Method for learning state estimation model, method for manufacturing state estimation device, state estimation system, and state estimation device

Abstract

Provided is a state estimation model capable of estimating the state of a battery more quickly than when parameters of a physical model are learned. The PU combines the models of the n battery cells defined by the cell model data to generate a battery model of the battery pack. The PU sets model specification data that specifies the battery model as a value that represents a normal battery pack and a value that represents an abnormal battery pack. The PU generates training data identifying the model by simulation using the battery model. In the recognition model, the current and voltage of the battery pack are input variables, and whether the battery pack is degraded or not is output and displayed. And the PU learns the recognition model by using the training data.

Description

Technical Field

[0001] The present invention relates to a method for learning a state estimation model, a method for manufacturing a state estimation device, a state estimation system, and a state estimation device. Background Art

[0002] For example, Patent Document 1 below describes a device for estimating non-uniformity of the salt concentration of a secondary battery. Specifically, this device uses a physical model of the secondary battery in the estimation of the salt concentration. In particular, this device includes a process of calculating parameters of a specified physical model by the least squares method based on measured values such as the voltage of the secondary battery as an input variable. Prior Art Documents Patent Documents

[0003] Patent Document 1: Japanese Unexamined Patent Application Publication No. 2022-139508 Summary of the Invention Problems to be Solved by the Invention

[0004] As described above, in order to calculate the parameters of a specified physical model based on the detection values of the actual state of a secondary battery, it is necessary to use the detection values over a long period of time. Therefore, it is difficult to meet the requirement of quickly estimating the battery state based on input variables. Means for Solving the Problems

[0005] Examples of the present disclosure are described below. Example 1. A method for learning a state estimation model for estimating the state of a battery is provided. The method for learning the state estimation model has steps of performing a model specification process, a training data generation process, and a learning process. The model specification process is a process of setting parameters of a specified input-output model, that is, specified parameters, to values respectively representing mutually different battery states. The input-output model is a model in which either the current or the voltage of the battery is included in the input and the other is included in the output. The training data generation process is a process of generating training data for the state estimation model. The training data includes data of input and output obtained by simulation using the input-output model and data representing the state of the battery represented by the values of the specified parameters of the input-output model. The learning process is a process of learning the state estimation model using the training data. The input variables of the state estimation model include the input and output of the input-output model, and the output of the state estimation model represents the state of the battery.

[0006] According to the state estimation model as the learned model, compared with the case of learning the parameters of a physical model, the state of the battery can be estimated quickly. In addition, in the above method, the state of the battery is represented by the specified parameters of the input-output model. Moreover, the behaviors of the current and voltage of the battery corresponding to the battery state can be obtained by using simulations of the input-output models that respectively represent the battery state. Therefore, in the above method, even without relying on the charge and discharge of an actual battery with a desired state, the behaviors of the current and voltage of the battery corresponding to the battery state can be obtained. And by using them as training data, a state estimation model can be generated. Therefore, according to the above method, it is possible to easily obtain the training data for generating the state estimation model.

[0007] Example 2. In the learning method of the state estimation model in Example 1 above, the battery is a battery pack in which a plurality of battery cells are connected in series, the input-output model is a model in which a plurality of single-cell models are connected in series, the single-cell model is a model in which any one of the current and voltage of the battery cell is included in the input and the other is included in the output, and the model specification process may include a process of representing the state of the battery pack by determining the values of the specified parameters for each of the single-cell models.

[0008] The above method includes a process of setting the values of the specified parameters for each of the single-cell models constituting the battery pack. Therefore, compared with the case where the values of the specified parameters of all the single-cell models are the same, the degree of freedom in representing the state of the battery pack can be increased.

[0009] Example 3. In the learning method of the state estimation model in Example 1 or Example 2 above, the state estimation model may be an identification model for estimating whether the deterioration of the battery has progressed. Example 4. In the learning method of the state estimation model in any one of Examples 1 to 3 above, in the training data used for the learning of the state estimation model through the above learning process, in addition to the above training data generated by the above training data generation process, it may also include data composed of a set of data obtained when actually charging and discharging the battery and data representing the state of the battery.

[0010] In the above method, by including the training data obtained from the actual battery, compared with the case of only using the data obtained through simulation, it is possible to perform learning more suitable for the state of the actual battery.

[0011] Example 5. In the learning method of the state estimation model according to any one of Examples 1 to 4 above, there are steps of performing individual data calculation processing, training data regeneration processing, and relearning processing. The above individual data calculation processing is a process of calculating the value of the above specified parameter by measuring the voltage and current of one battery. The above training data regeneration processing includes generating data of the input and output of the above input-output model by using simulation of the above input-output model specified by the value of the above specified parameter calculated by the above individual data calculation processing, and data representing the state of the above battery represented by the value of the above specified parameter of the above input-output model. The above relearning processing may be a process of relearning the above state estimation model by using the above training data generated by the above training data regeneration processing.

[0012] The accuracy of the state recognition model learned through the learning process depends on the training data generated through the training data generation process. Therefore, for example, even if the value of a certain specified parameter is set to a value representing a predetermined state, there is a possibility that the estimation results are different when the value of another specified parameter is slightly different. In contrast, in the above method, since one battery is used to set the value of another specified parameter, the possibility of different estimation results can be reduced. That is, in the above method, a state estimation model capable of accurately estimating the state of the above one battery can be obtained through relearning.

[0013] Example 6. A manufacturing method of a state estimation device for estimating the state of a battery is provided. In this manufacturing method, the above state estimation device is configured to perform input variable acquisition processing and state estimation processing. The above input variable acquisition processing is a process of acquiring the value of the input variable of the above state estimation model in the learning method of the state estimation model according to any one of Examples 1 to 5 above. The above state estimation processing is a process of estimating the state of the above battery by inputting the value of the above input variable acquired by the above input variable acquisition processing into the above state estimation model. This manufacturing method has each step in the learning method of the state estimation model according to any one of Examples 1 to 5 above.

