A method for learning a lifetime prediction model, a learning program for a lifetime prediction model, and an information processing device.

Abstract

Improve the development efficiency of lithium-ion batteries. [Solution] The method for predicting the lifespan of a lithium-ion battery of the present invention is carried out by a computer using the following steps: (a) acquiring learning data including life data and cycle measurement data that records the voltage applied to each cell, the current flowing through each cell, and the current capacity by performing a charge / discharge test on each cell of the battery, including a constant current charging step followed by a constant current discharging step; and (b) learning a lifespan prediction model for one or more predicted execution cycle numbers ni (i is an integer greater than or equal to 1), using learning cycle measurement data Ani, which is an initial number ni cycles extracted from the cycle measurement data of each cell of the learning data, as an explanatory variable and the life data of the learning data as a target variable, to obtain a set of learned lifespan prediction models Cni in a number corresponding to the number of predicted execution cycles ni, and the lifespan data included in the learning data is estimated based on the change in current capacity recorded in the cycle measurement data.

Description

[Technical Field]

[0001] The present invention relates to Learning methods for life expectancy prediction models, learning programs for life expectancy prediction models , and an information processing apparatus. [Background Art]

[0002] In the research and development of lithium-ion batteries, charge-discharge cycle tests are performed and the cycle life of battery cells is measured to evaluate the materials and design constituting the battery. 1A in FIG. 1 illustrates the cycle life obtained by a charge-discharge cycle test, wherein the cycle life is obtained as the number of cycles at which the discharge capacity retention rate falls below a threshold value. Most of the time spent on battery research and development is consumed in measuring cycle life, and in order to accelerate battery development, a method that can predict cycle life at an early stage, either directly or via the discharge capacity retention rate, is required.

[0003] Among these methods, as a method for directly and early predicting the cycle life of a battery cell, a method for predicting cycle life from initial results of a charge-discharge cycle test using machine learning technology is disclosed (Patent Documents 1 and 2). Also disclosed is a method for predicting cycle life that does not require a charge-discharge cycle test and uses machine learning technology that uses, as training data, data including design factors, process factors, and formation factors determined at the time of battery cell design (Patent Document 3). [Prior Art Literature] [Patent Documents]

[0004] [Patent Document 1] Japanese National Publication of International Patent Application No. 2010-539473 [Patent Document 2] Japanese Unexamined Patent Publication No. 2019-113524 [Patent Document 3] Japanese Unexamined Patent Publication No. 2013-217897 [Summary of the Invention] [Problems to be Solved by the Invention]

[0005] In the research and development of lithium-ion batteries, methods for predicting cycle life must satisfy the following two requirements from the perspective of improving development efficiency.

[0006] Firstly, materials and designs frequently change during research and development. To avoid having to start over from preparing training data for machine learning every time such changes occur, the training data needs to be independent of the materials and design that make up the battery.

[0007] Secondly, since the cycle life of current battery cells spans several thousand cycles, and errors in such long-term predictions have a significant impact, it is desirable that the predicted cycle life be in the form of a probability distribution.

[0008] The methods disclosed in Patent Documents 1 and 2 use measurement data such as voltage, current, and charge / discharge capacity acquired at the beginning of a charge / discharge cycle test as training data, and do not include information on the materials or design of the battery, thus satisfying the first requirement mentioned above. However, the output of the neural network used in the method disclosed in Patent Document 1 is a single value, and the method disclosed in Patent Document 2 does not present a procedure for predicting cycle life in the form of a probability distribution. Therefore, neither method satisfies the second requirement mentioned above.

[0009] The method disclosed in Patent Document 3 does not satisfy the first requirement mentioned above because it uses data that depends on the materials and design of the battery as training data.

[0010] This disclosure aims to provide a predictive technology that improves the development efficiency of lithium-ion batteries. [Means for solving the problem]

[0011] The present invention comprises the following configuration.

[0012] [1] Step (a) to acquire training data including battery cycle measurement data and lifespan data, (b) A step of training a lifetime prediction model using the training data for one or more predicted execution cycles to obtain a set of trained lifetime prediction models corresponding to each predicted execution cycle, Step (c) acquires predictive cycle measurement data for the battery to be predicted, sequentially up to the predicted number of execution cycles, A method for predicting the lifespan of a lithium-ion battery, characterized in that a computer performs the following steps: (d) inputting prediction cycle measurement data obtained up to the predicted number of execution cycles into a trained lifespan prediction model corresponding to the predicted number of execution cycles, and obtaining the probability distribution of lifespan at each predicted number of execution cycles as output.

[0013] [2] The lifetime prediction model is A data formatting unit that formats the cycle measurement data into fixed-length data, A feature extraction unit compresses the aforementioned fixed-length data into compressed data, The system includes a nonlinear transformation unit that converts the compressed data into nonlinear feature data mapped to a high-dimensional space, A method for predicting the lifespan of a lithium-ion battery according to [1], characterized in that a regression unit that takes the aforementioned nonlinear feature data as input and outputs a probability distribution of lifespan is connected in this order.

