Learning Method, State Estimation Method and State Estimation Device of State Estimation Model

Abstract

The present invention provides a learning method for a state estimation model, a state estimation method, and a state estimation device. The state (charge rate and degree of deterioration) of a secondary battery having various electrical characteristics is estimated with high precision during operation. The method includes the following steps: measuring state variables including terminal current and terminal voltage of the secondary battery at regular time intervals; preprocessing the state variables to calculate state estimation input data; and causing the state estimation model to learn the relationship between the state estimation input data and the charge rate and / or degree of deterioration of the secondary battery through machine learning. In the step of calculating the state estimation input data, a differential gradient, which is a change rate of a voltage difference of the terminal voltage with respect to a current difference of the terminal current, is calculated, and state estimation input data including time series data of the differential gradient is generated.

Description

Technical Field

[0001] The present invention relates to a learning method, a state estimation method, and a state estimation device for a state estimation model that estimates the state of a secondary battery during operation. Background Art

[0002] A secondary battery, which is a storage battery that can be repeatedly used by charging, is widely used in mobile bodies such as electric vehicles and electric bicycles, or in buildings. When using these secondary batteries, it is important to appropriately grasp the state of the secondary battery in order to determine the appropriate charging timing, replacement timing, etc. Here, the state of the secondary battery refers to the SOC (state of charge, remaining capacity, State Of Charge) and / or SOH (state of degradation, State Of Health).

[0003] Conventionally, the following technique has been known: two neural networks are used to automatically and accurately determine the degradation state and SOC of a secondary battery during operation in real time (Patent Document 1). In this technique, a first neural network and a second neural network are used. The first neural network is learned in such a way as to estimate the degradation state D (distinction between "normal", "caution", and "degraded") of the secondary battery based on a combination of time series of measured values of the operating parameters (voltage V, current I, internal impedance Z, temperature T) of the secondary battery. The second neural network is learned in such a way as to estimate the SOC of the secondary battery based on the measured values of the operating parameters and the estimated degradation state D.

[0004] On the other hand, the electrical characteristics of secondary batteries, such as SOC-OCV (open circuit voltage) characteristics, internal impedance characteristics, and / or the dependence of these characteristics on SOH, may vary depending on the manufacturer and / or model of the secondary battery. Therefore, the relationships between the voltage V, current I, and internal impedance Z of the secondary battery and the SOC and / or SOH (hereinafter, SOC, etc.) mostly vary greatly depending on the manufacturer and model of the secondary battery.

[0005] Therefore, in the above-described conventional technique in which the measured values of the voltage V, current I, internal impedance Z, and temperature of the secondary battery are directly input to the neural network during learning, one manufacturer and model of the secondary battery to be the estimation target (target secondary battery) are determined, and learning data for the neural network is collected using secondary batteries of the same manufacturer and model.

[0006] Moreover, although the above-mentioned existing neural network learned in this way can accurately estimate the SOC and the like of secondary batteries of the same manufacturer and the same model as the secondary battery to be estimated, it is difficult to accurately estimate the SOC and the like of secondary batteries of various manufacturers and / or models with different electrical characteristics as the estimation objects.

[0007] However, for example, when estimating the state of a secondary battery during the operation of a vehicle, it is convenient if the SOC and the like of secondary batteries of various manufacturers and models can be accurately estimated using a common one or a set of estimation models (e.g., neural network).

[0008] Prior Art Documents

[0009] Patent Documents

[0010] Patent Document 1: Japanese Patent Application Laid-Open No. 2003-249271 Summary of the Invention

[0011] Problems to be Solved by the Invention

[0012] The present invention has been made in view of the above circumstances, and an object thereof is to accurately estimate the state of charge (SOC) and / or state of health (SOH) of secondary batteries having various electrical characteristics with different manufacturers or models during the operation of the secondary batteries.

[0013] Means for Solving the Problems

[0014] One aspect of the present invention is a learning method for a state estimation model of a secondary battery, which is a machine learning-based learning method for a state estimation model that estimates the state of charge and / or state of health of a working secondary battery connected to a load or a charger. The learning method for the state estimation model of the secondary battery includes the following steps: measuring state variables including a terminal current and a terminal voltage of the working secondary battery at a predetermined time interval; preprocessing the state variables to calculate state estimation input data; and causing the state estimation model to learn the relationship between the state estimation input data and the state of charge and / or state of health of the working secondary battery through machine learning. In the step of performing the calculation, based on the time series data of the terminal current and the time series data of the terminal voltage, a difference in the terminal current, i.e., a current difference, and a difference in the terminal voltage, i.e., a voltage difference, are calculated, and based on the time series data of the current difference and the time series data of the voltage difference, a change rate of the voltage difference with respect to the current difference, i.e., a differential gradient, is calculated for a period from a past time traced back by a first predetermined time to the current time, and the state estimation input data including the time series data of the differential gradient is generated.

[0015] According to another aspect of the present invention, the state estimation input data further includes time series data of the open-circuit voltage of the secondary battery in operation, time series data of a first gradient change rate, and time series data of a second gradient change rate. The first gradient change rate is calculated based on time series data of a cumulative current value, which is the sum of the terminal current values continuously obtained during a period from the past traced back by a second specified time to the present, and time series data of a differential gradient change amount obtained by subtracting the differential gradient traced back by the second specified time from the current differential gradient. The first gradient change rate is the change rate of the differential gradient change amount with respect to the cumulative current value during a period from the past traced back by a third specified time to the present. The second gradient change rate is calculated based on time series data of an open-circuit voltage change amount obtained by subtracting the open-circuit voltage traced back by the second specified time from the current open-circuit voltage, and time series data of the differential gradient change amount. The second gradient change rate is the change rate of the differential gradient change amount with respect to the open-circuit voltage change amount during a period from the past traced back by a third specified time to the present.

[0016] According to another aspect of the present invention, the first gradient change rate and the second gradient change rate are calculated using the least squares method.

[0017] According to another aspect of the present invention, in the step of performing the calculation, time series data of the terminal current, time series data of the terminal voltage, and time series data of the differential gradient are used as voltage estimation input data to estimate the open-circuit voltage of the secondary battery in operation, and the state estimation input data is calculated using the estimated open-circuit voltage.

[0018] According to another aspect of the present invention, in the step of performing the calculation, the state estimation input data is calculated using the open-circuit voltage estimated by using a learned open-circuit voltage estimation model, where the learned open-circuit voltage estimation model has learned the relationship between the voltage estimation input data and the open-circuit voltage of the secondary battery in operation.

[0019] According to another aspect of the present invention, the differential gradient is calculated using the least squares method.

[0020] According to another aspect of the present invention, the current difference and the voltage difference are respectively the fourth-order differences of the time series data of the terminal current and the fourth-order differences of the time series data of the terminal voltage.

[0021] According to another aspect of the present invention, the state estimation model is constituted by an RNN (Recurrent Neural Network).

[0022] According to another aspect of the present invention, the intermediate layer of the RNN constituting the state estimation model is constituted by LSTM (Long Short-Term Memory) or GRU (Gated Recurrent Unit).

[0023] According to another aspect of the present invention, the state estimation model is constituted by a one-dimensional CNN (Convolutional Neural Network).

[0024] According to another aspect of the present invention, the state estimation model is generated by learning using time series data of state variables including terminal current and terminal voltage for each of a plurality of secondary batteries having different electrical characteristics to which a load or a charger is connected.

[0025] Another aspect of the present invention is a method for estimating the state of a secondary battery, the method for estimating the state of the secondary battery including the steps of: measuring state variables including terminal current and terminal voltage of the working secondary battery to which a load or a charger is connected at a predetermined time interval; preprocessing the state variables to calculate state estimation input data; and estimating the current charge rate and / or degree of deterioration of the working secondary battery based on the state estimation input data using the learned state estimation model obtained by the learning method of the state estimation model of any of the above secondary batteries. In the step of performing the calculation, based on the time series data of the terminal current and the time series data of the terminal voltage, the difference of the terminal current, i.e., the current difference, and the difference of the terminal voltage, i.e., the voltage difference, are calculated, and based on the time series data of the current difference and the time series data of the voltage difference, the change rate of the voltage difference with respect to the current difference, i.e., the differential gradient, is calculated for the period from the past traced back by the first predetermined time to the present, and the state estimation input data including the time series data of the differential gradient is generated.

[0026] Another aspect of the present invention is a state estimation device for a secondary battery, the state estimation device for the secondary battery having: a state observation unit that measures state variables including a terminal current and a terminal voltage of the working secondary battery at a prescribed time interval; a preprocessing unit that preprocesses the state variables measured by the state observation unit to calculate state estimation input data; and a state estimation unit that estimates a current charge rate and / or degree of deterioration of the working secondary battery based on the state estimation input data using a learned state estimation model obtained by a learning method based on the state estimation model of any of the above secondary batteries, wherein the preprocessing unit performs the following processing: based on the time series data of the terminal current and the time series data of the terminal voltage obtained by the state observation unit, calculates a difference in the terminal current, i.e., a current difference, and a difference in the terminal voltage, i.e., a voltage difference, and based on the time series data of the current difference and the time series data of the voltage difference, calculates a rate of change of the voltage difference with respect to the current difference, i.e., a differential gradient, within a period from a past time traced back by a first prescribed time to the present, and generates the state estimation input data including the time series data of the differential gradient.

[0027] Effects of the Invention

[0028] According to the present invention, it is possible to accurately estimate the charge rate (SOC) and / or degree of deterioration (SOH) of secondary batteries having various electrical characteristics and different manufacturers or models during operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flowchart showing steps of a learning method of a state estimation model according to a first embodiment of the present invention.

[0030] Figure 2 is a diagram showing the execution of Figure 1 a structure of a machine learning device of the learning method of the state estimation model shown.

[0031] Figure 3 is a diagram showing Figure 1 a flowchart of detailed processing in a step of calculating state estimation input data in the learning method of the state estimation model shown.

[0032] Figure 4 is a diagram for explaining Figure 3 the calculation of a current difference in the processing shown.

[0033] Figure 5 is a diagram for explaining Figure 3 the calculation of a voltage difference in the processing shown.

[0034] Figure 6 is a diagram for explainingFigure 3 A diagram for explaining the calculation of the differential gradient in the processing shown.

[0035] Figure 7 Yes, it is for Figure 3 A diagram for explaining the calculation of the cumulative current value, open circuit voltage change, and differential gradient change amount in the processing shown.

[0036] Figure 8 It is for Figure 3 A diagram for explaining the calculation of the first gradient change rate in the processing shown.

[0037] Figure 9 It is for Figure 3 A diagram for explaining the calculation of the second gradient change rate in the processing shown.

[0038] Figure 10 It shows Figure 2 An example of the structure of the open circuit voltage estimation model generated by the model learning unit of the machine learning device shown.