[0014] In the above manufacturing method, since the training data for generating the state estimation model can be easily obtained, the man-hours spent in manufacturing the state estimation device can be reduced. Example 7. A state estimation system for estimating the state of a battery is provided. The state estimation system is configured to perform model specification processing, training data generation processing, learning processing, input variable acquisition processing, and state estimation processing. The model specification processing is a process of setting the parameters of a specified input-output model, i.e., specified parameters, to values representing different battery states. The input-output model is a model in which either the current or voltage of the battery is included in the input and the other is included in the output. The training data generation processing is a process of generating training data for the state estimation model. The training data includes data of input and output obtained by simulation using the input-output model, and data representing the state of the battery represented by the values of the specified parameters of the input-output model. The learning processing is a process of learning the state estimation model using the training data. The input variables of the state estimation model include the input and output of the input-output model. The output of the state estimation model represents the state of the battery. The input variable acquisition processing is a process of acquiring the values of the input variables of the state estimation model. The state estimation processing is a process of estimating the state of the battery by inputting the values obtained by the input variable acquisition processing into the state estimation model.

[0015] In the above configuration, the state of the battery is represented by the specified parameters of the input-output model. And the behavior of the current and voltage of the battery corresponding to the state of the battery can be obtained by simulation using each input-output model representing the state of the battery. Therefore, in the above configuration, by preparing an actual battery with a desired state, the behavior of the current and voltage of the battery corresponding to the state of the battery can be obtained without relying on charge and discharge using them. And by using them as training data, a state estimation model can be generated. Therefore, according to the above configuration, it is possible to easily obtain training data for generating a state estimation model.

[0016] In addition, in the above configuration, when estimating the state of the battery using the state estimation model, the state can be estimated more quickly compared to the case of estimating the state by obtaining the parameters of the model of the specified battery that is the object of state estimation.

[0017] Example 8. The state estimation system of Example 7 described above includes a state estimation device and a learning device. The state estimation device is configured to perform the input variable acquisition processing and the state estimation processing described above. The learning device is configured to perform the model specification processing, the training data generation processing, and the learning processing described above.

[0018] According to the above configuration, by separately arranging the state estimation device that performs the estimation process using the state estimation model and the learning device, the place where the state estimation process is executed can be freely selected. Example 9. In the state estimation system of Example 8 above, the state estimation device is configured to perform individual data calculation processing, specified parameter transmission processing, and state estimation model reception processing. The learning device is configured to perform specified parameter reception processing, training data regeneration processing, relearning processing, and state estimation model transmission processing. The individual data calculation processing is a process of calculating the value of the specified parameter by measuring the voltage and current of an actual single battery. The specified parameter transmission processing is a process of transmitting the specified parameter calculated by the individual data calculation processing to the learning device. The specified parameter reception processing is a process of receiving the specified parameter transmitted by the specified parameter transmission processing. The training data regeneration processing is a process of generating the input and output data of the input-output model and the data representing the state of the battery represented by the value of the specified parameter of the input-output model by simulating the input-output model specified by the value of the specified parameter received by the specified parameter reception processing. The relearning processing is a process of relearning the state estimation model using the training data generated by the training data regeneration processing. The state estimation model transmission processing is a process of transmitting the state estimation model relearned by the relearning processing to the state estimation device. The state estimation model reception processing is a process of receiving the state estimation model transmitted by the state estimation model transmission processing.

[0019] The accuracy of the state estimation model learned through the learning process depends on the training data generated by the training data generation process. Therefore, for example, even if the value of a certain specified parameter is set to a value representing a predetermined state, there is a possibility that the estimation results are different when the value of another specified parameter is slightly different. In contrast, in the above configuration, the state estimation model is relearned using the specified parameter of an actual single battery. Thus, a state estimation model focused on the state of an actual single battery to be estimated can be obtained.

[0020] Example 10. Provide the state estimation device in the state estimation system of Example 9 above. Brief Description of the Drawings

[0021] Figure 1 is a block diagram showing the drive system of a vehicle in one embodiment. Figure 2 is a diagram showing Figure 1 the single cell model used by the learning device shown. Figure 3 is a diagram showingFigure 1 Flowchart of the process of the processing executed by the learning device shown Figure 4 It shows Figure 1 Flowchart of the process of the processing executed by the battery ECU shown Figure 5 It shows Figure 1 Flowchart of the process of the processing executed by the state estimation system shown Detailed implementation mode

[0022] An embodiment will be described below with reference to the drawings "Premise configuration" Figure 1 It shows the configuration of the state estimation system of the battery pack

[0023] The vehicle 2 is equipped with an electric generator 4. The rotating shaft of the electric generator 4 is mechanically connected to the drive wheels of the vehicle. The output voltage of the power conversion circuit 6 is applied to the terminals of the electric generator 4. The power conversion circuit 6 is configured to convert the terminal voltage of the battery pack 10, which is a DC voltage source, into an AC voltage and output it

[0024] The battery pack 10 is a series connection of battery cells 12(1), 12(2),... 12(n). The terminal voltage of the battery pack 10 can be, for example, several tens of volts to several hundreds of volts. The numbers in parentheses in the battery cells 12(1), 12(2),... 12(n) are individual identification numbers. Hereinafter, when summarizing the battery cells 12(1), 12(2),... 12(n), they are described as battery cells 12. As an example, the battery cell 12 is a lithium-ion secondary battery. As an example, the battery cell 12 has a structure in which a positive electrode plate, a separator, and a negative electrode plate are laminated and wound into a flat shape

[0025] The monitoring unit 20 is a circuit that monitors the states of the battery cells 12(1), 12(2),... 12(n) of the battery pack 10 The battery ECU 30 includes a PU 32, a storage device 34, and a communicator 36. The PU 32 is a software processing device such as a CPU and a GPU. The storage device 34 can be a storage medium such as an electrically rewritable non-volatile memory and a disk medium. The battery ECU 30 is a device that monitors the state of the battery pack 10

[0026] The battery ECU 30 refers to the charge and discharge current I of the battery pack 10 detected by the current sensor 22. In addition, the battery ECU 30 refers to the terminal voltages of the individual battery cells 12(1), 12(2), … 12(n), i.e., the individual cell voltages Vc(1), Vc(2), … Vc(n), detected by the monitoring unit 20. In addition, the battery ECU 30 refers to the temperature T of the battery pack 10 detected by the monitoring unit 20.