[0014] [3] The method for predicting the lifespan of a lithium-ion battery according to [2], characterized in that a dimensionality reduction method is used as the data compression method in the feature extraction unit.

[0015] [4] The method for predicting the lifespan of a lithium-ion battery according to [3], characterized in that the dimensionality reduction method is principal component analysis.

[0016] [5] Step (a) of acquiring training data including battery cycle measurement data and discharge capacity retention rate in each cycle, (b) A step in which, for one or more predicted execution cycles, the training data is used to train the discharge capacity maintenance rate in each cycle, and a set of trained discharge capacity maintenance rate prediction models corresponding to each predicted execution cycle is obtained. Step (c) acquires predictive cycle measurement data for the battery to be predicted, sequentially up to the predicted number of execution cycles, A method for predicting the discharge capacity maintenance rate of a lithium-ion battery, characterized in that a computer performs the following steps: (d) inputting prediction cycle measurement data obtained up to the predicted number of execution cycles into a trained discharge capacity maintenance rate prediction model for the corresponding predicted number of execution cycles, and obtaining as output the probability distribution of the discharge capacity maintenance rate for each cycle from each predicted number of execution cycles to the end of life.

[0017] [6] Step (a) to acquire training data including battery cycle measurement data and lifespan data, (b) A step of training a lifetime prediction model using the training data for one or more predicted execution cycles to obtain a set of trained lifetime prediction models corresponding to each predicted execution cycle, Step (c) acquires predictive cycle measurement data for the battery to be predicted, sequentially up to the predicted number of execution cycles, A lifetime prediction program that causes a computer to perform the following steps: (d) inputting prediction cycle measurement data obtained up to the predicted number of execution cycles into the corresponding trained lifetime prediction model for the predicted number of execution cycles, and obtaining the probability distribution of lifetime for each predicted number of execution cycles as output.

[0018] [7] Step (a) of acquiring training data including battery cycle measurement data and discharge capacity retention rate in each cycle, (b) A step in which, for one or more predicted execution cycles, the training data is used to train the discharge capacity maintenance rate in each cycle, and a set of trained discharge capacity maintenance rate prediction models corresponding to each predicted execution cycle is obtained. Step (c) of sequentially acquiring prediction cycle measurement data of a battery that is a prediction target up to said number of prediction execution cycles, A discharge capacity retention rate prediction program for causing a computer to execute step (d) of inputting the prediction cycle measurement data acquired up to the number of prediction execution cycles into a trained discharge capacity retention rate prediction model corresponding to the respective number of prediction execution cycles, and acquiring, as an output, a probability distribution of discharge capacity retention rate in each cycle from each number of prediction execution cycles to the end of service life,

[0019] [8] Step (a) of acquiring training data including cycle measurement data and service life data of a battery, Step (b) of training a service life prediction model using said training data for one or more numbers of prediction execution cycles, and obtaining a set of trained service life prediction models respectively corresponding to each number of prediction execution cycles, Step (c) of sequentially acquiring prediction cycle measurement data of a battery that is a prediction target up to said number of prediction execution cycles, An information processing apparatus that executes step (d) of inputting the prediction cycle measurement data acquired up to the number of prediction execution cycles into a trained service life prediction model corresponding to the respective number of prediction execution cycles, and acquiring, as an output, a probability distribution of service life for each number of prediction execution cycles,

[0020] [9] Step (a) of acquiring training data including cycle measurement data of a battery and discharge capacity retention rate in each cycle, Step (b) of training the discharge capacity retention rate in each cycle using said training data for one or more numbers of prediction execution cycles, and obtaining a set of trained discharge capacity retention rate prediction models respectively corresponding to each number of prediction execution cycles, Step (c) of sequentially acquiring prediction cycle measurement data of a battery that is a prediction target up to said number of prediction execution cycles, An information processing device that performs the following steps: (d) inputs the predictive cycle measurement data obtained up to the predicted number of execution cycles into the corresponding trained discharge capacity maintenance rate prediction model for the predicted number of execution cycles, and obtains as output the probability distribution of the discharge capacity maintenance rate for each cycle from each predicted number of execution cycles to the end of life. [Effects of the Invention]

[0021] Firstly, even if changes occur in materials or design, there is no need to re-prepare the training data for machine learning. Secondly, because the cycle life or discharge capacity retention rate is in the form of a probability distribution, the impact of errors in long-term predictions can be reduced.