[0039] Figure 11 It shows Figure 2 An example of the structure of the state estimation model generated by the model learning unit of the machine learning device shown.

[0040] Figure 12 A diagram showing an example of the state estimation of a secondary battery using the learned open circuit voltage estimation model and state estimation model.

[0041] Figure 13 A flowchart showing the steps of the state estimation method according to the second embodiment of the present invention.

[0042] Figure 14 It shows the execution of Figure 13 A diagram of the structure of the state estimation device for the state estimation method shown.

[0043] Figure 15 It is Figure 14 A functional block diagram of the processing device included in the state estimation device shown.

[0044] Explanation of reference numerals

[0045] 100: Machine learning device; 102, 404: Secondary battery; 104: Charger; 106: Load; 108: Switching switch; 110, 406: Characteristic measuring device; 112: Learning management device; 120, 420, 440: Processing device; 122, 428, 448: Storage device; 124, 430: Open-circuit voltage estimation model; 126, 432: State estimation model; 130: State variable measurement unit; 132: Input data generation unit; 134: Model learning unit; 200, 202, 204: Approximate straight line; 300, 310: Input layer; 302, 312: Intermediate layer; 304, 314: Output layer; 400: State estimation device; 402: Vehicle; 408: Power-on controller; 410: Rotating electric machine; 412: External charging device; 414: Travel control device; 422: State observation unit; 424: Preprocessing unit; 426: State estimation unit; 442: Motor control unit; 444: Charging control unit; 446: Notification control unit; 450: Display device; 452: Accelerator pedal sensor; 454: Brake pedal sensor; 456: Vehicle speed sensor; 500, 502, 504, 506, 508, 510; 512, 514, 516, 518: Processing; 600, 602, 604, 606: Line. DETAILED DESCRIPTION OF THE INVENTION

[0046] The inventors of the present invention have found that there is a correlation between the change patterns of the terminal current and terminal voltage of secondary batteries, i.e., the higher-order change patterns, and the internal states (OCV, SOC, and / or SOH) of the secondary batteries, at least among secondary batteries of the same type (for example, secondary batteries of the same type "lithium-ion battery"). Moreover, the inventors have obtained the following insight: By using the change rate (differential gradient described later) of the difference in the time series data of the terminal voltage (voltage difference) with respect to the difference in the time series data of the terminal current (current difference) as a parameter representing the higher-order change patterns of the terminal current and terminal voltage of the secondary battery and using it as the input to a model (such as a neural network), a model can be generated that can accurately estimate the states of secondary batteries with various electrical characteristics of different manufacturers or models. The present invention is based on such excellent insights.

[0047] Hereinafter, embodiments of the present invention will be described with reference to the drawings.

[0048] [First Embodiment]

[0049] Figure 1It is a diagram showing steps of a learning method of a state estimation model of a secondary battery according to a first embodiment of the present invention. The learning method of the state estimation model includes the following steps: a step (S100) of measuring state variables including a terminal current and a terminal voltage of an operating secondary battery connected to a load or a charger at a prescribed time interval; and a step (S102) of preprocessing the measured state variables to calculate state estimation input data. Further, the learning method of the open-circuit voltage estimation model includes a step (S104) of causing the state estimation model to learn, by machine learning, the relationship between the state estimation input data and a charge rate and / or a degree of deterioration, which are states of the operating secondary battery.

[0050] Figure 2 It is a diagram showing Figure 1 an example of the structure of a learning management device and a machine learning device that execute the learning method of the state estimation model shown. The learning management device and the machine learning device perform machine learning of an open-circuit voltage estimation model for estimating the open-circuit voltage of a secondary battery. The state estimation model and the open-circuit voltage estimation model are constituted by, for example, a neural network. The learning management device 112 controls the operation of the secondary battery 102 during the machine learning, and calculates measured values of the open-circuit voltage, SOC, and SOH as teaching data and provides them to the machine learning device 100.

[0051] The secondary battery 102 is charged by the charger 104 and discharged by being energized to the load 106. The charger 104 is, for example, a DC power supply, and the load 106 is, for example, a motor. Whether to charge the secondary battery 102 from the charger 104 or discharge the secondary battery 102 to the load 106 is selected by a changeover switch 108. A characteristic measurer 110 is inserted between the changeover switch 108 and the secondary battery 102.

[0052] The characteristic measurer 110 measures the current value of a prescribed state variable of the secondary battery 102. The prescribed state variable can include the terminal voltage Vte, the terminal current Ite, the internal impedance Z, and the temperature T (°C) of the surface of the case of the secondary battery 102. Here, the internal impedance Z can be measured according to the prior art, for example, by inputting an alternating current as a measurement signal to the secondary battery 102.

[0053] The terminal current Ite of the secondary battery 102 takes a positive value during discharge of the secondary battery 102 and a negative value during charging.

[0054] [1. Learning management device]

[0055] The learning management device 112 controls the charging and discharging of the secondary battery 102, and generates teaching data for learning the open-circuit voltage estimation model and the state estimation model and outputs it to the machine learning device 100. The learning management device 112 is, for example, a computer, starts operating according to an instruction from an operator, and gives instructions to start and stop power output to the charger 104 and instructions for switching operations to the changeover switch 108.

[0056] The learning management device 112 acquires the terminal current Ite, the terminal voltage Vte, and the internal impedance Z of the secondary battery 102 during charging and discharging from the characteristic measurer 110 at a prescribed time interval.

[0057] The learning management device 112 calculates the open-circuit voltage Voc of the secondary battery 102 based on the acquired terminal current Ite, terminal voltage Vte, and internal impedance Z, and generates time-series data of the open-circuit voltage Voc. The time-series data of the open-circuit voltage Voc is used as teaching data when learning the open-circuit voltage estimation model executed by the machine learning device 100 described later.

[0058] In addition, the learning management device 112 calculates the charge amount (full charge amount) when the secondary battery 102 is charged to the limit and the current charge amount based on the time-series data of the acquired terminal voltage Vte and terminal current Ite. In the present embodiment, the SOH is the above-mentioned full charge amount (unit: Ah), and the SOC is the ratio (%) of the current charge amount to the above-mentioned full charge amount.

[0059] [2. Machine learning device]

[0060] The machine learning device 100 includes a processing device 120 and a storage device 122. The storage device 122 is constituted by, for example, a volatile and / or non-volatile semiconductor memory and / or a hard disk device, etc. The storage device 122 stores the open-circuit voltage estimation model 124 and the state estimation model 126 generated by a model learning unit 134 described later.

[0061] The processing device 120 is, for example, a computer including a processor such as a CPU (Central Processing Unit). The processing device 120 may also have a structure including a ROM (Read Only Memory) in which a program is written, a RAM (Random Access Memory) for temporarily storing data, etc. Moreover, the processing device 120 includes a state variable measurement unit 130, an input data generation unit 132, and a model learning unit 134 as functional elements or functional units.

[0062] These functional elements included in the processing device 120 are implemented, for example, by the processing device 120, which is a computer, executing a program. Additionally, the above computer program can be pre-stored in any computer-readable storage medium. Instead, all or part of the above functional elements included in the processing device 120 can also be constituted by hardware including one or more electronic circuit components.

[0063] [Function of the state variable measurement unit]

[0064] The state variable measurement unit 130 executes Figure 1 the steps S100 shown. Specifically, the state variable measurement unit 130 obtains state variables including the terminal current Ite and the terminal voltage Vte of the secondary battery 102 to which the load 106 or the charger 104 is connected from the characteristic measurement unit 110 at a specified time interval. Thereby, the state variable measurement unit 130 measures the state variables at a specified time interval. The state variable measurement unit 130 may also measure the temperature T of the secondary battery 102 as a state variable at the above-specified time interval.

[0065] [Function of the input data generation unit]

[0066] The input data generation unit 132 generates voltage estimation input data for learning of the open-circuit voltage estimation model 124 based on the terminal current Ite and the terminal voltage Vte obtained by the state variable measurement unit 130. In addition, the input data generation unit 132 executes Figure 1 the steps S102 shown, for example, after the learning of the open-circuit voltage estimation model 124 is completed, and also generates state estimation input data for learning of the state estimation model 126 using the learned open-circuit voltage estimation model 124.

[0067] [2.2.1 Generation of voltage estimation input data]

[0068] The voltage estimation input data is the input data input to the open-circuit voltage estimation model 124 generated by the input data generation unit 132 for learning of the open-circuit voltage estimation model 124.

[0069] Specifically, first, the input data generation unit 132 calculates the difference of the terminal current Ite, that is, the current difference δIte, and the difference of the terminal voltage Vte, that is, the voltage difference δVte, based on the time series data of the terminal current Ite and the time series data of the terminal voltage Vte. In the present embodiment, the current difference δIte and the voltage difference δVte are respectively the fourth-order differences Δ 4 Ite of the time series data of the terminal current Ite and the fourth-order difference Δ 4 Vte of the time series data of the terminal voltage Vte.

[0070] In addition, based on the time series data of the calculated current difference δIte and the time series data of the voltage difference δVte described above, the input data generation unit 132 calculates the differential gradient Sdiff, which is the change rate of the voltage difference δVte with respect to the current difference δIte, within the period from the past traced back by a specified time T1 from the current time to the current time.

[0071] In addition, the calculation of the current difference δIte, the voltage difference δVte, and the differential gradient Sdiff will be described in detail in the description of the state estimation input data to be described later.

[0072] Moreover, the input data generation unit 132 generates voltage estimation input data including at least the following three time series data within the period from the past traced back by a specified time T2 from the current time to the current time, as input data for learning the open-circuit voltage estimation model 124.

[0073] Time series data of the terminal current Ite

[0074] Time series data of the terminal voltage Vte

[0075] Time series data of the differential gradient Sdiff

[0076] [2.2.2 Generation of State Estimation Input Data]

[0077] The state estimation input data is the input data input to the state estimation model 126 generated by the input data generation unit 132 in Figure 1 Step S102 of the learning method shown. In addition, after the learning of the open-circuit voltage estimation model is completed, the generation of the state estimation input data in step S102 is executed.

[0078] Figure 3 It shows Figure 1 The flowchart of the detailed processing in step S102 for calculating the state estimation input data. In step S102 for calculating the state estimation input data, the input data generation unit 132 first calculates the difference of the terminal current Ite, that is, the current difference δIte, and the difference of the terminal voltage Vte, that is, the voltage difference δVte, based on the time series data of the terminal current Ite and the time series data of the terminal voltage Vte (S200). Next, the input data generation unit 132 calculates the change rate of the voltage difference δVte with respect to the current difference δIte, that is, the differential gradient Sdiff, within the period from the past traced back by a specified time T1 from the current time to the current time (S202).