[0027] The battery ECU 30 can communicate with the upper ECU 40 via the in-vehicle network. The upper ECU 40 is a device that manages the power of the vehicle's drive system. The upper ECU 40 controls the driving force of the motor generator 4 by outputting an instruction to the MG ECU 44 via the in-vehicle network. The charge and discharge electric energy of the battery pack 10 is also controlled by controlling the driving force. Therefore, the upper ECU 40 controls the charge and discharge electric energy of the battery pack 10 by outputting an instruction to the MG ECU 44 via the in-vehicle network.

[0028] The MG ECU 44 operates the power conversion circuit 6 to control the control quantity of the motor generator 4 to be controlled, such as torque. The battery ECU 30 sets the maximum value Imax of the charging current at which lithium does not precipitate in the negative electrode of the battery pack 10 and outputs it to the upper ECU 40. The upper ECU 40 calculates the maximum value of the charging power of the battery pack 10, i.e., the maximum charging power Winmax, based on the maximum value Imax and outputs it to the MG ECU 44. The battery ECU 30 sets the maximum value Imax according to whether the battery pack 10 is deteriorated. The battery ECU 30 determines whether the battery pack 10 is deteriorated based on the identification model specified by the identification model data 34a stored in the storage device 34.

[0029] The battery ECU 30 can communicate with a learning device 60 outside the vehicle 2 via the wireless network 50. The learning device 60 can communicate with multiple vehicles 2. The learning device 60 includes a PU 62, a storage device 64, and a communication device 66. The PU 62 is a software processing device such as a CPU and a GPU. The storage device 64 can be a storage medium such as an electrically rewritable non-volatile memory and a disk medium. In the storage device 64, monomer model data 64a that defines a model of the individual battery cell 12, i.e., a monomer model, is stored. The learning device 60 uses the monomer model to generate training data for generating the above-mentioned identification model data 34a.

[0030] A detailed description thereof will be given below. "Monomer model" Figure 2 The monomer model is shown.

[0031] The negative electrode solid-phase diffusion model M10 is a process of calculating the lithium-ion concentration on the surface of the negative electrode of the battery cell 12 using a mathematical model based on the charge-discharge current I and temperature T as input variables.

[0032] The negative electrode potential mapping M12 is a process of calculating the negative electrode potential of the battery cell 12 based on the lithium-ion concentration on the surface of the negative electrode as an input variable. Specifically, the negative electrode potential mapping M12 includes a process of calculating the SOC of the negative electrode active material based on the lithium-ion concentration on the surface of the negative electrode as an input variable through the following formula.

[0033] SOC of negative electrode active material = {lithium-ion concentration on the surface of the negative electrode} / cnmax In the above formula, the maximum lithium-ion concentration cnmax of the negative electrode, which is a specified parameter of a specified model, is used.

[0034] The negative electrode potential mapping M12 includes a process of calculating the negative electrode potential based on the SOC of the negative electrode active material as an input variable. Specifically, as an example, this process is a process of performing a mapping operation on the negative electrode potential using the PU62 in a state where the mapping data is stored in the storage device 64. The mapping data is data with the SOC of the negative electrode active material as the input variable and the negative electrode potential as the output variable.

[0035] Here, the mapping data is a data group of discrete values of the input variable and the values of the output variable corresponding to the values of the input variable respectively. In addition, when any one of the value of the input variable and the value of the input variable of the mapping data is the same, the mapping operation may be a process where the value of the output variable of the corresponding mapping data is the operation result. In addition, when none of the value of the input variable and the value of the input variable of the mapping data is the same, the mapping operation may be a process where the value obtained by interpolation of the values of the multiple output variables included in the mapping data is the operation result. Alternatively, when none of the value of the input variable and the value of the input variable of the mapping data is the same, the mapping operation may also be a process where the value of the output variable of the mapping data corresponding to the value of the input variable closest to the value of the input variable included in the mapping data is the operation result.

[0036] The negative electrode reaction overvoltage model M14 is a process of calculating the overvoltage generated by the negative electrode of the battery cell 12, that is, the negative electrode reaction overvoltage, using a mathematical model based on the charge-discharge current I and temperature T as input variables.

[0037] The negative electrode potential calculation process M16 is a process of substituting the value obtained by adding the negative electrode potential calculated by the negative electrode potential mapping M12 and the negative electrode reaction overvoltage into the estimated value of the negative electrode potential. The positive electrode solid-phase diffusion model M18 is a process of calculating the lithium-ion concentration on the surface of the positive electrode of the battery cell 12 using a mathematical model based on the charge and discharge current I and the temperature T as input variables.

[0038] The positive electrode potential mapping M20 includes a process of calculating the positive electrode potential using a mathematical model based on the lithium-ion concentration on the surface of the positive electrode as an input variable. Specifically, the positive electrode potential mapping M20 includes a process of calculating the state of charge (SOC) of the positive electrode active material based on the lithium-ion concentration on the surface of the positive electrode as an input variable by the following formula.

[0039] State of charge of positive electrode active material = 1 - {lithium-ion concentration on the surface of the positive electrode} / cpmax In the above formula, the maximum lithium-ion concentration cpmax of the positive electrode, which is a specified parameter of a specified model, is used.

[0040] The positive electrode potential mapping M20 includes a process of calculating the positive electrode potential based on the state of charge (SOC) of the positive electrode active material as an input variable. Specifically, as an example, this process is a process of performing a mapping operation on the positive electrode potential using the PU62 in a state where the mapping data is stored in the storage device 64. The mapping data is data with the charge and discharge current I and the temperature T as input variables and the positive electrode potential as an output variable.

[0041] The positive electrode reaction overvoltage model M22 is a process of calculating the overvoltage that can be generated by the positive electrode of the battery cell 12, i.e., the positive electrode reaction overvoltage, using a mathematical model based on the charge and discharge current I and the temperature T as input variables.