[0022] In other words, this disclosure provides a predictive technology that can improve the development efficiency of lithium-ion batteries. [Brief explanation of the drawing]

[0023] [Figure 1] Figure 1A is a graph showing the relationship between the discharge capacity retention rate obtained by charge-discharge cycle testing and the cycle life. Figure 1B is a graph showing the discharge capacity retention rate up to the predicted number of cycles and the probability distribution of the cycle life obtained using the combination of the cycle measurement results up to the predicted number of cycles and the present invention. [Figure 2] Figure 2 is a flowchart showing the flow of the life prediction method according to the first embodiment of the present invention. [Figure 3] Figure 3 shows the configuration of cycle measurement data and lifetime data used in the first embodiment of the present invention. [Figure 4] Figure 4 is a block diagram schematically showing the configuration of the lifetime prediction model used in the first embodiment of the present invention. [Figure 5] Figure 5 shows the hardware configuration of the information processing device that executes the lifetime prediction model used in the present invention. [Figure 6] Figure 6 is a graph plotting a selection of the cycle measurement data used in the example. [Figure 7] Figure 7A is a graph plotting the current capacity in the constant current discharge step of the charge-discharge cycle against voltage, before processing in step (4) of the data shaping unit C-1. Figure 7B is a graph plotting the current capacity obtained by resampling at equally spaced voltage sampling points against voltage. [Figure 8] Figure 8 is a graph showing the lifespan prediction results obtained in the example. [Figure 9] Figure 9 is a graph showing the discharge capacity maintenance rate up to the predicted number of cycles and the predicted discharge capacity maintenance rate beyond the predicted number of cycles. [Figure 10] Figure 10 is a flowchart showing the flow of a life prediction method according to a second embodiment of the present invention. [Figure 11] Figure 11 shows the configuration of cycle measurement data and discharge capacity maintenance rate data used in the second embodiment of the present invention. [Figure 12] Figure 12 is a block diagram schematically showing the configuration of the lifetime prediction and discharge capacity maintenance rate prediction model used in the second embodiment of the present invention. [Modes for carrying out the invention]

[0024] The embodiments will be described below with reference to the attached drawings. To facilitate understanding of the explanation, the same reference numerals are used for identical components in each drawing whenever possible, and redundant explanations are omitted.

[0025] [First Embodiment] The lithium-ion battery life prediction method of the present invention includes the steps of: (a) acquiring training data including battery cycle measurement data and life data; (b) training a life prediction model using the training data for one or more predicted execution cycle counts to obtain a set of trained life prediction models corresponding to each predicted execution cycle count; (c) sequentially acquiring prediction cycle measurement data for the battery to be predicted up to the predicted execution cycle count; and (d) inputting the prediction cycle measurement data acquired up to the predicted execution cycle count into the corresponding trained life prediction models for each predicted execution cycle count to obtain the probability distribution of life at each predicted execution cycle count as an output.

[0026] Figure 2 is a flowchart illustrating steps (a) to (d) in detail, where step (a) corresponds to step S1, step (b) to steps S31 to S32, step (c) to step S41, and step (d) to step S42. The lithium-ion battery life prediction method of the present invention will now be explained based on Figure 2.

[0027] As shown in Figure 2, the lithium-ion battery life prediction method of the present invention is performed in the following steps: Step S1: acquire training cycle measurement data A and training life data B; Step S2: determine one or more predicted execution cycle numbers ni (i=1,2,…); Step S31: extract measurement data Ani for the initial ni cycles from training cycle measurement data A; Step S32: train a life prediction model C using the extracted measurement data Ani and training life data B, and acquire a trained life prediction model Cni; Step S41: acquire prediction cycle measurement data Dni up to the predicted execution cycle number ni; Step S42: input the prediction cycle measurement data Dni into the trained life prediction model Cni, and acquire the predicted life probability distribution Eni; Step S42: perform these steps for each predicted execution cycle number ni determined in Step S2.

[0028] In step S1, charge-discharge cycle tests are performed on multiple battery cells to obtain learning cycle measurement data A and learning lifespan data B. It is desirable to use 50 or more cells for acquiring learning data. Furthermore, in order to obtain learning lifespan data B, it is desirable to continue the charge-discharge cycle tests until each cell reaches its cycle life.

[0029] Each cycle in the aforementioned charge-discharge cycle test must include a constant-current charging step followed by a constant-current discharging step. A constant-voltage charging step may be included between the constant-current charging and constant-current discharging steps. Furthermore, a pause step may be included between each charge-discharge step.

[0030] The aforementioned cycle life is the number of cycles at which the discharge capacity falls below a predetermined threshold. For example, this threshold is 80% of the discharge capacity measured in the first cycle.

[0031] Figure 3 shows the structure of cycle measurement data and lifetime data.

[0032] Cycle measurement data is obtained by performing charge-discharge cycle tests on multiple cells, and recording the sampling times t1, t2, ... and the physical properties at that time for each cycle. The sampling times should be appropriately selected to adequately track changes in physical properties and do not necessarily have to be at fixed intervals. The physical properties recorded must include the voltage applied to the cell and the current flowing through the cell, and at least one measurable physical and chemical property exhibited by the cell. Examples of physical properties to be recorded include voltage, current, and current capacity.

[0033] The lifespan data represents the cycle life of each cell from which the cycle measurement data was acquired.