[0079] In addition, the input data generation unit 132 inputs the time series data of the terminal current Ite, the terminal voltage Vte, and the differential gradient Sdiff for the period from the past traced back a specified time T2 from the current to the present to the open-circuit voltage estimation model 430, and estimates the current open-circuit voltage Voc (S204). Next, the input data generation unit 132 subtracts the differential gradient Sdiff from the past traced back a specified time T3 from the current from the current differential gradient Sdiff, and calculates the differential gradient change amount Dsd (S206).

[0080] Next, the input data generation unit 132 calculates the cumulative current value ΣIte, which is the sum of the terminal current Ite measured during the period from the past traced back a specified time T3 from the current to the present (S208). Then, the input data generation unit 132 calculates the change rate of the differential gradient change amount Dsd with respect to the cumulative current value ΣIte for the period from the past traced back a specified time T4 from the current to the present, as the first gradient change rate R1 (S210).

[0081] In addition, the input data generation unit 132 subtracts the open-circuit voltage Voc from the past traced back a specified time T3 from the current from the current open-circuit voltage Voc to calculate the open-circuit voltage change amount Dvoc (S212). Next, the input data generation unit 132 calculates the change rate of the differential gradient change amount Dsd with respect to the open-circuit voltage change amount Dvoc for the period from the past traced back a specified time T4 from the current to the present, as the second gradient change rate R2 (S214).

[0082] Then, the input data generation unit 132, for example, generates state estimation input data including the following four time series data for the period from the past traced back a specified time T5 from the current to the present (S216), and ends the process.

[0083] Time series data of the differential gradient Sdiff

[0084] Time series data of the open-circuit voltage Voc

[0085] Time series data of the first gradient change rate R1

[0086] Time series data of the second gradient change rate R2

[0087] Hereinafter, specific calculation methods for the current difference δIte, the voltage difference δVte, the differential gradient Sdiff, the cumulative current value ΣIte, the open-circuit voltage change amount Dvoc, the differential gradient change amount Dsd, the first gradient change rate R1, the second gradient change rate R2, and the voltage estimation input data and the state estimation input data will be described.

[0088] [2.2.2.1 Calculation of Current Difference δIte]

[0089] Figure 4 is a diagram for explaining the calculation of the current difference δIte. In the Figure 4 shown table, the leftmost column is set as the 1st column, and the columns to the right are the 2nd column, 3rd column, …… 6th column in sequence. Figure 4 The 1st column of the table represents the moment or the index (number) of the moment when the state variable measurement unit 130 repeatedly acquires the terminal current Ite at time intervals dt. The 2nd column is the time series data of the terminal current Ite, representing the terminal current Ite acquired at each moment.

[0090] The 3rd column, 4th column, 5th column, and 6th column respectively represent the first-order difference Δ 1 Ite, second-order difference Δ 2 Ite, third-order difference Δ 3 Ite, and fourth-order difference Δ 4 Ite of the terminal current Ite calculated based on the terminal current Ite in the 2nd column.

[0091] The h-order difference Δ n Ite(t h )(h = 1, 2, …… 4) at the current moment t n is calculated by the following formula.

[0092] Δ h Ite(t n ) = Δ h-1 Ite(t n ) - Δ h-1 Ite(t n-1 )

[0093] Here, h = 1, 2, 3, 4. In addition, let Δ 0 Ite(t n ) = Ite(t n ).

[0094] That is, the first-order difference Δ n Ite(t 1 ) at the moment t n is calculated by subtracting the terminal current Ite(t n ) at the moment t n from the terminal current Ite(t n-1 ) at the moment t n-1 . In addition, the second-order difference Δ n Ite(t 2 ) at the moment t n is the first-order difference Δ n Ite(t 1 ) at the moment t n) Subtract the moment t n-1 The first-order difference Δ 1 Ite(t n-1 ) is calculated.

[0095] Similarly, the third-order difference Δ n of the moment t 3 Ite(t n ) is obtained by subtracting the second-order difference Δ n Ite(t 2 ) of the moment t n from the second-order difference Δ n-1 Ite(t 2 ) of the moment t n-1 . The fourth-order difference Δ n of the moment t 4 Ite(t n ) is obtained by subtracting the third-order difference Δ n Ite(t 3 ) of the moment t n from the third-order difference Δ n-1 Ite(t 3 ) of the moment t n-1 ) and is calculated.

[0096] In the present embodiment, the input data generation unit 132 sets the fourth-order difference Δ 4 Ite of the terminal current Ite at each moment as the current difference δIte. That is,

[0097] δIte(t) = Δ 4 Ite(t) t = t n , t n-1 ,....

[0098] [2.2.2.2 Calculation of Voltage Difference δVte]

[0099] The input data generation unit 132 calculates the voltage difference δVte of the terminal voltage Vte in the same manner as the above-described current difference. Figure 5 is a diagram showing the calculation steps of the voltage difference δVte. In Figure 5 the shown table, the leftmost column is set as the first column, and the columns to the right are the second column, the third column,... the sixth column in sequence. Figure 5 The first column of the table of

[0100] represents the moment when the state variable measurement unit 130 repeatedly acquires the terminal voltage Vte at time intervals dt or the index (number) of that moment. The second column is the time series data of the terminal voltage Vte, representing the terminal voltage Vte acquired at each moment.

[0100] The third column, the fourth column, the fifth column, and the sixth column respectively represent the first-order difference Δ 1Vte, second-order difference Δ 2 Vte, third-order difference Δ 3 Vte, and fourth-order difference Δ 4 Vte.

[0101] Current time t n The h-order difference Δ h Vte(t n )(h = 1, 2, …… 4) is calculated by the following formula.

[0102] Δ h Vte(t n ) = Δ h-1 Vte(t n ) - Δ h-1 Vte(t n-1 )

[0103] Here, h = 1, 2, 3, 4.

[0104] In addition, let Δ 0 Vte(t n ) = Vte(t n ).

[0105] That is, the first-order difference Δ n Vte(t 1 ) at time t n is calculated by subtracting the terminal voltage Vte(t n ) at time t n from the terminal voltage Vte(t n-1 ) at time t n-1 . In addition, the second-order difference Δ n Vte(t 2 ) at time t n is calculated by subtracting the first-order difference Δ n Vte(t 1 ) at time t n from the first-order difference Δ n-1 Vte(t 1 ) at time t n-1 .

[0106] Similarly, the third-order difference Δ n Vte(t 3 ) at time t n is calculated by subtracting the second-order difference Δ n Vte(t 2 ) at time t n from the second-order difference Δ n-1 Vte(t 2 ) at time t n-1 , and the fourth-order difference Δ n Vte(t 4Vte(t n ) is calculated by subtracting the third-order difference Δ n Vte(t 3 ) at time t n from the third-order difference Δ n-1 Vte(t 3 ) at time t n-1 .

[0107] In this embodiment, the input data generation unit 132 sets the fourth-order difference Δ 4 Vte of the terminal voltage Vte at each time as the voltage difference δVte. That is,

[0108] δVte(t) = Δ 4 Vte(t) at t = t n , t n-1 ,....

[0109] [2.2.2.3 Calculation of differential gradient Sdiff]

[0110] The differential gradient Sdiff is the rate of change of the voltage difference δVte with respect to the current difference δIte during the period from the past traced back by a specified time T1 to the present. Specifically, as shown in Figure 4 and Figure 5 , the input data generation unit 132 extracts k1 (k1 = n - m + 1) current differences δIte and voltage differences δVte during the period from the past time t m to the current time t n . Among them, the period from the past time t m to the current time t n corresponds to the period from the past traced back by a specified time T1 to the present. Moreover, the input data generation unit 132 calculates the differential gradient Sdiff, which is the rate of change of the voltage difference δVte with respect to the current difference δIte, according to the data set (δIte, δVte) at each time composed of the extracted δIte and δVte by the least squares method.

[0111] More specifically, as shown in Figure 6 , the slope of the approximate straight line (regression straight line) 200 of the above k1 data sets (δIte, δVte) on the two-dimensional plane with the current difference δIte as the horizontal axis and the voltage difference δVte as the vertical axis (the black circles inside the illustrated dotted ellipse) corresponds to the differential gradient Sdiff. That is, when the approximate straight line 200 is given by δVte = a1×δIte + b1, the slope a1 of this approximate straight line 200 corresponds to the differential gradient Sdiff. Here, the approximate straight line is calculated by the least squares method, for example.

[0112] [Calculation of cumulative current value ΣIte, open-circuit voltage change Dvoc, and differential gradient change Dsd]

[0113] Next, the calculation of the cumulative current value ΣIte, the open-circuit voltage change Dvoc, and the differential gradient change Dsd will be described.

[0114] Figure 7 It is a diagram for explaining the calculation of the cumulative current value ΣIte, the open-circuit voltage change Dvoc, and the differential gradient change Dsd. In Figure 7 the shown table, the leftmost column is set as the 1st column, and to the right are the 2nd column, 3rd column, ……, 7th column in sequence. Figure 7 The 1st column of the table represents the time or the index (number) of the moment when the state variable measurement unit 130 repeatedly acquires the terminal current Ite at time intervals dt. The 2nd column is the time series data of the terminal current Ite, representing the terminal current Ite acquired at each moment. In addition, the 3rd column represents the cumulative current value ΣIte calculated based on the time series data of the terminal current Ite in the 2nd column.

[0115] The 4th column is the time series data of the open-circuit voltage Voc, representing the open-circuit voltage Voc acquired at each moment. In this embodiment, the open-circuit voltage Voc is calculated and acquired using the learned open-circuit voltage estimation model 124. In addition, the 5th column represents the open-circuit voltage change Dvoc calculated based on the time series data of the open-circuit voltage Voc in the 4th column.

[0116] The 6th column is the differential gradient Sdiff calculated as described above, and the 7th column represents the differential gradient change Dsd calculated based on the time series data of the differential gradient Sdiff in the 6th column.

[0117] Current moment t n The cumulative current value ΣIte(t n ) is the sum of the terminal currents Ite measured during the period from the past moment t p retrospective from the current by a specified time T3 to the current moment t n and is calculated by the following formula.

[0118]

[0119] Current moment t n The open-circuit voltage change Dvoc(t n ) is obtained by subtracting the open-circuit voltage Voc(t n ) at the past moment t n retrospective from the current by a specified time T3 from the open-circuit voltage Voc(t p ) at the current moment t p) The obtained value is calculated by the following formula.

[0120] Dvoc(t n ) = Voc(t n ) - Voc(t p )

[0121] The differential gradient change amount Dsd(t n ) at the current time t n is obtained by subtracting the differential gradient Sdiff(t n ) at the current time t n from the differential gradient Sdiff(t p ) at the past time t p which is traced back from the current time by a specified time T3, and is calculated by the following formula.