[0042] The positive electrode potential calculation process M24 is a process of substituting the value obtained by adding the positive electrode potential calculated by the positive electrode potential mapping M20 and the positive electrode reaction overvoltage into the estimated value of the positive electrode potential. The liquid-phase diffusion model M26 is a process of calculating the salt concentration in the electrolyte in the thickness direction of the battery cell 12 using a mathematical model based on the charge and discharge current I and the temperature T as input variables. As an example, the mathematical model here is the liquid-phase diffusion model. Specifically, the liquid-phase diffusion model M26 includes a process of performing a mapping operation on the negative electrode boundary salt concentration and the positive electrode boundary salt concentration based on the state of charge (SOC) and the temperature T as input variables. The negative electrode boundary salt concentration is the salt concentration at the boundary between the negative electrode and the negative electrode current collector. In addition, the positive electrode boundary salt concentration is the salt concentration at the boundary between the positive electrode and the positive electrode current collector. The liquid-phase diffusion model M26 includes a process of calculating the salt concentration in the electrolyte in the thickness direction of the battery cell 12 using the liquid-phase diffusion model with the negative electrode boundary salt concentration and the positive electrode boundary salt concentration as boundary conditions based on the charge and discharge current I and the temperature T as input variables. The thickness direction of the battery cell 12 refers to the direction parallel to the shortest side when the electrode body formed by laminating and winding the positive electrode plate, the separator, and the negative electrode plate is regarded as a cuboid.

[0043] The liquid-phase potential difference model M28 is a process of calculating the potential difference in the electrolyte in the thickness direction of the battery cell 12, i.e., the liquid-phase potential difference, based on the salt concentration in the above-mentioned electrolyte as an input variable. The component resistance model M30 is a process of calculating the voltage reduction amount caused by the resistance of the components included in the battery cell 12 using a mathematical model based on the charge / discharge current I and the temperature T as input variables. As an example, the component resistance model M30 includes a process of calculating the component resistance R using the proportionality coefficient α and the intercept b by "R = α·T + b". The component resistance model M30 includes a process of substituting the value obtained by multiplying the component resistance R by the charge / discharge current I into the voltage reduction amount caused by the resistance of the components.

[0044] The liquid resistance model M32 is a process of calculating the voltage reduction amount in the electrolyte using a mathematical model based on the voltage reduction amount caused by the resistance of the components as an input variable. The film resistance model M34 is a process of calculating the voltage reduction amount caused by the film formed on the negative electrode surface using a mathematical model based on the voltage reduction amount in the electrolyte as an input variable.

[0045] The voltage calculation process M40 is a process of substituting the value obtained by subtracting the sum of the estimated value of the negative electrode potential and the voltage reduction amount based on the film from the sum of the estimated value of the positive electrode potential and the liquid-phase potential difference into the voltage estimated value. It should be noted that in the monomer model data 64a, the parameters of the monomer model, i.e., the specified parameters, are set in a manner corresponding to the behavior of the central characteristic product of the battery cell 12. That is, for example, the maximum lithium ion concentration cnmax in the negative electrode, the maximum lithium ion concentration cpmax in the positive electrode, the proportionality coefficient α, and the intercept b are set corresponding to the central characteristic product.

[0046] "Generation process of the identification model" Figure 3 shows the process of the generation process of the identification model. Figure 3 The shown process is realized by the PU62 repeatedly executing the program stored in the storage device 64 at a predetermined cycle, for example. It should be noted that hereinafter, the step numbers of each process are represented by numbers with the prefix "S".

[0047] Figure 3 In the shown series of processes, the PU62 first sets the specified parameters (S10) of the battery model that defines the normal battery pack 10. The battery model of the battery pack 10 is a model of n Figure 2A model in which the single-cell models of the battery cells 12 shown are connected in series. The battery model of the normal battery pack 10 can be specified, for example, only by the parameters specified by the single-cell model data 64a. That is, the specified parameters of each of the n battery cells 12 can be made values for simulating the behavior of the central characteristic product of all the battery cells 12. In addition, in the battery model of the normal battery pack 10, for at least one of the battery cells 12(1) to 12(n), it may include specified parameters whose change amount compared to the central characteristic product is below a predetermined value. Here, the predetermined value is set to a value considered to have no deterioration. It should be noted that there are preferably multiple types of battery models for the normal battery pack 10. In the process of S10, the PU62 sets the deviation of the state of charge SOC of each of the n battery cells 12 to a value within the allowable range.

[0048] Next, the PU62 sets the specified parameters of the battery model of the battery pack 10 specified as abnormal (S12). Among the specified parameters of the battery model of the battery pack 10 specified as abnormal, for at least one of the battery cells 12(1) to 12(n), it includes specified parameters whose change amount compared to the central characteristic product is above a specified value. The specified value is set to a value considered to have deteriorated. There are preferably multiple types of battery models for the abnormal battery pack 10.

[0049] In the process of S12, as an example, the PU62 represents the abnormal battery pack 10 by the maximum negative electrode lithium ion concentration cnmax, the maximum positive electrode lithium ion concentration cpmax, the proportionality coefficient α, the intercept b, and the state of charge SOC of each battery cell 12.

[0050] That is, the PU62 represents the abnormal battery pack 10 by setting the deviation of the full charge capacity determined by the maximum negative electrode lithium ion concentration cnmax and the maximum positive electrode lithium ion concentration cpmax of the n battery cells 12 to exceed the allowable range. In addition, the PU62 represents the abnormal battery pack 10 by setting the full charge capacity determined by the maximum negative electrode lithium ion concentration cnmax and the maximum positive electrode lithium ion concentration cpmax to below a predetermined value. Here, the number of battery cells 12 with a full charge capacity below the predetermined value is 1 or more and n or less. The PU62 can set the maximum negative electrode lithium ion concentration cnmax and the maximum positive electrode lithium ion concentration cpmax in a manner that covers all patterns where the number of battery cells 12 with a full charge capacity below the predetermined value is 1 to n.