[0034] If the charge-discharge cycle test is not performed until the cycle life is reached when acquiring the training cycle measurement data A and training lifetime data B in step S1, the cycle life may be estimated using the change in discharge capacity up to the measured charge-discharge cycles and a combination of a mathematical model or time series data analysis method, and the estimated cycle life may be used as training lifetime data B. Examples of mathematical models include the square root law and power law known for capacity decay in lithium-ion batteries. Examples of time series data analysis methods include the autoregressive integrated moving average (ARIMA) model, Kalman filter, and Gaussian process regression.

[0035] Stored measurement data may be used as the learning cycle measurement data A and learning lifetime data B acquired in step S1.

[0036] The number of predicted execution cycles determined in step S2 is the number of charge-discharge cycles in the charge-discharge cycle test that are repeated to accumulate measurement data until the predicted execution of the cell whose lifespan is to be predicted. Multiple predicted execution cycles may be selected and determined. Alternatively, a predetermined number of predicted execution cycles may be used, and step S2 may be omitted.

[0037] Steps S31 and S32 are repeated for one or more predicted execution cycle numbers ni (i=1,2,…) determined in step S2.

[0038] In step S31, the cycle measurement data up to the predicted number of execution cycles ni is extracted from the training cycle measurement data A obtained in step S1 and named training cycle measurement data Ani. If there are cells in training cycle measurement data A and training lifetime data B whose measurements have finished before the predicted number of execution cycles ni, the data from those cells cannot be used and is therefore excluded.

[0039] In the following step S32, the extracted training cycle measurement data Ani is used as the explanatory variable and the training lifetime data B as the dependent variable to train the lifetime prediction model C, and a trained lifetime prediction model Cni with a predicted number of execution cycles ni is obtained.

[0040] In steps S31 and S32, a trained lifetime prediction model Cni is prepared, which is trained on cycle measurement data up to the initial ni cycles corresponding to the predicted number of execution cycles ni. In other words, if multiple predicted number of execution cycles are selected, multiple lifetime prediction models Cni are trained for each initial ni cycle corresponding to the predicted number of execution cycles ni.

[0041] The initial ni cycles refer to, for example, the initial 200 cycles, meaning cycles 1 through 200. The lifetime prediction model C200 is trained using training cycle measurement data A200, which is extracted from training cycle measurement data A to represent the initial 200 cycles from cycles 1 to 200, and training lifetime data B. Once the trained lifetime prediction model C200 receives prediction cycle measurement data D200 up to the predicted number of cycles executed (200), it outputs the probability distribution E200 of the predicted lifetime.

[0042] The target variable, cycle lifetime, is always a positive value. Therefore, when training the lifetime prediction model C in step S32, it is desirable to transform the training lifetime data B into a logarithm before training, rather than training the model directly with B. This ensures that the prediction result is also a logarithm of the lifetime, guaranteeing that the predicted lifetime obtained by the inverse transformation will be positive.

[0043] As shown in Figure 4, the lifespan prediction model used in the lithium-ion battery lifespan prediction method of the present invention can be configured by connecting the data shaping unit C-1, the feature extraction unit C-2, the nonlinear transformation unit C-3, and the regression unit C-4 in this order. The lifespan prediction model receives prediction cycle measurement data up to the prediction execution cycle number ni as input and outputs a probability distribution of lifespan. Below, as an example, the elements constituting the lifespan prediction model will be explained one by one.

[0044] The data formatting unit C-1 receives cycle measurement data up to the predicted number of execution cycles ni as input and outputs fixed-length data in matrix format where the number of rows is the number of cells and the number of columns is fixed length. The fixed-length data in matrix format is generated by formatting the cycle measurement data of each cell into data with a fixed number of columns using the formatting processes (1) to (7) below, and then stacking these in the row direction.

[0045] (1) The formatted data may include physical properties at specific points in time during the charge-discharge cycle test and values ​​calculated using them. Items of values ​​to be included in the formatted data include, for example, the voltage at the start of charging in the first cycle, the voltage at the start of charging in the second cycle, ..., the voltage at the start of charging in the nith cycle, the charging time in the first cycle, the charging time in the second cycle, ..., the charging time in the nith cycle, the charging capacity in the first cycle, the charging capacity in the second cycle, ..., the charging capacity in the nith cycle, the discharge capacity in the first cycle, the discharge capacity in the second cycle, ..., the discharge capacity in the nith cycle, the initial Coulomb efficiency, the slope and intercept of the cycle number-voltage-start plot near the nith cycle, the slope and intercept of the cycle number-time plot near the nith cycle, the slope and intercept of the charge number-capacity plot near the nith cycle, and the slope and intercept of the cycle number-discharge capacity plot near the nith cycle. All of these are single values, and the length of each item is 1.

[0046] (2) The formatted data may include physical property values ​​at specific points in time for each cycle of the charge-discharge cycle test. The items of values ​​to be included in the formatted data include, for example, the charging start voltage for each cycle, the charging time for each cycle, the charging capacity for each cycle, and the discharging capacity for each cycle. In this case, the length of each item is the predicted number of cycles to be performed, ni.