[0122] Dsd(t n ) = Sdiff(t n ) - Sdiff(t p )

[0123] [Calculation of the first gradient change rate R1 and the second gradient change rate R2 in 2.2.2.5]

[0124] Next, the calculation of the first gradient change rate R1 and the second gradient change rate R2 will be described.

[0125] The first gradient change rate R1 is the change rate of the differential gradient change amount Dsd with respect to the cumulative current value ΣIte during the period from the past time traced back from the current by a specified time T4 to the current. Specifically, as Figure 7 shown, the input data generation unit 132 extracts k2 (k2 = n - q + 1) cumulative current values ΣIte and differential gradient change amounts Dsd during the period from the past time t q to the current time t n , where the period from the past time t q to the current time t n corresponds to the period from the past time traced back from the current by a specified time T4 to the current. Moreover, the input data generation unit 132 calculates the first gradient change rate R1, which is the change rate of the differential gradient change amount Dsd with respect to the cumulative current value ΣIte, based on the data groups (ΣIte, Dsd) at each time of the extracted ΣIte and Dsd by the least squares method.

[0126] More specifically, as Figure 8As shown, the slope of the approximate line (regression line) 202 of the above k2 data sets (ΣIte, Dsd) on the two-dimensional plane with the cumulative current value ΣIte as the horizontal axis and the differential gradient change amount Dsd as the vertical axis (the black circles within the dotted ellipse shown in the figure) corresponds to the first gradient change rate R1. That is, when the approximate line 202 is given by Dsd = a2 × ΣIte + b2, the slope a2 of this approximate line 202 corresponds to the first gradient change rate R1. Here, the approximate line is calculated, for example, by the least squares method.

[0127] The second gradient change rate R2 is the change rate of the differential gradient change amount Dsd with respect to the open-circuit voltage change amount Dvoc during the period from the past that traces back a specified time T4 from the current time to the current time. Specifically, as Figure 7 shown, the input data generation unit 132 extracts the open-circuit voltage change amount Dvoc and the differential gradient change amount Dsd of k2 (k2 = n - q + 1) during the period from the past time t q to the current time t n . Among them, the period from the past time t q to the current time t n corresponds to the period from the past that traces back a specified time T4 from the current time to the current time. Moreover, the input data generation unit 132 calculates the second gradient change rate R2, which is the change rate of the differential gradient change amount Dsd with respect to the open-circuit voltage change amount Dvoc, by the least squares method according to the data sets (Dvoc, Dsd) of each moment of the extracted Dvoc and Dsd.

[0128] More specifically, as Figure 9 shown, the slope of the approximate line (regression line) 204 of the above k2 data sets (Dvoc, Dsd) on the two-dimensional plane with the open-circuit voltage change amount Dvoc as the horizontal axis and the differential gradient change amount Dsd as the vertical axis (the black circles within the dotted ellipse shown in the figure) corresponds to the second gradient change rate R2. That is, when the approximate line 204 is given by Dsd = a3 × Dvoc + b3, the slope a3 of this approximate line 204 corresponds to the second gradient change rate R2. Here, the approximate line is calculated, for example, by the least squares method.

[0129] [2.2.2.6 Voltage Estimation Input Data]

[0130] As described above, the voltage estimation input data is composed of time series data of the terminal current Ite, the terminal voltage Vte, and the differential gradient Sdiff during the period from the past that traces back a specified time T2 from the current time to the current time. When the current time is set as t n and the past time that traces back a specified time T2 from the current time is set as t rWhen, the voltage estimation input data is expressed by the following formula.

[0131]

[0132] Wherein,

[0133] V Ite(t n ) = (Ite(t r ), Ite(t r+1 ), Ite(t r+2 ),...., Ite(t n ))

[0134] V Vte(t n ) = (Vte(t t ), Vte(t r+1 ), Vte(t r+2 ),...., Vte(t n ))

[0135] V Sdiff(t n ) = (Sdiff(t r ), Sdiff(t r+1 ), Sdiff(t r+2 ),...., Sdiff(t n ))

[0136] Here, the time series data of the terminal current Ite V Ite(t n ), the time series data of the terminal voltage Vte V Vte(t n ), and the time series data of the differential gradient Sdiff V Sdiff(t n ) are first-order tensors having n - r + 1 values from time t r to time t n as elements for the terminal current Ite, the terminal voltage Vte, and the differential gradient Sdiff. Therefore, the voltage estimation input data V x1(t n ) is a second-order tensor.

[0137] [2.2.2.7 State Estimation Input Data]

[0138] As described above, the state estimation input data is composed of the time series data of the differential gradient Sdiff, the open-circuit voltage Voc, the first gradient change rate R1, and the second gradient change rate R2 during the period from the past traced back by a specified time T5 to the present. When the current time is set as tn Let the time in the past when a specified time T5 has been traced from the current time be t s At this time, the state estimation input data is represented by the following formula

[0139]

[0140] where

[0141] V Sdiff(t n ) = (Sdiff(t s ), Sdiff(t s+1 ), Sdiff(t s+2 ),...., Sdiff(t n ))

[0142] V Voc(t n ) = (Voc(t s ), Voc(t s+1 ), Voc(t s+2 ),...., Voc(t n ))

[0143] V R1(t n ) = (R1(t s ), R1(t s+1 ), R1(t s+2 ),...., R1(t n ))

[0144] V R2(t n ) = (R2(t s ), R2(t s+1 ), R2(t s+2 ),...., R2(t n ))

[0145] Here, the time series data of the differential gradient Sdiff V Sdiff(t n ), the time series data of the open circuit voltage Voc V Voc(t n ), the time series data of the first gradient change rate R1 V R1(t n ) and the time series data of the second gradient change rate R2 V R2(t n ) are respectively the differential gradient Sdiff, open circuit voltage Voc, first gradient change rate R1, and second gradient change rate R2 from time t s to time tn The n - s + 1 values are used as the first - order tensor of the elements. Thus, the state - estimation input data V x2(t n ) is a second - order tensor.

[0146] [Function of the model learning unit]

[0147] The model learning unit 134 generates an open - circuit voltage estimation model 124 through machine learning. In addition, the model learning unit 134 performs Figure 1 the step S104 shown in the figure, and makes the state - estimation model 126 learn through machine learning.

[0148] [Generation of the open - circuit voltage estimation model]

[0149] The model learning unit 134 uses the voltage - estimation input data (described above) generated by the input - data generation unit 132 to generate an open - circuit voltage estimation model 124 through machine learning. At this time, the model learning unit 134, for example, obtains the time - series data of the open - circuit voltage Voc of the secondary battery 102 from the learning management device 112, and uses the obtained time - series data of the open - circuit voltage Voc as the teaching data to perform the above - mentioned machine learning.

[0150] Figure 10 FIG. is a diagram showing the structure of the open - circuit voltage estimation model 124 generated by the model learning unit 134. The open - circuit voltage estimation model 124 is composed of a neural network and has an input layer 300, an intermediate layer 302, and an output layer 304. The open - circuit voltage estimation model 124 is, for example, an RNN (Recurrent Neural Network).

[0151] The input layer 300 receives the voltage - estimation input data which is the second - order tensor shown as Equation (1) above. The intermediate layer 302 includes LSTM (Long Short Term Memory) configured as multiple layers in the present embodiment. However, the intermediate layer 302 is not limited to LSTM. For example, the intermediate layer 302 may also be composed of GRU (Gated Recurrent Unit).

[0152] The output layer 304 outputs an estimated value of the open - circuit voltage Voc of the secondary battery 102 at time t n as an output y1(t n ). That is, the output y1(t n ) is the open - circuit voltage Voc(t n ) as a scalar.

[0153] [Generation of the state - estimation model]

[0154] The model learning unit 134 executes Figure 1 Step S104 shown above. Using the state estimation input data generated by the input data generation unit 132, the state estimation model 126 is learned through machine learning. At this time, the model learning unit 134 obtains, for example, the time series data of the SOC and SOH of the secondary battery 102 calculated by the learning management device 112, and uses the obtained time series data of the SOC and SOH as teaching data to perform the above-mentioned machine learning.

[0155] Figure 11 FIG. is a diagram showing the structure of the state estimation model 126 generated by the model learning unit 134. The state estimation model 126 is composed of a neural network having an input layer 310, an intermediate layer 312, and an output layer 314. For example, the state estimation model 126 is an RNN.

[0156] The input layer 310 receives the state estimation input data of the second-order tensor shown as Equation (2) above. The intermediate layer 312 includes LSTM layers in this embodiment. However, the intermediate layer 312 is not limited to LSTM. For example, the intermediate layer 312 may also be composed of GRU.

[0157] The output layer 314 outputs the estimated value of the SOC and the estimated value of the SOH of the secondary battery 102 at time t n as outputs V y2(t n ). That is, the output V y2(t n ) is a first-order tensor having the SOC(t n ) and SOH(t n ) as elements, where SOC and SOH are scalars.

[0158] The open-circuit voltage estimation model 124 and the state estimation model 126 generated as described above do not directly use the terminal current Ite and the terminal voltage Vte of the secondary battery as inputs, but use the differential gradient Sdiff as the input. The differential gradient Sdiff is the change gradient of the voltage difference δVte calculated from the time series data of the terminal voltage Vte with respect to the current difference δIte calculated from the time series data of the terminal current Ite. That is, the open-circuit voltage estimation model 124 and the state estimation model 126 do not learn the relationship between the change patterns of the terminal current Ite and the terminal voltage Vte themselves and the SOC, etc., but learn the change pattern of the change of the terminal current Ite and the terminal voltage Vte, that is, the relationship between the higher-order change pattern and the SOC, etc.

[0159] As described above, regarding the high-order change patterns of such terminal current and terminal voltage, if they are at least the same type of secondary battery (for example, the same secondary battery within the type of "lithium-ion battery"), there is a correlation with the internal state of the secondary battery. Therefore, the open-circuit voltage estimation model 124 and the state estimation model 126 generated as described above can accurately estimate the open-circuit voltage, charge rate (SOC), and / or degree of deterioration (SOH) of secondary batteries with various electrical characteristics that are different in terms of manufacturer or model during the operation of these secondary batteries.

[0160] [3. Secondary Battery for Model Learning]

[0161] As the secondary battery 102 for model learning, it is desirable to use a variety of multiple secondary batteries with different electrical characteristics due to different manufacturers or models. Thereby, an open-circuit voltage estimation model 124 and a state estimation model 126 with less variation in estimation accuracy for different manufacturers or models can be generated. For example, during the learning of the open-circuit voltage estimation model 124 and the state estimation model 126, it is desirable to use multiple secondary batteries with different electrical characteristics such as SOC-OCV characteristics, SOC-internal impedance characteristics, and / or their SOH dependencies.