[0051] In addition, PU62 identifies an abnormal battery pack 10 by setting it such that the deviation of the charging rate SOC of each of the n battery cells 12 exceeds the allowable range. In addition, PU62 identifies an abnormal battery pack 10 by means of battery cells 12 in which the component resistance R determined by the proportionality coefficient α and the intercept b is above the threshold. Here, the number of battery cells 12 in which the component resistance R is above the threshold is 1 or more and n or less. PU62 can set the proportionality coefficient α and the intercept b in such a way as to cover all cases where the number of battery cells 12 in which the component resistance R is above the threshold is 1 to n. In addition, PU62 identifies an abnormal battery pack 10 by setting it such that the deviation of the component resistance R determined by the proportionality coefficient α and the intercept b exceeds the allowable range.

[0052] PU62 simulates the terminal voltage of the battery pack 10 (S14) by inputting the current and temperature according to a predetermined pattern into either a normal battery model or an abnormal battery model related to the battery pack 10.

[0053] PU62 stores, in the storage device 64, in association with a label variable indicating whether the data is generated by a normal model, the sets of the current, temperature, terminal voltage of the battery cell 12, and terminal voltage of the battery pack 10 obtained through the process of S14 (S16). One of the label variable and the set of data associated with the label variable constitutes one training data.

[0054] It should be noted that the input variables of the recognition model are not limited to only the values at one time point of the current of the battery cell 12, the terminal voltage of the battery cell 12, and the terminal voltage of the battery pack 10. For example, the input variables of the recognition model may include time-series data composed of the values of the current of the battery cell 12, the terminal voltage of the battery cell 12, and the terminal voltage of the battery pack 10 at mutually different time points. In this case, the training data includes the label variable, the time-series data associated with the label variable, and the temperature.

[0055] In addition, PU62 acquires, from the outside, training data obtained from the actual battery pack 10 and stores it in the storage device 64 (S18). The training data may include the current, terminal voltage, and temperature when the central characteristic product of the battery pack 10 is actually charged and discharged, and the terminal voltage of the battery cells 12 constituting the battery pack 10. In addition, the training data may include the current, terminal voltage, and temperature when the deliberately deteriorated battery pack 10 is actually charged and discharged, and the terminal voltage of the battery cells 12 constituting the battery pack 10.

[0056] Next, the PU 62 learns an identification model (S20) using the training data stored in the storage device 64 through the processes of S16 and S18. Here, in fact, only a part of the data stored in the storage device 64 through the processes of S16 and S18 can be used as training data, and the remaining data can be used as verification data and detection data. Additionally, this step S20 may include a process of changing the data used as training data among the data stored in the storage device 64 through the processes of S16 and S18, for example, by performing cross-validation.

[0057] The identification model can be a non-parametric model such as a support vector machine. In this case, the identification model data 34a that defines the identification model can be support vectors extracted from the training data. Additionally, the identification model can also be a parametric model such as a neural network. In this case, the identification model data 34a can be the weight coefficients and bias amounts of the neural network.

[0058] It should be noted that the PU 62 learns the identification model in such a way that the output of the identification model is made consistent with the label variable by using a known cross-entropy or the like for the evaluation function. That is, for example, when the identification model is a neural network, the PU 62 updates the weight coefficients and bias coefficients based on the evaluation function by means of the error backpropagation method or the like.

[0059] When the learning is completed, the PU 62 stores the identification model data 34a in the storage device 64 (S22). The identification model data 34a thus stored in the storage device 64 is stored in the storage device 34 mounted on each vehicle.

[0060] It should be noted that when the PU 62 completes the process of S22, the series of processes shown Figure 3 is temporarily terminated. "Limit processing regarding power limit" Figure 4 shows the process of the processing related to the limitation of the power limit performed by the battery ECU 30. Figure 4 The series of processes shown is implemented by the PU 32 repeatedly executing the program stored in the storage device 34 at a predetermined cycle, for example.

[0061] In Figure 4In the series of processes shown, PU32 first obtains the charge-discharge current I, temperature T, terminal voltage V of the battery pack 10, and individual cell voltages Vc(1) to Vc(n) (S30), which are input variables of the recognition model. Here, the terminal voltage V of the battery pack 10 can be a value obtained by adding the individual cell voltages Vc(1) to Vc(n) by PU32. It should be noted that the charge-discharge current I, temperature T, terminal voltage V of the battery pack 10, and individual cell voltages Vc(1) to Vc(n) obtained through the process of S30 are not limited to the sampled values at one time point. The variable of at least one of the charge-discharge current I, temperature T, terminal voltage V of the battery pack 10, and individual cell voltages Vc(1) to Vc(n) obtained through the process of S30 can also be the sampled values at multiple time points. That is, for example, when the input variables of the recognition model include the time-series data of the charge-discharge current I, terminal voltage V, and individual cell voltages Vc(1) to Vc(n), the process of S30 is to obtain this time-series data.

[0062] Next, PU32 determines whether there is an abnormality in the battery pack 10 by substituting the values of the variables obtained through the process of S30 into the recognition model specified by the recognition model data 34a (S32). Next, PU32 determines whether it is the power limit restriction mode (S34). The power limit restriction mode is a mode that limits the output of the battery pack 10 to the lower side.

[0063] When PU32 determines that it is not the power limit restriction mode (S34: No), it is determined whether the output of the recognition model shows a deterioration abnormality of the battery pack 10 (S36). When the output of the recognition model shows a deterioration abnormality of the battery pack 10, PU32 switches the control mode of the battery pack 10 to the power limit restriction mode (S40). And PU32 substitutes the limit value IL into the maximum value Imax (S42).

[0064] On the other hand, when the output of the recognition model does not show a deterioration abnormality of the battery pack 10 (S36: No), PU32 substitutes the normal value IH into the maximum value Imax (S38). The normal value IH is set to a value larger than the limit value IL.

[0065] When PU32 determines that it is in the power limit restriction mode (S34: Yes), the output of the recognition model determines whether the battery pack 10 is normal (S44). When PU32 determines that it is normal (S44: Yes), the count C is incremented (S46). After that, PU32 determines whether the count C is equal to or greater than the threshold Cth (S48). This process determines whether the power limit restriction mode is released because the battery pack 10 is considered normal. When PU32 determines that the count C is equal to or greater than the threshold Cth (S48: Yes), the process proceeds to S38. In this case, PU32 initializes the count C. On the other hand, when PU32 determines that the count C is less than the threshold Cth (S48: No), and when the determination in the process of S44 is negative, the process proceeds to S42.