[0047] (3) The data for the constant current charging step of the charge-discharge cycle is processed as a function of voltage. If there are any physical properties that show a significant change in response to the voltage change, resampling is performed at a predetermined number of voltage resampling points. By performing this process for all cycles, each physical property is converted into data with a length of the predicted number of execution cycles ni × the number of voltage resampling points. Current capacity is an example of a physical property that shows a significant change in response to the voltage change.

[0048] (4) The data from the constant current discharge step of the charge-discharge cycle is processed in the same way as the data from the constant current charge step.

[0049] (5) If the charge-discharge cycle includes a constant voltage charging step, the data is processed as a function of current. If there are physical properties that show a significant change in response to the change in current, resampling is performed at a predetermined number of current resampling points. By performing this process for all cycles, each physical property is converted into data with a length of the predicted number of execution cycles ni × the number of current resampling points. Examples of physical properties that show a significant change in response to the change in current include current capacity and time.

[0050] (6) If the charge / discharge cycle includes a pause step, delete that data.

[0051] (7) The values ​​of the items obtained in steps (1) to (6) above are concatenated in the column direction to generate one row of fixed-length data in matrix format.

[0052] The resampling method performed in the data shaping unit C-1 can be the simplest linear interpolation, or it may be a polynomial interpolation of second order or higher, or a spline interpolation, etc.

[0053] The feature extraction unit C-2 receives fixed-length data output from the data formatting unit C-1 as input, compresses the fixed-length data using a data compression method, and outputs it as compressed data. Dimensionality reduction methods are examples of data compression methods, and principal component analysis is preferred as the dimensionality reduction method. When using principal component analysis, it is preferable to compress the data for each item linked in the column direction within the matrix-format data. However, items with a length of 1 in the column direction are left as they are without data compression. Principal component analysis extracts principal components from fixed-length data. The number of principal components to be extracted must be predetermined for each item. It is preferable to select a sufficiently large number of principal components so that the cumulative contribution rate of the extracted principal components is 90% or more.

[0054] The nonlinear transformation unit C-3 receives the compressed data output from the feature extraction unit C-2 as input, applies a nonlinear transformation, and outputs it as nonlinear feature data mapped to a high-dimensional space. The subsequent regression unit C-4 receives the nonlinear feature data as input and outputs it as a probability distribution of lifetimes.

[0055] The methods used in the nonlinear transformation unit C-3 and the regression unit C-4 may be set individually, but it is more preferable to apply Gaussian process regression, which has both functions, to C-3 and C-4. If set individually, it is preferable to use the kernel method in the nonlinear transformation unit C-3. The regression unit C-4 must output a probability distribution, and methods other than Gaussian process regression, such as Bayesian ridge, can be used.

[0056] When applying Gaussian process regression to the nonlinear transformation unit C-3 and regression unit C-4, and when applying a kernel method to the nonlinear transformation unit C-3, a nonlinear kernel such as a radial basis function kernel is used to ensure that the transformation is nonlinear.

[0057] By repeating steps S31 and S32, a set of trained lifetime prediction models Cni corresponding to one or more predicted execution cycle numbers ni (i=1,2,…) is obtained. For one or more of these ni and Cni, steps S41 and S42 are repeated as follows.

[0058] In step S41, charge-discharge cycle tests are performed on the battery cells whose lifespan is to be predicted up to the predicted number of cycles ni, and predictive cycle measurement data Dni is obtained. The composition of the physical properties of the predictive cycle measurement data Dni encompasses the composition of the physical properties of the training cycle measurement data Ani.

[0059] In the following step S42, the prediction cycle measurement data Dni is input to the trained lifetime prediction model Cni with the corresponding number of predicted execution cycles ni, and the lifetime probability distribution Eni is obtained as output. If the prediction cycle measurement data Dni contains physical property values ​​that are not present in the training cycle measurement data Ani, those items are removed from Dni before being input to the lifetime prediction model Cni.

[0060] Steps S41 and S42 may be repeated for all one or more predicted execution cycle numbers ni. However, if the operator determines that they have gained sufficient confidence regarding the lifespan of the target cell based on the lifespan prediction results for an early predicted execution cycle number, they may omit predictions for subsequent predicted execution cycle numbers ni. In that case, the operator can terminate the charge-discharge cycle test of the target cell earlier.

[0061] Figure 5 shows the hardware configuration of the information processing device that executes the life prediction model described above. As shown in Figure 5, the information processing device 500 includes a processor 501, memory 502, auxiliary storage device 503, I / F (Interface) device 504, communication device 505, and drive device 506. The hardware components of the information processing device 500 are interconnected via a bus 507.

[0062] The processor 501 has various computing devices such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit). The processor 501 reads various programs (for example, a lifespan prediction program) into the memory 502 and executes them.

[0063] Memory 502 has main memory devices such as ROM (Read Only Memory) and RAM (Random Access Memory). The processor 501 and memory 502 form a so-called computer, and the computer realizes the above functions by the processor 501 executing various programs read from memory 502.

[0064] The auxiliary storage device 503 stores various programs and various data used when those programs are executed by the processor 501.