[0162] [4. Operation Mode of Secondary Battery during Model Learning]

[0163] Regarding the operation mode (charge-discharge process) of the secondary battery during model learning, it is not only monotonously discharging or charging between the fully charged state (SOC = 100%) and the fully discharged state (SOC = 0%), but also desirable to randomly perform charge and discharge and / or alternately perform charge and discharge according to a specified criterion. Such a specified criterion can be a criterion corresponding to the use of the secondary battery to be estimated. For example, in the case of assuming a vehicle secondary battery as the estimation object, a criterion adjusted in a manner that follows the typical charge-discharge cycles during vehicle driving in various traffic scenarios such as urban areas, mountainous areas, rural areas, and highways can be used.

[0164] [5. Collection of Learning Data]

[0165] In this embodiment, the machine learning device 100 obtains the state variables (Ite, Vte) of the secondary battery 102 that form the basis of the learning data for the open-circuit voltage estimation model 124 and the state estimation model 126, as well as the time-series data of SOC and SOH as teaching data, from the characteristic measuring device 110, and calculates them through the learning management device 112 and immediately uses them for the learning of the open-circuit voltage estimation model 124 and the state estimation model 126. However, the time-series data of these state variables and teaching data do not necessarily have to be immediately used for learning.

[0166] The learning management device 112 may also cause the secondary battery to operate in advance to obtain and store the time series data of the state variables and the time series data of the teaching data. The machine learning device 100 may also obtain from the learning management device 112 the time series data of the state variables and the time series data of the teaching data stored by the learning management device 112, and perform learning of the open circuit voltage estimation model 124 and the state estimation model 126.

[0167] In addition, regarding the time series data of the state variables and the time series data of the teaching data, as long as the error from the actual data is within a range that is practically acceptable, they may also be generated by simulating the charge and discharge characteristics grasped from design materials such as the equivalent circuit of the secondary battery 102 using a computer.

[0168] [6. Example of State Estimation Using the State Estimation Model]

[0169] Next, an example of state estimation of the secondary battery using the state estimation model learned by the learning method of this embodiment will be described. Figure 12 It is a diagram showing an example of state estimation of the secondary battery using the learned open circuit voltage estimation model and state estimation model.

[0170] The learning data for both the open circuit voltage estimation model and the state estimation model are generated by simulating the charge and discharge characteristics of dozens of sample secondary batteries for vehicles with different electrical characteristics using a computer. Specifically, for each of dozens of sample secondary batteries with different SOC-OCV characteristics, internal impedance characteristics, and capacity characteristics (SOH) as electrical characteristics, the terminal current Ite, terminal voltage Vte, SOC, and SOH at every specified time interval dt when charging and discharging are performed according to a specified charge and discharge process are calculated through the simulation of the above computer.

[0171] The above charge and discharge process not only monotonously discharges or charges the sample secondary battery between the fully charged state (SOC = 100%) and the fully discharged state (SOC = 0%), but also is adjusted in a manner that follows the typical charge and discharge cycles during vehicle driving in various traffic scenarios such as urban areas, mountainous areas, rural areas, and highways.

[0172] The sample secondary battery is a lithium-ion battery. In addition, the measurement interval dt of the state variables is 100 ms, and the specified times T1, T2, T3, T4, and T5 when calculating the voltage estimation input data and state estimation input data of the above open circuit voltage estimation model and state estimation model are 5 seconds, 5 seconds, 300 seconds, 200 seconds, and 5 seconds respectively. Additionally, the numerical values of these times are an example, and the specified times T1, T2, T3, T4, and T5 may also be set to time values different from the above values.

[0173] Figure 12 Shows the estimation results of SOC and SOH obtained using the learned state estimation model, as well as the simulated values of SOC and SOH, during the period when a secondary battery arbitrarily selected from the above sample secondary batteries (hereinafter referred to as "target secondary battery") is discharged from the fully charged state to the fully discharged state.

[0174] In Figure 12 , the horizontal axis is the elapsed time after the start of discharge when the secondary battery starts to discharge from the fully charged state, and the vertical axis is the SOC (%) and SOH (Ah) of the target secondary battery. The state estimation input data provided to the state estimation model in the estimation of SOC and SOH is calculated based on Ite and Vte at every predetermined time interval dt during the discharge of the target secondary battery, which are calculated according to the charge-discharge characteristics of the target secondary battery through simulation.

[0175] In Figure 12 , the lines 600 and 602 formed by the set of gray points are the estimated values of SOC and SOH estimated by the state estimation model, respectively. In addition, the lines 602 and 604 formed by the set of black points are the SOC and SOH calculated through simulation according to the charge-discharge characteristics of the target secondary battery, respectively.

[0176] According to Figure 12 the comparison between the line 600 and the line 604 shown, and the comparison between the line 602 and the line 606, it can be seen that the state estimation model learned by using the learning method shown in this embodiment accurately estimates the SOC and SOH for the target secondary battery. In particular, although the state estimation model used for this estimation is generated using the learning data of dozens of sample secondary batteries with different electrical characteristics, the estimated values of SOC and SOH do not diverge but converge into a single line (lines 600 and 602), accurately estimating the SOC and SOH for a specific target secondary battery. Therefore, it can be seen that the state estimation method of this embodiment uses multiple secondary batteries with different electrical characteristics for learning, and thus can accurately estimate the states of various working secondary batteries with different manufacturers and models, that is, SOC and SOH.

[0177] [Second Embodiment]

[0178] Next, the second embodiment of the present invention will be described. Figure 13It is a diagram showing steps of a method for estimating the state of a secondary battery according to an embodiment of the present invention. The state estimation method includes the following steps: a step (S300) of measuring state variables including terminal current and terminal voltage of an operating secondary battery connected to a load or a charger at a prescribed time interval; and a step (S302) of preprocessing the measured state variables to calculate state estimation input data. In addition, the state estimation method includes a step (S304) of estimating the state of the operating secondary battery according to the state estimation input data by using a state estimation model learned by the learning method of the first embodiment described above. The above state is SOC and SOH in the present embodiment.

[0179] Figure 13 The state estimation method shown, for example, in Figure 14 is executed in the state estimation device 400 shown. The state estimation device 400 is mounted, for example, on a vehicle 402 which is an electric vehicle, and estimates the state of an operating secondary battery 404 which is an in-vehicle battery of the vehicle 402. The secondary battery 404 is connected to a rotating electric machine 410 via a characteristic measurer 406 and a power supply controller 408.

[0180] The rotating electric machine 410 functions as a motor that is powered by discharging from the secondary battery 404 to drive the wheels of the vehicle 402, and also functions as a generator that generates electricity by the rotational force transmitted from the wheels to charge the secondary battery 404.

[0181] The characteristic measurer 406 measures the current values of state variables of the secondary battery 404 including terminal current Ite and terminal voltage Vte. The power supply controller 408 controls the amount of power supplied from the secondary battery 404 to the rotating electric machine 410 and the amount of power supplied from the rotating electric machine 410 to the secondary battery 404 under the control of a travel control device 414 mounted on the vehicle 402. In addition, when an external charging device 412 located outside the vehicle 402 is connected to the vehicle 402, the power supply controller 408 controls the amount of power supplied from the external charging device 412 to the secondary battery 404 under the control of the travel control device 414. The external charging device 412 is, for example, a charger at a charging station. In addition, when another generator driven by an internal combustion engine is mounted on the vehicle 402, the power supply controller 408 can also control the amount of power supplied from the generator to the secondary battery.

[0182] The travel control device 414 obtains estimated values of the current SOC and SOH indicating the state of the secondary battery 404 from the state estimation device 400, controls the operation of the rotating electric machine 410 based on the obtained SOC and SOH, and notifies the user.

[0183] Specifically, the driving control device 414 has a processing device 440 and a storage device 448. The storage device 448 is, for example, a semiconductor memory and stores data required in the processing of the processing device 440.

[0184] The processing device 440 is, for example, a computer including a processor such as a CPU. The processing device 440 may also have a structure including a ROM in which a program is written, a RAM for temporarily storing data, and the like. Moreover, the processing device 440 includes a motor control unit 442, a charge control unit 444, and a notification control unit 446 as functional elements or functional units.

[0185] These functional elements included in the processing device 440 are realized, for example, by the processing device 440, which is a computer, executing a program. Additionally, the above-mentioned computer program can be pre-stored in any computer-readable storage medium. Instead, all or part of the above-mentioned functional elements included in the processing device 440 can also be constituted by hardware including one or more electronic circuit components.

[0186] The motor control unit 442 detects the amount of depression of an accelerator pedal (not shown) of the vehicle 402 based on the accelerator pedal sensor 452. When the accelerator pedal is depressed, the driving control device 414 instructs the power supply controller 408 to supply power from the secondary battery 404 to the rotary electric machine 410, causing the rotary electric machine 410 to operate as a motor and driving the vehicle 402. In addition, the driving control device 414 controls the rotational speed of the rotary electric machine 410 via the power supply controller 408 so that the speed of the vehicle 402 obtained from the vehicle speed sensor 456 becomes a speed corresponding to the amount of depression of the above-mentioned accelerator pedal.

[0187] At this time, the motor control unit 442 limits, based on the current SOC estimated value obtained from the state estimation device 400, for example, the upper limit value (maximum power supply current) of the current supplied from the secondary battery 404 to the rotary electric machine 410 during acceleration or constant-speed driving of the vehicle 402. That is, the motor control unit limits the discharge of the secondary battery 404, for example, to limit the generated torque of the rotary electric machine 410, and determines the maximum power supply current in such a way that the fuel efficiency (e.g., the driving distance per 1 kWh) determined according to the characteristics of the secondary battery 404 and the rotary electric machine 410 is not less than a specified value.

[0188] The charging control unit 444 determines whether the brake pedal (not shown) of the vehicle 402 is depressed through the brake pedal sensor 454. Then, when the brake pedal is depressed, the charging control unit 444 instructs the motor control unit 442 to stop energizing the rotating electric machine 410 from the secondary battery 404. Then, the charging control unit 444 instructs the power supply controller 408 to perform power supply from the rotating electric machine 410 to the secondary battery 404, causing the rotating electric machine 410 to operate as a generator and charging the secondary battery 404 from the rotating electric machine 410 (so-called regenerative braking operation).

[0189] In addition, when the external charging device 412 is connected to the vehicle 402, the charging control unit 444 controls the power supply amount from the external charging device 412 to the secondary battery 404 via the power supply controller 408.

[0190] The notification control unit 446 performs prescribed display on the display device 450 based on the current SOC estimated value and SOH estimated value obtained from the state estimation device 400. For example, the notification control unit 446 only displays the obtained current SOC estimated value and SOH estimated value on the display device 450. In addition, for example, when the SOC estimated value is lower than a prescribed value, the notification control unit 446 displays a message on the display device 450 proposing to the driver of the vehicle 402 to charge at a charging station. Or, for example, when the SOH estimated value is lower than a prescribed value, the notification control unit 446 displays a message on the display device 450 proposing to the driver of the vehicle 402 to replace the secondary battery 404.

