Learning method, state estimation method, and state estimation device for state estimation model of secondary battery

1. A learning method of a state estimation model of a first secondary battery by a first processor of a machine learning device, the learning method using machine learning, the state estimation model estimating a charge rate and/or a deterioration degree of the operating first secondary battery, the first secondary battery being connected to a load or a charger, the state estimation model being stored in a memory, the first secondary battery being connected to a characteristic measuring instrument that measures present values of state variables including terminal currents and terminal voltages of the operating first secondary battery, the machine learning device being connected to the characteristic measuring instrument, the method comprising:

a step of measuring the state variables at predetermined time intervals;

a step of calculating state estimation input data by preprocessing the state variables;

a step of training the state estimation model to learn a relationship of the state estimation input data with the charge rate and/or the deterioration degree of the operating first secondary battery, by machine learning; and

a step of suggesting replacing the operating first secondary battery based on the deterioration degree of the operating first secondary battery,

wherein the step of calculating:

uses time-series data of the terminal currents and time-series data of the terminal voltages to calculate current differences and voltage differences, each current difference being a difference in the terminal currents at each predetermined time interval, each voltage difference being a difference in the terminal voltages at each predetermined time interval;

uses time-series data of the current differences and time-series data of the voltage differences to calculate a difference gradient, the difference gradient being a change rate of the voltage differences with respect to the current differences as a slope of an approximate straight line when on a two-dimensional plane comprising the current difference and the voltage difference, points are plotted at coordinates each of which is determined by the current difference and the voltage difference at each predetermined time interval in a period to a present from a past that goes back a first predetermined time from the present; and

generates the state estimation input data including time-series data of the difference gradients

wherein the state estimation input data further includes

time-series data of open circuit voltages of the operating first secondary battery,

time-series data of first gradient change rates, and

time-series data of second gradient change rates,

each first gradient change rate being determined as follows:

an integrated current value is determined to be a sum of the terminal current values acquired continuously in a period to a present from a past that goes back a second predetermined time from the present;

a difference gradient change amount is determined by subtracting the difference gradient in the past that goes back the second predetermined time from the present, from the present difference gradient;

time-series data of the integrated current values and time-series data of the difference gradient change amounts determines a change rate of the difference gradient change amounts with respect to the integrated current values; and

the first gradient change rate is determined to be the change rate of the difference gradient change amounts with respect to the integrated current values, in a period to the present from a past that goes back a third predetermined time from the present, and

each second gradient change rate being determined as follows:

an open circuit voltage change amount is determined by subtracting the open circuit voltage in the past that goes back the second predetermined time from the present, from the present open circuit voltage;

time-series data of the open circuit voltage change amounts and time-series data of the difference gradient change amounts determines a change rate of the difference gradient change amounts with respect to the open circuit voltage change amounts; and

the second gradient change rate is determined to be the change rate of the difference gradient change amounts with respect to the open circuit voltage change amounts, in the period to the present from the past that goes back the third predetermined time from the present.

2. The learning method of the state estimation model of the first secondary battery according to claim 1 , wherein

each first gradient change rate and each second gradient change rate are calculated using the least squares method.

3. The learning method of the state estimation model of the first secondary battery according to claim 1 , wherein

the step of calculating:

uses time-series data of the terminal currents and time-series data of the terminal voltages, and timeseries data of the difference gradients, as voltage estimation input data, to estimate an open circuit voltage of the operating first secondary battery; and

uses the estimated open circuit voltage to calculate the state estimation input data.

4. The learning method of the state estimation model of the first secondary battery according to claim 3 , wherein

the step of calculating uses the open circuit voltage estimated using a trained open circuit voltage estimation model, and thereby calculates the state estimation input data, the trained open circuit voltage estimation model having learned a relationship between the voltage estimation input data and an open circuit voltage of the operating first secondary battery.

5. The learning method of the state estimation model of the first secondary battery according to claim 1 , wherein

the difference gradient is calculated using the least squares method.

6. The learning method of the state estimation model of the first secondary battery according to claim 1 , wherein

each current difference and each voltage difference are respectively a fourth-order difference of time-series data of the terminal currents and a fourth-order difference of time-series data of the terminal voltages.

7. The learning method of the state estimation model of the first secondary battery according to claim 1 , wherein

the state estimation model is configured of an RNN (Recurrent Neural Network).

8. The learning method of the state estimation model of the first secondary battery according to claim 7 , wherein

an intermediate layer of the RNN configuring the state estimation model is configured of an LSTM (Long Short Term Memory) or a GRU (Gated Recurrent Unit).

9. The learning method of the state estimation model of the first secondary battery according to claim 1 , wherein

the state estimation model is configured of a one-dimensional CNN (Convolutional Neural Network).

10. The learning method of the state estimation model of the first secondary battery according to claim 1 , wherein

the state estimation model is generated by learning using time-series data of state variables including terminal currents and terminal voltages for each of a plurality of first secondary batteries with different electrical characteristics, the first secondary batteries each being connected to a load or a charger.