[0065] The I / F device 504 is a connection device that connects the operating device 511 and the display device 512 to the information processing device 500. The communication device 505 is a communication device for communicating with an external device (not shown) via a network. The drive device 506 is a device for setting the recording medium 513.

[0066] The various programs to be installed on the auxiliary storage device 503 are installed, for example, when the distributed recording medium 513 is set in the drive device 506 and the various programs recorded on the recording medium 513 are read by the drive device 506. Alternatively, the various programs to be installed on the auxiliary storage device 503 may be installed by downloading them from the network via the communication device 505.

[0067] [Examples] The following describes an example based on this embodiment.

[0068] In this embodiment, in step S1, charge-discharge cycle tests were performed on 104 lithium-ion battery cells, and learning cycle measurement data A and learning life data B were obtained. The charge-discharge cycle test consisted of five steps: a constant current charge step, a constant voltage charge step, a pause step, a constant current discharge step, and a pause step. Three physical properties were recorded to be included in the cycle measurement data: the voltage applied to the cell, the current flowing through the cell, and the current capacity.

[0069] Figure 6 is a graph plotting a portion of the cycle measurement data used in the embodiment. Specifically, data from the learning cycle measurement data A acquired in the embodiment was extracted from a single cell, from before the start of the first cycle to after the end of the second cycle. The data is plotted with time on the horizontal axis and voltage, current, and current capacity on the vertical axis. As shown in Figure 6, in the charge-discharge cycle test of this embodiment, the voltage range is from 2.8V to 4.2V, and the current range is from -50mA to 50mA. The plots are smooth curves or straight lines within each step, indicating that the sampling time was appropriately selected to adequately track the changes in physical properties.

[0070] In this embodiment, the cycle life was defined as the number of cycles at which the cell's discharge capacity falls below 80% of the first cycle's discharge capacity. The charge-discharge cycle tests to obtain the learning life data B in step S1 ranged from 2,000 cycles for the shortest to 4,000 cycles for the longest. For some cells that did not reach their cycle life, the cycle life was estimated using a combination of the measured discharge capacity changes and a power law. The estimated cycle life was used as the learning life data B, provided that the root mean square error of the curve fitting using the power law was less than 0.001 mAh. As a result, the number of cells included in the learning cycle measurement data A and learning life data B in this embodiment was 70, and the range of lifespans included in the learning life data B was approximately from 500 cycles to 6,000 cycles.

[0071] In this embodiment, ten predicted execution cycles ni were selected as the number of predicted execution cycles ni determined in step S2, ranging from 100 cycles to 1,000 cycles in increments of 100 cycles.

[0072] For the 10 predicted execution cycle numbers ni determined in step S2, step S31 was performed to extract the cycle measurement data from the training cycle measurement data A up to the predicted execution cycle number ni as training cycle measurement data Ani, and step S32 was repeated to train a lifetime prediction model C using the extracted training cycle measurement data Ani as an explanatory variable and the training lifetime data B transformed into a logarithmic scale as the target variable, thereby obtaining a trained lifetime prediction model Cni for the predicted execution cycle number ni.

[0073] The details of the lifetime prediction model in this embodiment are as follows.

[0074] The data formatting unit C-1 receives cycle measurement data up to the predicted execution cycle number ni as input and outputs fixed-length data in matrix format where the number of rows is the number of cells and the number of columns is fixed length. In this embodiment, the fixed-length data is generated by formatting the cycle measurement data of each cell into data with a fixed number of columns using the formatting processes (1) to (7) below, and then stacking them in the row direction.

[0075] (1) Physical property values ​​at specific points in the charge-discharge cycle test and values ​​calculated using them include: voltage at the start of charging in the first cycle, voltage at the start of charging in the second cycle, voltage at the start of charging in the nth cycle, total charging time in the first cycle, total charging time in the second cycle, total charging time in the nth cycle, constant voltage charging time in the first cycle, constant voltage charging time in the second cycle, constant voltage charging time in the nth cycle, total charge capacity in the first cycle, total charge capacity in the second cycle, total charge capacity in the nth cycle, constant voltage charging capacity in the first cycle, constant voltage charging capacity in the second cycle, constant voltage charging capacity in the nth cycle, discharge capacity in the first cycle, and 2nd cycle The data includes 33 items: the discharge capacity of the first cycle, the discharge capacity of the second cycle, the discharge capacity retention rate of the second cycle, the slope and intercept of the cycle count-start voltage plot around the second cycle, the slope and intercept of the cycle count-total charging time plot around the second cycle, the slope and intercept of the constant voltage charging time plot around the second cycle, the slope and intercept of the total charging capacity plot around the second cycle, the slope and intercept of the constant voltage charging capacity plot around the second cycle, the slope and intercept of the discharge capacity plot around the second cycle, and the slope and intercept of the discharge capacity retention rate plot around the second cycle. All of these are single values, and the length of each item is 1.