[0191] The state estimation device 400 executes Figure 13 the above-described state estimation method of the secondary battery, estimates the SOC and SOH of the working secondary battery 404, and outputs the current SOC estimated value and SOH estimated value to the driving control device 414.

[0192] Specifically, the state estimation device 400 includes a processing device 420 and a storage device 428. The storage device 428 is composed of non-volatile and volatile semiconductor memories. In the storage device 428, the learned open-circuit voltage estimation model 124 and state estimation model 126 obtained by the learning method shown in the first embodiment are pre-stored as the open-circuit voltage estimation model 430 and state estimation model 432.

[0193] The processing device 420 is a computer including a processor such as a CPU, for example. The processing device 420 may also have a structure including a ROM in which a program is written, a RAM for temporarily storing data, and the like. Moreover, the processing device 420 includes a state observation unit 422, a preprocessing unit 424, and a state estimation unit 426 as functional elements or functional units.

[0194] These functional elements included in the processing device 420 are implemented, for example, by the processing device 420, which is a computer, executing a program. Additionally, the above-mentioned computer program can be pre-stored in any computer-readable storage medium. Instead, all or part of the above-mentioned functional elements included in the processing device 420 can also be constituted by hardware including one or more electronic circuit components.

[0195] Figure 15 The functional block diagram of the processing device 420 having a state observation unit 422, a preprocessing unit 424, and a state estimation unit 426 is shown. In Figure 15 this, the dashed rectangles respectively represent the processing in the preprocessing unit 424.

[0196] The state observation unit 422 executes Figure 13 the step S300 shown. Specifically, the state observation unit 422 acquires the state variables of the secondary battery 404 including the terminal current Ite(t) and the terminal voltage Vte(t) of the working secondary battery 404 from the characteristic measurer 406 at a prescribed time interval. Thereby, the state observation unit 422 obtains the time series data of the state variables measured at the prescribed time interval.

[0197] The preprocessing unit 424 executes Figure 13 the step S302 shown. Specifically, the preprocessing unit 424 preprocesses the state variables acquired by the state observation unit 422 and calculates the state estimation input data for the state estimation model 432. Specifically, the preprocessing unit 424 calculates the difference in the terminal current Ite, i.e., the current difference δIte, and the difference in the terminal voltage Vte, i.e., the voltage difference δVte ( Figure 15 the process 500 shown). Then, the preprocessing unit 424 calculates the rate of change of the voltage difference δVte with respect to the current difference δIte, i.e., the difference gradient Sdiff ( Figure 15 the process 502 shown).

[0198] In addition, the preprocessing unit 424 inputs the time series data of the terminal current Ite, the terminal voltage Vte, and the calculated Sdiff, respectively, within the period from the past traced back a prescribed time T2 from the current time to the current time, into the open-circuit voltage estimation model 430 ( Figure 15 the process 504), and calculates the estimated value of the current open-circuit voltage Voc of the secondary battery 404 through the open-circuit voltage estimation model 430 ( Figure 15 the process 506).

[0199] In addition, the preprocessing unit 424 calculates the cumulative current value ΣIte, which is the sum of the terminal currents Ite continuously obtained during the period from the past that is traced back by a specified time T3 from the current time to the current time ( Figure 15 processing 508). In addition, the preprocessing unit 424 calculates the differential gradient change amount Dsd obtained by subtracting the differential gradient Sdiff from the past that is traced back by a specified time T3 from the current differential gradient Sdiff ( Figure 15 processing 510).

[0200] Then, the preprocessing unit 424 calculates the change rate of the differential gradient change amount Dsd with respect to the cumulative current value ΣIte during the period from the past that is traced back by a specified time T4 from the current time to the current time, as the first gradient change rate R1 ( Figure 15 processing 512).

[0201] In addition, the preprocessing unit 424 subtracts the open-circuit voltage Voc from the past that is traced back by a specified time T3 from the current open-circuit voltage Voc, thereby calculating the open-circuit voltage change amount Dvoc ( Figure 15 processing 514).

[0202] In addition, the preprocessing unit 424 calculates the change rate of the differential gradient change amount Dsd with respect to the open-circuit voltage change amount Dvoc during the period from the past that is traced back by a specified time T4 from the current time to the current time, as the second gradient change rate R2 ( Figure 15 processing 516).

[0203] Then, the preprocessing unit 424 uses the time-series data during the period from the past that is traced back by a specified time T5 from the current time to the current time for the differential gradient Sdiff, the open-circuit voltage Voc, the first gradient change rate R1, and the second gradient change rate R2 respectively as the state estimation input data of the state estimation model 432 ( Figure 15 processing 518).

[0204] In addition, the specific calculation methods of the current difference δIte, the voltage difference δVte, the differential gradient Sdiff, the open-circuit voltage change amount Dvoc, the cumulative current value ΣIte, the differential gradient change amount Dsd, the first gradient change rate R1, and the second gradient change rate R2 in the preprocessing unit 424, and the structure of the state estimation input data are the same as the calculation methods and structures described in the first embodiment.

[0205] The state estimation unit 426 executes Figure 13Step S304 shown. Specifically, the state estimation unit 426 uses the state estimation input data calculated by the preprocessing unit 424, and estimates the SOC and SOH as the current state of the secondary battery 404 through the learned state estimation model 432. The state estimation unit 426 outputs the estimated current SOC and SOH values as the SOC estimation value and the SOH estimation value to the driving control device 414.

[0206] In addition, the present invention is not limited to the structure of the above-described embodiments, and can be implemented in various ways without departing from its gist.

[0207] For example, the machine learning device 100 in the above-described first embodiment and the state estimation device 400 in the second embodiment use the learned open-circuit voltage estimation model to obtain the open-circuit voltage of the secondary battery required for learning or state estimation of the state estimation model. However, when obtaining the open-circuit voltage of the working secondary battery in the first and second embodiments, it is not necessary to use the open-circuit voltage estimation model.

[0208] Alternatively, for example, according to the prior art, an alternating current as a measurement signal is input to the secondary battery 102 to measure the internal impedance Z, and the open-circuit voltage of the working secondary battery is calculated based on the measured internal impedance Z, the current terminal current, the terminal voltage, and the impedance of the load. Or, for example, the current internal impedance Z can be calculated based on the temperature dependence of the typical internal impedance Z of the secondary battery and the current temperature, and the open-circuit voltage of the working secondary battery can be calculated based on the calculated internal impedance Z, the current terminal current, the terminal voltage, and the impedance of the load.

[0209] Furthermore, in the method for estimating the state of the working secondary battery in the present embodiment, the state estimation model is learned in a manner of estimating both the SOC and SOH of the secondary battery. However, the state estimation model can also be learned in a manner of estimating only one of the SOC and SOH.

[0210] Furthermore, in the first embodiment, the time series data of the differential gradient Sdiff, the open-circuit voltage Voc, the first gradient change rate R1, and the second gradient change rate R2 are used as the state estimation input data input to the state estimation model 126. However, only the time series data of the differential gradient Sdiff can also be used as the state estimation input data.

[0211] However, in order to accurately estimate the SOC and SOH for a wider range of secondary batteries with different manufacturers and models, it is preferable to also use the time series data of three other input variables (open-circuit voltage Voc, first gradient change rate R1, and second gradient change rate R2) to enable learning of the differences in SOC vs OCV characteristics, etc. between different secondary batteries such as different models.

[0212] In addition, time series data of the terminal current Ite during the period from the past traced back a specified time T5 to the present can be appended to the state estimation input data input to the state estimation model 126. Thereby, the accuracy of state estimation of the secondary battery by the state estimation model can be further improved. Here, the time series data of the terminal current Ite is represented by the following formula.

[0213] V Ite(t n ) = (Ite(t s ), Ite(t s+1 ), Ite(t s+2 ),...., Ite(t n ))

[0214] In addition, in the present embodiment, the current difference δIte and the voltage difference δVte are respectively the fourth-order differences Δ 4 Ite of the terminal current and the fourth-order difference Δ 4 Vte of the terminal voltage. However, the current difference δIte and the voltage difference δVte do not necessarily have to be fourth-order differences. For example, even if the current difference δIte and the voltage difference δVte are the first-order differences Δ 1 Ite and Δ 1 Vte, the state estimation model 126 can learn the relationship between the operation of the change (gradient) of the terminal current with respect to the terminal voltage and the SOC and SOH. However, higher-order differences above the fourth order can extract more common change patterns of the terminal current and the terminal voltage between secondary batteries with different electrical characteristics. Therefore, from the perspective of more accurately estimating the SOC and SOH for secondary batteries with different manufacturers or models, it is preferable.

[0215] In addition, time series data of the temperature of the secondary battery 102 can be appended to the input data of each of the open-circuit voltage estimation model 124 and the state estimation model 126. Thereby, the estimation accuracy of the SOC and SOH can be further improved.

[0216] In addition, in the above-described embodiment, the open-circuit voltage estimation model 124 and the state estimation model 126 are RNNs that are easy to process data continuous in time series as inputs. However, the structures of the open-circuit voltage estimation model and the state estimation model are not limited to RNNs.

[0217] For example, the open-circuit voltage estimation model 124 and the state estimation model 126 can also both be constituted by a one-dimensional CNN (Convolutional Neural Network). In this case, the voltage estimation input data and the state estimation input data (Equations (1) and (2)) represented by two-dimensional tensors can be respectively input into the open-circuit voltage estimation model 124 and the state estimation model 126.

[0218] In addition, in the above-described embodiment, as an example of the device for performing the step S304 of estimating the state of the secondary battery during the estimation operation, a state estimation device 400 for estimating the state of the secondary battery 404 during the estimation operation mounted on the vehicle 402 is shown. However, the step S304 of estimating the state of the secondary battery during the estimation operation is not limited to the secondary battery for vehicles, and can be used for state estimation of secondary batteries for any applications such as mobile phones, bicycles, and households.

[0219] In addition, in the above-described embodiment, the state estimation device 400 is implemented as a single centralized device that only performs state estimation. However, this is only an example, and the step S304 of estimating the state of the secondary battery during the estimation operation can be executed in other devices having functions other than the state estimation of the secondary battery. For example, the step S304 of estimating the state of the secondary battery during the estimation operation can be executed in a controller that controls the load of the secondary battery. As a specific example, for example, in Figure 14 , the state observation unit 422, the preprocessing unit 424, and the state estimation unit 426 included in the processing device 420 of the state estimation device 400 can also be implemented in the processing device 440 of the travel control device 414. In this case, the open-circuit voltage estimation model 430 and the state estimation model 432 stored in the storage device 428 are stored in the storage device 448 of the travel control device 414.

