation state from usage information using a machine learning model trained to estimate the degradation state from usage information relating to the vehicle's usage history; and a second estimation unit that estimates the degradation state using a physical model that estimates the degradation state from information relating to the battery state when the accuracy of the degradation state estimated using the machine learning model is less than a predetermined accuracy.

Images