[0076] (2) The physical properties of each cycle in the charge-discharge cycle test included seven items: the voltage at the start of charging for each cycle, the total charging time for each cycle, the constant voltage charging time for each cycle, the total charging capacity for each cycle, the constant voltage charging capacity for each cycle, the discharge capacity for each cycle, and the discharge capacity retention rate for each cycle. The length of each item is the predicted number of cycles to be performed, ni.

[0077] (3) The current capacity data in the constant current charging step of the charge-discharge cycle was considered as a function of voltage, and linear interpolation was performed on 101 voltage sampling points that were sampled at equal intervals over the voltage range of 2.8V to 4.2V in this embodiment. By performing this process for all cycles, the data was converted into data with a length of ni (predicted number of execution cycles) × 101 (number of resampling points for voltage).

[0078] (4) The current capacity data in the constant current discharge step of the charge-discharge cycle was processed in the same way as the current capacity data in the constant current charge step.

[0079] (5) The current capacity and time data in the constant voltage charging step of the charge-discharge cycle were considered as functions of current, and linear interpolation was performed on 101 current sampling points that were sampled at equal intervals over the current range of -50mA to 50mA in this embodiment. By performing this process for all cycles, the data for current capacity and time were converted into data with a length of ni (predicted number of execution cycles) × 101 (number of resampling points for voltage).

[0080] (6) Data for the pause step in the charge / discharge cycle has been deleted.

[0081] (7) The 33 items of length 1 obtained by the formatting processes in (1) to (6) above, the 7 items of length ni, and the 4 items of length ni × 10¹ were concatenated in the column direction to generate one row of fixed-length data in matrix format.

[0082] Figure 7A is a graph plotting the current capacity at the constant current discharge step of the charge-discharge cycle against voltage, before processing by the data shaping process (4) of the data shaping unit C-1. The three plots represent current capacity at different cycle counts, and the number of samples and sampling positions at voltage differ between the plots. Figure 7B is a graph plotting the current capacity obtained by applying the data shaping process (4) to the data shown in Figure 7A and resampling at a common voltage sampling point, against voltage.

[0083] The feature extraction unit C-2 receives fixed-length data output from the data formatting unit C-1 as input, applies a data compression method to it, and outputs compressed data. In this embodiment, principal component analysis was adopted as the data compression method, and principal component analysis was applied separately to each item linked in the column direction within the matrix-format fixed-length data to obtain compressed data. At this time, 33 items with a column length of only 1 were left as they were without applying principal component analysis. The number of principal components extracted in the principal component analysis was selected to be sufficiently large so that the cumulative contribution rate of the extracted principal components was 90% or more. Specifically, for the 7 items whose column length is the number of predicted execution cycles ni, 10 principal components were extracted from each, and for the 4 items whose column length is the number of predicted execution cycles ni × 10¹, 20 principal components were extracted from each.

[0084] The nonlinear transformation unit C-3 receives the compressed data output from the feature extraction unit C-2 as input, applies a nonlinear transformation, and outputs it as nonlinear feature data. The subsequent regression unit C-4 receives the nonlinear feature data as input and outputs it as a probability distribution of lifetimes. In this embodiment, Gaussian process regression, which has both functions, is adopted as the method used in the nonlinear transformation unit C-3 and the regression unit C-4, and the sum of the radial basis function kernel, constant kernel, and white kernel is used as the kernel.

[0085] By repeating steps S31 and S32, a set of trained lifetime prediction models Cni corresponding to each of the 10 predicted execution cycle numbers ni was obtained. For these 10 ni and Cni, steps S41 and S42 were repeated in this embodiment as follows.

[0086] In step S41 of this embodiment, charge-discharge cycle tests were performed on the two cells whose lifespan was to be predicted up to the predicted number of cycles ni, and prediction cycle measurement data Dni was obtained. The structure of the prediction cycle measurement data Dni is the same as the structure of the learning cycle measurement data Ani in this embodiment.

[0087] In the subsequent step S42, the prediction cycle measurement data Dni is input to the trained lifetime prediction model Cni with the corresponding number of predicted execution cycles ni, and the probability distribution Eni of the logarithm of the lifetime is obtained as output.

[0088] Figure 8 shows graphs plotted for two cells, obtained by calculating the mean and 95% confidence interval from the logarithmic probability distribution Eni of the lifetime, and then inversely transforming the logarithm. The squares "□" represent the mean lifetime, and the upper and lower whiskers represent the 95% confidence interval of the lifetime. The two cells targeted for prediction are shown separately in Figure 8, 8A and Figure 8, 8B. The cell corresponding to Figure 8, 8A reached its lifespan at 668 cycles in the charge-discharge cycle test, so predictions were not made beyond 700 cycles. The actual lifetime of 668 cycles is shown as a dashed line in Figure 8, 8A. The cell corresponding to Figure 8, 8B did not reach its cycle lifetime even at the final predicted number of cycles, 1,000 cycles. For verification, the charge-discharge cycle test was continued up to 1,400 cycles, and the actual cycle lifetime was estimated to be 4,650 cycles using the combination of the obtained discharge capacity and the power law. The root mean square error of the curve fitting was less than 0.001 mAh. The estimated actual lifespan of 4,650 cycles, obtained through estimation, is shown as a dashed line in 8B of Figure 8.