[0066] It should be noted that when PU32 completes the processes of S38 and S42, the series of processes shown Figure 4 is temporarily terminated. "Model update process" Figure 5 shows the process related to the update of the recognition model data 34a and the single-cell model data 34b stored in the storage device 34. Figure 5 The process shown on the left side is implemented by PU32 repeatedly executing the program stored in the storage device 34 at a predetermined cycle, for example. Figure 5 The process shown on the right side is implemented by PU62 repeatedly executing the program stored in the storage device 64 at a predetermined cycle, for example. It should be noted that the single-cell model data 34b stored in the storage device 34 when the vehicle 2 is shipped is the same as the single-cell model data 64a.

[0067] In Figure 5 In the series of processes shown, PU32 first obtains the charge and discharge current I, temperature T, terminal voltage V of the battery pack 10, and single-cell voltages Vc(1) to Vc(n) (S50). Next, PU32 uses the single-cell model specified by the single-cell model data 34b to calculate the estimated value Ve of the terminal voltage of the battery pack 10 based on the charge and discharge current I and temperature T as input variables (S52). In addition, PU32 calculates the estimated values Vce(1) to Vce(n) of the single-cell voltages Vc(1) to Vc(n). Next, PU32 updates the specified parameters and the state of charge SOC of each battery cell 12 in such a way as to reduce the difference between the estimated value and the detected value of the voltage. That is, PU32 updates the specified parameters and the state of charge SOC of each battery cell 12 in such a way as to reduce the differences between the single-cell voltages Vc(1) to Vc(n) and the estimated values Vce(1) to Vce(n), and between the terminal voltage V and the estimated value Ve.

[0068] In the process of S54, as an example, the PU62 updates the maximum negative electrode lithium ion concentration cnmax, the maximum positive electrode lithium ion concentration cpmax, the proportionality coefficient α, the intercept b, and the charging rate SOC of each battery cell 12. This process can be, for example, a process in which the PU62 searches for a solution that minimizes the difference between the estimated value and the detected value of the voltage when various settings are made for the values of the above-mentioned specified parameters and the charging rate SOC.

[0069] It should be noted that the process of S54 is actually executed using multiple sampled values at time points when the charge and discharge current I, the temperature T, the terminal voltage V of the battery pack 10, and the cell voltages Vc(1) to Vc(n) of the battery pack 10 are different from the corresponding estimated values.

[0070] After that, the PU32 determines whether the update of the model has been completed (S56). For example, when the absolute value of the difference between the estimated value Ve and the terminal voltage V is below the threshold, the PU32 can determine that the update of the model has been completed. When the PU62 determines that the update of the model has been completed (S56: Yes), the updated specified parameters are sent to the learning device 60 together with the identifier of the vehicle by operating the communication device 36 (S58).

[0071] On the contrary, the PU62 of the learning device 60 receives the sent identifier of the vehicle and the specified parameters (S70). Then, after updating the cell model data 64a with the received specified parameters, the PU62 performs a process equivalent to the process of S14 using the updated cell model (S72). Incidentally, the PU62 performs the processes of S10 and S12 using both the updated cell model data 64a and the data of the values of the specified parameters that deviate from the cell model data 64a. That is, for example, when the value of the specified parameter of the cell model data 64a is within the normal range although it deviates from the value of the central characteristic product, the PU62 generates a normal battery model using this value. For example, even when the proportionality coefficient α and the intercept b shown in the cell model data 64a deviate from the values of the central characteristic product, when they are within the normal range, the model generated by the PU62 using the proportionality coefficient α and the intercept b shown in the cell model data 64a is included in the normal battery model. In addition, the battery model in which, for example, the proportionality coefficient α and the intercept b shown in the cell model data 64a are used and the charging rate SOC is deviated is included in the abnormal battery model.

[0072] After that, the PU62 stores the data obtained by simulation as training data in the storage device 64 in the same way as the process of S16 (S74). After that, the PU72 re-learns the recognition model using the data stored by the process of S74 (S76). After that, the PU62 sends the data of the re-learned specified recognition model to the battery ECU 30 by operating the communication device 66 (S78).

[0073] Note that, in the case where PU62 of the learning device 60 finishes the process of S78, a series of processes shown on the right side of Figure 5 is temporarily terminated. On the other hand, PU32 of the battery ECU 30 receives data of a specified recognition model (S60). After that, PU32 updates the recognition model data 34a stored in the storage device 34 (S62). That is, the recognition model data 34a stored in the storage device 34 is replaced with the received data. Note that, in the case where PU32 finishes the process of S62 and in the case where a negative determination is made in the process of S56, a series of processes shown on the left side of Figure 5 is temporarily terminated.

[0074] "Functions and effects of the present embodiment" PU62 of the learning device 60 generates a model of the battery pack 10 by connecting in series n single models defined by the single model data 64a. In particular, PU62 represents both a normal battery pack 10 and an abnormal battery pack 10 by setting the specified parameters of the specified model and the SOC of each battery cell. PU62 calculates estimated values of the terminal voltages of the n battery cells 12 and the terminal voltage of the battery pack 10 based on the charge and discharge current I and the temperature T as input variables using the model of the battery pack 10. PU62 associates a label variable indicating whether the model of the battery pack 10 used in the calculation of the recognition estimated values is a normal model or an abnormal model with the charge and discharge current I and the temperature T, and the estimated values of the terminal voltages of the corresponding battery cells 12 and the terminal voltage of the battery pack 10. PU62 generates a recognition model by using the associated data as training data.

[0075] PU32 of the battery ECU 30 determines whether the battery pack 10 is abnormal by inputting the single cell voltages Vc(1) to Vc(n), the terminal voltage V of the battery pack 10, the charge and discharge current I, and the temperature T into the generated recognition model. After that, in the case where an abnormality is determined, PU32 restricts the output of the battery pack 10 to suppress the precipitation of lithium ions in the battery cells 12.

[0076] Here, the time required for determining whether there is an abnormality using the recognition model is shorter than the time required for obtaining the detection values of the sensors for calculating the specified parameter values of the single model. Therefore, by using the recognition model, the state of the battery pack 10 can be estimated earlier than the process of updating the specified parameters of the single model.