[0220] As described above, the learning method of the state estimation model of the secondary battery according to the first embodiment includes the following step S100: Measuring state variables including the terminal current Ite and the terminal voltage Vte of the working secondary battery 102 connected to the load 106 or the charger 104 at a prescribed time interval dt. In addition, this learning method includes the following steps: a step S102 of preprocessing the state variables to calculate state estimation input data; and a step S104 of making the state estimation model 126 learn the relationship between the state estimation input data and the charge rate SOC and / or the degree of deterioration SOH of the working secondary battery 102 through machine learning. Moreover, in the step S102 of calculating the state estimation input data, based on the time series data of the terminal current Ite and the time series data of the terminal voltage Vte, the difference in the terminal current Ite, i.e., the current difference δIte, and the difference in the terminal voltage Vte, i.e., the voltage difference δVte, are calculated. In addition, in the step S102, based on the time series data of the current difference δIte and the time series data of the voltage difference δVte, the rate of change of the voltage difference δVte with respect to the current difference δIte, i.e., the differential gradient Sdiff, within the period from the past traced back by the first prescribed time T1 to the current time is calculated. Moreover, in the step S102, state estimation input data including the time series data of the differential gradient Sdiff is generated. V Sdiff(t n ) of the state estimation input data V x2(t n ).

[0221] According to this configuration, a state estimation model can be generated that can accurately estimate the state of the charge rate and / or the degree of deterioration of these secondary batteries during the operation of secondary batteries with various electrical characteristics of different manufacturers or models.

[0222] In addition, the state estimation input data V x2(t n ) also includes the time series data of the open circuit voltage Voc of the working secondary battery 404 V Voc(t n ), the time series data of the first gradient change rate R1 V R1(t n ), and the time series data of the second gradient change rate R2 V R2(t n ).

[0223] Here, the first gradient change rate R1 is calculated based on the time-series data of the cumulative current value ΣIte, which is the sum of the values of the terminal current Ite continuously obtained during the period from the past that is the second specified time T3 back from the current to the current, and the time-series data of the differential gradient change amount Dsd obtained by subtracting the differential gradient Sdiff from the past that is the second specified time T3 back from the current from the current differential gradient Sdiff. The first gradient change rate R1 is the change rate of the differential gradient change amount Dsd with respect to the cumulative current value ΣIte during the period from the past that is the third specified time T4 back from the current to the current.

[0224] In addition, the second gradient change rate R2 is calculated based on the time-series data of the open-circuit voltage change amount Dvoc obtained by subtracting the open-circuit voltage Voc from the past that is the second specified time T3 back from the current from the current open-circuit voltage Voc, and the time-series data of the differential gradient change amount Dsd. The second gradient change rate R2 is the change rate of the differential gradient change amount Dsd with respect to the open-circuit voltage change amount Dvoc during the period from the past that is the third specified time T4 back from the current to the current.

[0225] With this structure, the estimation accuracy of the charging rate and / or the degradation degree of the secondary battery can be improved.

[0226] In addition, the state estimation input data V x2(t n ) also includes the time-series data of the terminal current Ite V Ite(t n ). With this structure, the estimation accuracy of the charging rate and / or the degradation degree of the secondary battery estimated by the generated state estimation model can be improved.

[0227] In addition, the differential gradient Sdiff, the first gradient change rate R1, and the second gradient change rate R2 are calculated using the least squares method. With this structure, it is possible to prevent a decrease in the estimation accuracy of the SOC and / or SOH due to measurement errors of the state variables.

[0228] In addition, in the step S102 of calculating the state estimation input data, the time-series data of the terminal current Ite V Ite(t n ), the time-series data of the terminal voltage Vte V Vte(t n ), and the time-series data of the differential gradient Sdiff V Sdiff(t n ) are used as the voltage estimation input data to estimate the open-circuit voltage Voc of the working secondary battery 404. Moreover, in the step S102, the above-estimated open-circuit voltage Voc is used to calculate the state estimation input data Vx2(t n )。

[0229] According to this structure, it is possible to accurately estimate the open-circuit voltage of secondary batteries with various electrical characteristics from different manufacturers and models, and further improve the estimation accuracy of the charging rate and / or degradation degree of the secondary batteries in the state estimation model.

[0230] In addition, in step S102 of calculating the state estimation input data, the open-circuit voltage estimated by the learned open-circuit voltage estimation model 430 is used to calculate the state estimation input data, where the open-circuit voltage estimation model 430 has learned the relationship between the above voltage estimation input data and the open-circuit voltage of the working secondary battery. According to this structure, it is possible to more accurately estimate the open-circuit voltage of secondary batteries with various electrical characteristics from different manufacturers or models, and further improve the estimation accuracy of the charging rate and / or degradation degree of the secondary batteries in the state estimation model.

[0231] In addition, the differential gradient Sdiff is calculated using the least squares method. According to this structure, in the generated state estimation model, it is possible to prevent a decrease in the estimation accuracy of SOC and / or SOH caused by measurement errors of state variables.

[0232] In addition, the current difference δIte and the voltage difference δVte are respectively the fourth-order differences Δ 4 Ite of the time series data of the terminal current Ite and the fourth-order difference Δ 4 Vte of the time series data of the terminal voltage Vte. According to this structure, through the generated state estimation model, using the high-order change patterns of the terminal current Ite and the terminal voltage Vte that can be more commonly possessed by secondary batteries with different electrical characteristics, it is possible to more accurately estimate the SOC and / or SOH of secondary batteries with various electrical characteristics from different manufacturers or models.

[0233] In addition, the state estimation model 126 is composed of an RNN or a one-dimensional CNN. In addition, the intermediate layer of the RNN constituting the state estimation model 126 can be composed of an LSTM or a GRU. According to this structure, it is possible to efficiently process the time series data of multiple variables and perform effective learning for the state estimation model.

[0234] In addition, the state estimation model 126 is generated by learning using the time series data of state variables including the terminal current Ite and the terminal voltage Vte of each of the multiple secondary batteries 102 with different electrical characteristics connected to the load 106 or the charger 104. According to this structure, a state estimation model can be generated that can accurately estimate the charging rate and / or degradation degree state of secondary batteries with various electrical characteristics from different manufacturers or models.

[0235] In addition, the method for estimating the state of the secondary battery according to the second embodiment includes the following step S300: measuring state variables including the terminal current Ite and the terminal voltage Vte of the working secondary battery 102 connected to the load 106 or the charger 104 at a prescribed time interval dt. Further, this learning method includes: a step S302 of preprocessing the state variables to calculate state estimation input data; and a step S304 of estimating the current charge rate and / or degree of deterioration of the working secondary battery 102 based on the state estimation input data using the learned state estimation model 432 based on the learning method of the first embodiment. Moreover, in the step S302 of calculating the state estimation input data, based on the time series data of the terminal current Ite and the time series data of the terminal voltage Vte, the difference of the terminal current Ite, i.e., the current difference δIte, and the difference of the terminal voltage Vte, i.e., the voltage difference δVte ( Figure 15 processing 500). Further, in the step S302, based on the time series data of the current difference δIte and the time series data of the voltage difference δVte, the change rate of the voltage difference δVte with respect to the current difference δIte, i.e., the difference gradient Sdiff, within the period from the past of the first prescribed time T1 traced back from the current time to the current time is calculated ( Figure 15 processing 502). Then, in the step S302, state estimation input data including the time series data of the difference gradient Sdiff ( V Sdiff(t n ) is generated ( V x2(t n )) ( Figure 15 processing 518).

[0236] In addition, the method for estimating the state of the secondary battery according to the second embodiment is executed, for example, by a state estimation device 400. The state estimation device 400 includes a state observation unit 422 that measures state variables including the terminal current Ite and the terminal voltage Vte of the working secondary battery 404 at a prescribed time interval dt. Further, the state estimation device 400 includes a preprocessing unit 424 that preprocesses the state variables measured by the state observation unit 422 to calculate state estimation input data. And the state estimation device 400 includes a state estimation unit 426 that estimates the current charge rate and / or degree of deterioration of the working secondary battery 102 based on the state estimation input data using the learned state estimation model 432 based on the learning method of the first embodiment.

[0237] Moreover, the preprocessing unit 424 calculates the difference in terminal current, i.e., current difference δIte, and the difference in terminal voltage, i.e., voltage difference δVte, based on the time-series data of the terminal current Ite and the time-series data of the terminal voltage Vte obtained by the state observation unit 422. In addition, the preprocessing unit 424 calculates the rate of change of the voltage difference δVte with respect to the current difference δIte, i.e., differential gradient Sdiff, during the period from the past traced back by the first specified time T1 from the current time to the current time, based on the time-series data of the current difference δIte and the time-series data of the voltage difference δVte. Moreover, the preprocessing unit 424 generates time-series data including the differential gradient Sdiff. V Sdiff(t n ) of the state estimation input data V x2(t n ).

[0238] With these configurations, it is possible to accurately estimate the state of the charging rate and / or degradation degree of secondary batteries with various electrical characteristics, which may differ in manufacturer or model, during operation.

[0239] [Configuration supported by the above embodiment]

[0240] The above embodiment supports the following configuration.

[0241] (Configuration 1) A learning method for a state estimation model of a secondary battery, which is a machine learning-based learning method for a state estimation model that estimates the charging rate and / or degradation degree of a secondary battery during operation while connected to a load or a charger. The learning method for the state estimation model of the secondary battery includes the following steps: measuring state variables including terminal current and terminal voltage of the secondary battery during operation at a specified time interval; preprocessing the state variables to calculate state estimation input data; and causing the state estimation model to learn the relationship between the state estimation input data and the charging rate and / or degradation degree of the secondary battery during operation through machine learning. In the step of performing the calculation, based on the time-series data of the terminal current and the time-series data of the terminal voltage, the difference in the terminal current, i.e., current difference, and the difference in the terminal voltage, i.e., voltage difference, are calculated. Based on the time-series data of the current difference and the time-series data of the voltage difference, the rate of change of the voltage difference with respect to the current difference, i.e., differential gradient, during the period from the past traced back by the first specified time from the current time to the current time is calculated, and the state estimation input data including the time-series data of the differential gradient is generated.