11. A method of learning a state estimation model of a first secondary battery by a first processor of a machine learning device, and estimating a state of a second secondary battery by a second processor, the second secondary battery and the second processor being mounted in a vehicle powered by the secondary battery, the learning method using machine learning, the state estimation model estimating a charge rate and/or a deterioration degree of the operating first secondary battery, the first secondary battery being connected to a load or a charger, the state estimation model being stored in a memory, the first secondary battery being connected to a characteristic measuring instrument that measures present values of state variables including terminal currents and terminal voltages of the operating first secondary battery, the machine learning device being connected to the characteristic measuring instrument, the method comprising:

a step of measuring the state variables at predetermined time intervals, the state variables including terminal currents and terminal voltages of the operating second secondary battery, the second secondary battery being connected to a load or a charger of the vehicle;

a step of calculating state estimation input data by preprocessing the state variables;

a step of training the state estimation model to learn a relationship of the state estimation input data with the charge rate and/or the deterioration degree of the operating first secondary battery, by machine learning;

a step of estimating a present charge rate and/or a present deterioration degree of the operating second secondary battery, from the state estimation input data, using the state estimation model trained by a learning method of the state estimation model of the first secondary battery; and

a step suggesting replacing the second secondary battery in the vehicle based on the present deterioration degree of the operating second secondary battery,

wherein the step of calculating:

uses time-series data of the terminal currents and time-series data of the terminal voltages to calculate current differences and voltage differences, each current difference being a difference in the terminal currents at each predetermined time interval, each voltage difference being a difference in the terminal voltages at each predetermined time interval;

uses time-series data of the current differences and time-series data of the voltage differences to calculate a difference gradient, the difference gradient being a change rate of the voltage differences with respect to the current differences in a period to a present from a past that goes back a first predetermined time from the present, as a slope of an approximate straight line when on a two-dimensional plane comprising the current difference and the voltage difference, points are plotted at coordinates each of which is determined by the current difference and the voltage difference at each predetermined time interval in a period to a present from a past that goes back a first predetermined time from the present; and

generates the state estimation input data including time-series data of the difference gradients.

12. A state estimation device of a second secondary battery, the device comprising a second processor, the second secondary battery and the second processor being mounted in a vehicle powered by the second secondary battery,

wherein the second processor is configured to:

measure state variables at predetermined time intervals, the state variables including terminal currents and terminal voltages of an operating second secondary battery;

perform preprocessing of the measured state variables, to calculate state estimation input data;

estimate a present charge rate and/or a present deterioration degree of the operating second secondary battery, from the state estimation input data, using a state estimation model trained by a learning method of a state estimation model of the first secondary battery; and

suggest replacing the second secondary battery in the vehicle based on the present deterioration degree of the operating second secondary battery,

wherein in the preprocessing, the second processor:

uses time-series data of the measured terminal currents and time-series data of the measured terminal voltages, to calculate current differences and voltage differences, each current difference being a difference in the terminal currents, each voltage difference being a difference in the terminal voltages;

uses time-series data of the current differences and time-series data of the voltage differences to calculate a difference gradient, the difference gradient being a change rate of the voltage differences with respect to the current differences in a period to a present from a past that goes back a first predetermined time from the present; and

generates the state estimation input data including time-series data of the difference gradients, and

wherein the state estimation model is trained by a learning method of the state estimation model of the first secondary battery by a first processor of a machine learning device, the learning method using machine learning, the state estimation model estimating a charge rate and/or a deterioration degree of the operating first secondary battery, the first secondary battery being connected to a load or a charger, the state estimation model being stored in a memory, the first secondary battery being connected to a characteristic measuring instrument that measures present values of state variables including terminal currents and terminal voltages of the operating first secondary battery, the machine learning device being connected to the characteristic measuring instrument, the method comprising:

a step of measuring the state variables at predetermined time intervals;

a step of calculating state estimation input data by preprocessing the state variables; and

a step of training the state estimation model to learn a relationship of the state estimation input data with the charge rate and/or the deterioration degree of the operating first secondary battery, by machine learning,

wherein the step of calculating:

uses time-series data of the terminal currents and time-series data of the terminal voltages to calculate current differences and voltage differences, each current difference being a difference in the terminal currents at each predetermined time interval, each voltage difference being a difference in the terminal voltages at each predetermined time interval;

uses time-series data of the current differences and time-series data of the voltage differences to calculate a difference gradient, the difference gradient being a change rate of the voltage differences with respect to the current differences as a slope of an approximate straight line when on a two-dimensional plane comprising the current difference and the voltage difference, points are plotted at coordinates each of which is determined by the current difference and the voltage difference at each predetermined time interval in a period to a present from a past that goes back a first predetermined time from the present; and

generates the state estimation input data including time-series data of the difference gradients.