[0077] According to the present embodiment described above, the following functions and effects are further obtained. (1) The PU62 of the learning device 60 independently sets the values of the specified parameters for each monomer model of the specified battery monomers 12(1) to 12(n). Thereby, compared with the case where the values of the specified parameters of each monomer model of the specified battery monomers 12(1) to 12(n) are made common, the degree of freedom of expressing the state of the battery pack 10 can be improved.

[0078] (2) In the learning of the recognition model, the PU62 of the learning device 60 uses not only the simulation results of the battery model but also the measured values of the behavior of the actual battery pack 10 as training data. Thereby, the accuracy of the output of the recognition model can be improved.

[0079] (3) The PU62 of the learning device 60 uses the specified parameters of the monomer model calculated by the PU32 of the battery ECU 30 to generate training data for re-learning. After that, the PU62 learns the recognition model inherent to the battery pack 10 mounted on one vehicle. Thereby, compared with the recognition model generated by the Figure 3 processing, a model capable of accurately determining whether there is an abnormality in the battery pack 10 mounted on one vehicle can be realized.

[0080] That is, for example, in a battery model that represents the abnormal deviation of the state of charge SOC of the battery monomers 12(1) to 12(n), the value of the specified parameter may be different from the value of the specified parameter that accurately represents the battery pack 10 to be estimated. In this case, the accuracy of determining whether there is an abnormal deviation in the state of charge SOC of the battery monomers 12(1) to 12(n) by the recognition model generated by the Figure 3 processing may be reduced. In contrast, in the present embodiment, the PU62 of the learning device 60 re-learns the recognition model using the value of the specified parameter that accurately represents the battery pack 10 to be estimated. Therefore, it is possible to accurately determine whether there is an abnormality in the battery pack 10 to be estimated.

[0081] (4) The learning device 60 performs re-learning of the recognition model. Thereby, compared with the case where the battery ECU 30 performs re-learning of the recognition model, the arithmetic load of the battery ECU 30 can be reduced.

[0082] <Corresponding relationship> The correspondence between the matters in the above-described embodiments and the matters described in the column of "Means for Solving the Problem" above is as follows. Hereinafter, the correspondence is shown for each number of the solutions described in the column of "Means for Solving the Problem". [1, 2, 3] The model specification process corresponds to the processes of S10 and S12. The training data generation process corresponds to the process of S14. The learning process corresponds to the process of S20. The input-output model corresponds to the model in which single models are connected in series. [4] The training data corresponds to the data obtained by the process of S18. [5] The individual data calculation process corresponds to the processes of S52 to S56. The training data regeneration process corresponds to the process of S72. The relearning process corresponds to the process of S76. [6] The input variable acquisition process corresponds to the process of S30. The state estimation process corresponds to the process of S32. [7, 8] The model specification process corresponds to the processes of S10 and S12. The training data generation process corresponds to the process of S14. The learning process corresponds to the process of S20. The input-output model corresponds to the model in which single models are connected in series. The input variable acquisition process corresponds to the process of S30. The state estimation process corresponds to the process of S32. [9, 10] The individual data calculation process corresponds to the processes of S52 to S56. The specified parameter transmission process corresponds to the process of S58. The state estimation model reception process corresponds to the process of S60. The specified parameter reception process corresponds to the process of S70. The training data regeneration process corresponds to the process of S72. The relearning process corresponds to the process of S76. The state estimation model transmission process corresponds to the process of S78.

[0083] <Other Embodiments> It should be noted that this embodiment can be implemented with the following modifications. This embodiment and the following modification examples can be implemented in combination with each other within a technically non-contradictory range.

[0084] "Regarding the Training Data" · The temperature included in the training data is not limited to the temperature T of the battery pack 10. For example, it can be a temperature independently set for each of the battery cells 12(1) to 12(n). When the training data includes data in which the battery cells 12(1) to 12(n) have mutually different temperatures, the conditions that can actually occur in the battery pack 10 can be represented with higher accuracy.

[0085] "Regarding the State Estimation Model" · The state estimation model does not have to be an identification model. For example, the state estimation model can be a regression model that outputs the full charge capacity of battery cells 12(1) to 12(n). In other words, the state to be estimated by the state estimation model can be the full charge capacity. Additionally, for example, the state estimation model can be a regression model that outputs a variable value representing the degree of deviation of the state of charge (SOC) of battery cells 12(1) to 12(n). Here, the variable value representing the degree of deviation can be, for example, the maximum value of the absolute value of the difference between the SOC of battery cell 12 and the average value of the SOCs of battery cells 12(1) to 12(n), etc.

[0086] "Regarding the battery" · The battery that is the object of state estimation based on the state estimation model does not have to be a battery pack.

[0087] "Regarding the input-output model" · The temperature T input to the cell model is not limited to the temperature T of the battery pack 10. For example, if the temperatures of battery cells 12(1) to 12(n) can be obtained separately, it can be the temperature inherent to the battery cell 12 that is the object.

[0088] · As the cell model in which the input includes current, it is not limited to Figure 2 the model illustrated in · As the model that includes either the current or voltage of the battery in the input and includes the other in the output, it is not limited to the model in which the input includes current. For example, it can also be a model in which the input includes voltage.

[0089] "Regarding the state estimation system" · Figure 5 The processes of S50 to S56, S62, S72 to S76 of

[0090] · Figure 5 The processes of S52 to S56, S72 to S78 of

[0091] · The state estimation system does not have to include the battery ECU 30 mounted on the vehicle and the learning device 60 that can communicate with multiple vehicles. The state estimation system can be configured as, for example, a control device provided in a power generation device equipped with a battery.

[0092] "Regarding the state estimation device" · The state estimation device is not limited to a device that executes software processing. For example, it may include a dedicated hardware circuit such as an ASIC that executes at least a part of the processing performed in the above-described embodiments. That is, the state estimation processing may include a processing circuit having any one of the following configurations (a) to (c). (a) A processing circuit including a processing device that executes all of the above processing according to a program, and a program storage device such as a storage device that stores the program. (b) A processing circuit including a processing device and a program storage device that execute a part of the above processing according to a program, and a dedicated hardware circuit that executes the remaining processing. (c) A processing circuit including a dedicated hardware circuit that executes all of the above processing. Here, there may be a plurality of software execution devices including a processing device and a program storage device. In addition, there may be a plurality of dedicated hardware circuits.