[0242] (Structure 2) A learning method for a state estimation model of a secondary battery according to Structure 1, wherein the state estimation input data further includes time series data of the open-circuit voltage of the secondary battery in operation, time series data of a first gradient change rate, and time series data of a second gradient change rate. The first gradient change rate is calculated based on time series data of the cumulative current value, which is the sum of the terminal current values continuously obtained during the period from the past traced back by a second specified time to the current time, and time series data of the differential gradient change amount obtained by subtracting the differential gradient from the past traced back by the second specified time from the current differential gradient. It is the change rate of the differential gradient change amount with respect to the cumulative current value during the period from the past traced back by a third specified time to the current time. The second gradient change rate is calculated based on time series data of the open-circuit voltage change amount obtained by subtracting the open-circuit voltage from the past traced back by the second specified time from the current open-circuit voltage, and time series data of the differential gradient change amount. It is the change rate of the differential gradient change amount with respect to the open-circuit voltage change amount during the period from the past traced back by a third specified time to the current time.

[0243] (Structure 3) A learning method for a state estimation model of a secondary battery according to Structure 2, wherein the first gradient change rate and the second gradient change rate are calculated using the least squares method.

[0244] (Structure 4) A learning method for a state estimation model of a secondary battery according to any one of Structures 1 to 3, wherein in the step of performing the calculation, time series data of the terminal current, time series data of the terminal voltage, and time series data of the differential gradient are used as voltage estimation input data to estimate the open-circuit voltage of the secondary battery in operation, and the state estimation input data is calculated using the estimated open-circuit voltage.

[0245] (Structure 5) A learning method for a state estimation model of a secondary battery according to Structure 4, wherein in the step of performing the calculation, the state estimation input data is calculated using the open-circuit voltage estimated by using a learned open-circuit voltage estimation model, and the learned open-circuit voltage estimation model has learned the relationship between the voltage estimation input data and the open-circuit voltage of the secondary battery in operation.

[0246] (Structure 6) A learning method for a state estimation model of a secondary battery according to any one of Structures 1 to 5, wherein the differential gradient is calculated using the least squares method.

[0247] (Structure 7) A learning method for a state estimation model of a secondary battery according to any one of Structures 1 to 6, wherein the current difference and the voltage difference are respectively the fourth-order differences of the time-series data of the terminal current and the fourth-order differences of the time-series data of the terminal voltage.

[0248] (Structure 8) A learning method for a state estimation model of a secondary battery according to any one of Structures 1 to 7, wherein the state estimation model is constituted by an RNN (Recurrent Neural Network).

[0249] (Structure 9) A learning method for a state estimation model of a secondary battery according to Structure 8, wherein the intermediate layer of the RNN constituting the state estimation model is constituted by an LSTM (Long Short-Term Memory) or a GRU (Gated Recurrent Unit).

[0250] (Structure 10) A learning method for a state estimation model of a secondary battery according to any one of Structures 1 to 8, wherein the state estimation model is constituted by a one-dimensional CNN (Convolutional Neural Network).

[0251] (Structure 11) A learning method for a state estimation model of a secondary battery according to any one of Structures 1 to 10, wherein the state estimation model is generated by learning using time-series data of state variables including terminal current and terminal voltage for each of a plurality of secondary batteries having different electrical characteristics connected to a load or a charger.

[0252] (Structure 12) A state estimation method for a secondary battery, wherein the state estimation method for the secondary battery includes the following steps: measuring state variables including terminal current and terminal voltage of the working secondary battery connected to a load or a charger at a prescribed time interval; preprocessing the state variables to calculate state estimation input data; and using the learned state estimation model obtained by the learning method for the state estimation model of the secondary battery according to any one of Structures 1 to 11, estimating the current charge rate and / or degree of deterioration of the working secondary battery based on the state estimation input data. In the step of performing the calculation, based on the time-series data of the terminal current and the time-series data of the terminal voltage, calculate the difference of the terminal current, i.e., the current difference, and the difference of the terminal voltage, i.e., the voltage difference, and based on the time-series data of the current difference and the time-series data of the voltage difference, calculate the rate of change of the voltage difference with respect to the current difference, i.e., the differential gradient, from the past up to the current time traced back by a first prescribed time, and generate the state estimation input data including the time-series data of the differential gradient.

[0253] (Structure 13) A state estimation device for a secondary battery, the state estimation device for the secondary battery having: a state observation unit that measures state variables including a terminal current and a terminal voltage of the working secondary battery at a prescribed time interval; a preprocessing unit that preprocesses the state variables measured by the state observation unit to calculate state estimation input data; and a state estimation unit that estimates a current charge rate and / or degree of deterioration of the working secondary battery based on the state estimation input data using a learned state estimation model obtained by a learning method of the state estimation model of the secondary battery according to any one of Structures 1 to 11, wherein the preprocessing unit performs the following processing: calculates a difference in the terminal current, i.e., a current difference, and a difference in the terminal voltage, i.e., a voltage difference, based on the time series data of the terminal current and the time series data of the terminal voltage obtained by the state observation unit, calculates a rate of change of the voltage difference with respect to the current difference, i.e., a differential gradient, within a period from a past that is a first prescribed time back from the present to the present based on the time series data of the current difference and the time series data of the voltage difference, and generates the state estimation input data including the time series data of the differential gradient.

Claims

1. A learning method for a state estimation model of a secondary battery, which is a machine learning-based learning method for a state estimation model that estimates the charging rate and / or degree of deterioration of a working secondary battery connected to a load or a charger, wherein, The learning method includes the following steps: Measuring state variables including terminal current and terminal voltage of the secondary battery in the operation at a specified time interval; Preprocessing the state variables to calculate state estimation input data; and Making the state estimation model learn the relationship between the state estimation input data and the charging rate and / or degradation degree of the secondary battery in the operation through machine learning, In the step of performing the calculation, Calculating the difference of the terminal current, i.e., current difference, and the difference of the terminal voltage, i.e., voltage difference, according to the time series data of the terminal current and the time series data of the terminal voltage, Calculating the rate of change of the voltage difference with respect to the current difference, i.e., differential gradient, within the period from the past of the first specified time traced back from the current time to the current time based on the time series data of the current difference and the time series data of the voltage difference, Generating the state estimation input data including the time series data of the differential gradient.

2. The learning method for a state estimation model of a secondary battery according to claim 1, wherein, The state estimation input data further includes the time series data of the open-circuit voltage of the secondary battery in the operation, the time series data of the first gradient change rate, and the time series data of the second gradient change rate, The first gradient change rate is calculated based on the time series data of the cumulative current value, which is the sum of the values of the terminal current continuously obtained within the period from the past of the second specified time traced back from the current time to the current time, and the time series data of the differential gradient change amount obtained by subtracting the differential gradient from the past of the second specified time traced back from the current time from the current differential gradient, and is the rate of change of the differential gradient change amount with respect to the cumulative current value within the period from the past of the third specified time traced back from the current time to the current time, The second gradient change rate is calculated based on the time series data of the open-circuit voltage change amount obtained by subtracting the open-circuit voltage from the past of the second specified time traced back from the current time from the current open-circuit voltage, and the time series data of the differential gradient change amount, and is the rate of change of the differential gradient change amount with respect to the open-circuit voltage change amount within the period from the past of the third specified time traced back from the current time to the current time.

3. The learning method for a state estimation model of a secondary battery according to claim 2, wherein, The first gradient change rate and the second gradient change rate are calculated using the least squares method.

4. The learning method for a state estimation model of a secondary battery according to claim 1, wherein, In the step of performing the calculation, Using the time series data of the terminal current, the time series data of the terminal voltage, and the time series data of the differential gradient as voltage estimation input data to estimate the open-circuit voltage of the secondary battery in the operation, Calculating the state estimation input data using the estimated open-circuit voltage.

5. The learning method for a state estimation model of a secondary battery according to claim 4, wherein, In the step of performing the calculation, the state estimation input data is calculated using the open-circuit voltage estimated by the learned open-circuit voltage estimation model, where the learned open-circuit voltage estimation model has learned the relationship between the voltage estimation input data and the open-circuit voltage of the secondary battery in the operation.

6. The learning method for a state estimation model of a secondary battery according to claim 1, wherein, The differential gradient is calculated using the least squares method.

7. The learning method for a state estimation model of a secondary battery according to claim 1, wherein, The current difference and the voltage difference are respectively the fourth-order differences of the time-series data of the terminal current and the fourth-order differences of the time-series data of the terminal voltage.

8. The learning method for a state estimation model of a secondary battery according to claim 1, wherein, The state estimation model is composed of an RNN (Recurrent Neural Network).

9. The learning method for a state estimation model of a secondary battery according to claim 8, wherein, The middle layer of the RNN constituting the state estimation model is composed of an LSTM (Long Short Term Memory) or a GRU (Gated Recurrent Unit).

10. The learning method of the state estimation model of the secondary battery according to claim 1, wherein, The state estimation model is composed of a one-dimensional CNN (Convolutional Neural Network).

11. The learning method of the state estimation model of the secondary battery according to claim 1, wherein, The state estimation model is generated by learning using time-series data of state variables including terminal current and terminal voltage for each of a plurality of secondary batteries with different electrical characteristics to which a load or a charger is connected.

12. A state estimation method of a secondary battery, wherein, The method for estimating the state of the secondary battery includes the following steps: Measuring state variables including terminal current and terminal voltage of the working secondary battery to which a load or a charger is connected at a prescribed time interval; Preprocessing the state variables to calculate state estimation input data; And Using the learned state estimation model obtained by the learning method of the state estimation model of the secondary battery according to any one of claims 1 to 11, estimating the current charge rate and / or degree of deterioration of the working secondary battery based on the state estimation input data, In the step of performing the calculation, Calculating a difference in the terminal current, i.e., a current difference, and a difference in the terminal voltage, i.e., a voltage difference, based on the time-series data of the terminal current and the time-series data of the terminal voltage, Calculating a difference gradient, which is a change rate of the voltage difference with respect to the current difference, within a period from the past traced back by a first prescribed time to the present based on the time-series data of the current difference and the time-series data of the voltage difference, Generating the state estimation input data including the time-series data of the difference gradient.

13. A state estimation device of a secondary battery, the state estimation device of the secondary battery having: A state observation unit that measures state variables including terminal current and terminal voltage of the working secondary battery at a prescribed time interval; A preprocessing unit that preprocesses the state variables measured by the state observation unit to calculate state estimation input data; And A state estimation unit that uses the learned state estimation model obtained by the learning method of the state estimation model of the secondary battery according to any one of claims 1 to 11, and estimates the current charge rate and / or degree of deterioration of the working secondary battery based on the state estimation input data, Wherein the preprocessing unit performs the following processing: Calculating a difference in the terminal current, i.e., a current difference, and a difference in the terminal voltage, i.e., a voltage difference, based on the time-series data of the terminal current and the time-series data of the terminal voltage obtained by the state observation unit, Based on the time series data of the current difference and the time series data of the voltage difference, calculate the rate of change of the voltage difference with respect to the current difference, that is, the differential gradient, within the period from the past at the first specified time traced back from the current to the current time. Generate the state estimation input data including the time series data of the differential gradient.

Images