[0093] "Regarding the learning device" · The learning device is not limited to a device that executes software processing. For example, it may include a dedicated hardware circuit such as an ASIC that executes at least a part of the processing performed in the above-described embodiments. That is, the learning processing may include a processing circuit having any one of the following configurations (a) to (c). (a) A processing circuit including a processing device that executes all of the above processing according to a program, and a program storage device such as a storage device that stores the program. (b) A processing circuit including a processing device and a program storage device that execute a part of the above processing according to a program, and a dedicated hardware circuit that executes the remaining processing. (c) A processing circuit including a dedicated hardware circuit that executes all of the above processing. Here, there may be a plurality of software execution devices including a processing device and a program storage device. In addition, there may be a plurality of dedicated hardware circuits.

Claims

1. A learning method for a state estimation model, which is a learning method for a state estimation model for estimating a state of a battery, wherein: The method includes steps of executing model specification processing, training data generation processing and learning processing. The model defining process is a process of setting parameters defining the input-output model, that is, defining parameters, to values ​​representing different battery states. The input-output model is a model in which either one of the current and the voltage of the battery is included in the input, and either the other is included in the output. The training data generation process is a process of generating training data for the state estimation model. The training data includes input and output data obtained by simulation using the input-output model, and data indicating the state of the battery represented by the value of the predetermined parameter of the input-output model. The learning process is a process of learning the state estimation model through the training data. The input variables of the state estimation model include the input and output of the input-output model. The output of the state estimation model represents the state of the battery.

2. The learning method of the state estimation model according to claim 1, wherein: The battery is a battery pack in which a plurality of battery cells are connected in series. The input-output model is a model in which multiple monomer models are connected in series. The cell model is a model in which any one of the current and the voltage of the battery cell is included in the input, and any other one of the current and the voltage of the battery cell is included in the output. The model specification process includes a process of expressing the state of the battery pack by determining a value of the specified parameter for each of the cell models.

3. The learning method of the state estimation model according to claim 1, wherein: The state estimation model is a recognition model that estimates whether degradation of the battery has progressed.

4. The learning method of the state estimation model according to claim 1, wherein: The training data used for learning the state estimation model through the learning process includes, in addition to the training data generated through the training data generation process, data consisting of a group of data obtained when an actual battery is charged and discharged and data representing the state of the battery.

5. The learning method of the state estimation model according to claim 1, wherein: The method includes the steps of executing individual data calculation processing, training data regeneration processing, and relearning processing. The individual data calculation process is a process of calculating the value of the predetermined parameter by measuring the voltage and current of one battery. The training data regeneration process includes a process of generating data of input and output of the input-output model and data indicating the state of the battery represented by the value of the prescribed parameter of the input-output model by using a simulation of the input-output model prescribed by the value of the prescribed parameter calculated by the individual data calculation process, The relearning process is a process of relearning the state estimation model using the training data generated by the training data regeneration process.

6. A method for manufacturing a state estimation device, which is a method for manufacturing a state estimation device for estimating a state of a battery, wherein: The state estimation device is configured to execute input variable acquisition processing and state estimation processing. The input variable acquisition process is a process of acquiring the value of the input variable of the state estimation model in the state estimation model learning method according to claim 1. The state estimation process is a process of estimating the state of the battery by inputting the value of the input variable acquired by the input variable acquisition process into the state estimation model. This method includes the steps of the state estimation model learning method according to claim 1.

7. A state estimation system, which is a state estimation system for estimating a state of a battery, wherein: The system is configured to execute model definition processing, training data generation processing, learning processing, input variable acquisition processing, and state estimation processing. The model defining process is a process of setting parameters defining the input-output model, that is, defining parameters, to values ​​representing different battery states. The input-output model is a model in which either one of the current and the voltage of the battery is included in the input, and either the other one is included in the output. The training data generation process is a process of generating training data for a state estimation model. The training data includes input and output data obtained by simulation using the input-output model, and data indicating the state of the battery represented by the value of the predetermined parameter of the input-output model. The learning process is a process of learning the state estimation model through the training data. The input variables of the state estimation model include the input and output of the input-output model. The output of the state estimation model represents the state of the battery, The input variable acquisition process is a process of acquiring the value of the input variable of the state estimation model. The state estimation process is a process of estimating the state of the battery by inputting the value acquired by the input variable acquisition process into the state estimation model.

8. The state estimation system according to claim 7, wherein: The system has a state estimation device and a learning device. The state estimation device is configured to execute the input variable acquisition process and the state estimation process. The learning device is configured to execute the model definition process, the training data generation process, and the learning process.

9. The state estimation system according to claim 8, wherein: The state estimation device is configured to execute individual data calculation processing, prescribed parameter transmission processing, and state estimation model reception processing. The learning device is configured to perform a predetermined parameter receiving process, a training data regeneration process, a relearning process, and a state estimation model transmission process. The individual data calculation process is a process of calculating the value of the predetermined parameter by measuring the voltage and current of one battery. The predetermined parameter transmission process is a process of transmitting the predetermined parameter calculated by the individual data calculation process to the learning device. The predetermined parameter receiving process is a process of receiving the predetermined parameter sent by the predetermined parameter sending process. The training data regeneration process is a process of generating data of input and output of the input-output model and data indicating the state of the battery expressed by the value of the prescribed parameter of the input-output model by using a simulation of the input-output model prescribed by the value of the prescribed parameter received by the prescribed parameter receiving process, The relearning process is a process of relearning the state estimation model using the training data generated by the training data regeneration process. The state estimation model transmission process is a process of transmitting the state estimation model relearned by the relearning process to the state estimation device. The state estimation model receiving process is a process of receiving the state estimation model transmitted by the state estimation model transmitting process. 10 . A state estimation device, which is the state estimation device in the state estimation system according to claim 9 .

Images