[0089] In both Figure 8A and Figure 8B, the average predicted lifespan "□" is close to the actual or estimated actual lifespan of the target cell, and it can be seen that the predicted 95% confidence interval converges to the actual or estimated actual lifespan as the predicted number of cycles increases. If the operator determines that the 95% confidence interval has become sufficiently narrow, predictions for subsequent predicted number of cycles can be omitted, and the charge / discharge cycle test can be terminated early.

[0090] [Second Embodiment] In the first embodiment described above, the prediction item based on predictive cycle measurement data was described in which the probability distribution of lifetime was predicted. However, the prediction item is not limited to the probability distribution of lifetime, and for example, it may be configured to further predict the predicted discharge capacity maintenance rate in each cycle from the predicted number of execution cycles to reaching the predicted lifetime, or beyond the predicted lifetime. The second embodiment will now be described, focusing on the differences from the first embodiment.

[0091] First, we will explain the discharge capacity maintenance rate in each cycle, which is a new prediction item in the second embodiment. Figure 9 shows the discharge capacity maintenance rate up to the predicted number of execution cycles and the predicted discharge capacity maintenance rate after the predicted number of execution cycles.

[0092] Similar to Figure 1, the horizontal axis in Figure 9 represents the number of cycles, and the vertical axis represents the discharge capacity retention rate.

[0093] In Figure 9, the symbol 910 represents a graph of prediction cycle measurement data acquired up to the predicted number of execution cycles.

[0094] Furthermore, symbol 920 represents a plot of the predicted discharge capacity maintenance rate at any number of cycles after the predicted execution cycle number (for example, cycle number = 200, 300, 400, etc.), and a line graph connecting each plot.

[0095] Furthermore, symbols 921 and 922 represent line graphs connecting the upper and lower limits of the 95% confidence intervals when the predicted discharge capacity maintenance rate is calculated for any number of cycles after the predicted number of execution cycles.

[0096] In this way, by displaying the predicted discharge capacity maintenance rate for each cycle until the predicted lifespan is reached, the user can understand the trend of the predicted discharge capacity maintenance rate until the predicted lifespan is reached.

[0097] In the example in Figure 9, the probability distribution of lifespan is shown by combining the mean lifespan (symbol 930) and the 95% confidence interval for lifespan (symbol 931).

[0098] Next, the flow of the life prediction method of the second embodiment will be described. Figure 10 is a flowchart showing the flow of the life prediction method of the second embodiment of the present invention. The differences from the flowchart described using Figure 2 in the first embodiment are steps S12, S33, and S43.

[0099] In step S12, the discharge capacity retention rate B' for each cycle is calculated, which was obtained in step S11 by continuing the charge-discharge cycle test until the cycle life was reached.

[0100] In step S33, the extracted training cycle measurement data Ani is used as the explanatory variable and the discharge capacity maintenance rate B' as the objective variable to train the discharge capacity maintenance rate prediction model C', and the trained discharge capacity maintenance rate prediction model C'ni for the predicted number of cycles ni is obtained.

[0101] In steps S31 and S33, a trained discharge capacity maintenance rate prediction model C'ni is prepared, which learns from cycle measurement data up to the initial ni cycles corresponding to the predicted number of execution cycles ni. That is, if multiple predicted number of execution cycles are selected, multiple discharge capacity maintenance rate prediction models C'ni are trained for each initial ni cycles corresponding to the predicted number of execution cycles ni.

[0102] As mentioned above, the initial ni cycles refer to, for example, the initial 200 cycles, meaning cycles 1 through 200. The discharge capacity maintenance rate prediction model C'200 is trained using training cycle measurement data A200, which is extracted from training cycle measurement data A for the initial 200 cycles from cycles 1 to 200, and the discharge capacity maintenance rate B' for each cycle. Once the trained discharge capacity maintenance rate prediction model C'200 receives prediction cycle measurement data D200 up to the predicted number of cycles executed (200), it outputs the probability distribution E'200 of the predicted discharge capacity maintenance rate.

[0103] The target variable, the predicted discharge capacity maintenance rate, is usually greater than zero and less than the discharge capacity maintenance rate at the predicted number of cycles. Therefore, when training the discharge capacity maintenance rate prediction model C' in step S33, it is desirable to train the model with a value converted by equation (1) below, rather than directly training the discharge capacity maintenance rate B'. In this case, the predicted discharge capacity maintenance rate will also be a value converted by equation (1), and it is possible to guarantee that the predicted discharge capacity maintenance rate obtained by the inverse conversion will be greater than zero and less than the discharge capacity maintenance rate at the predicted number of cycles. Equation (1): Converted value = log(ymi / (yni-ymi)) However, ymi is the predicted discharge capacity retention rate at the predicted number of cycles mi, and yni is the discharge capacity retention rate at the predicted number of cycles ni.

[0104] In step S43, the predictive cycle measurement data Dni is input to the trained discharge capacity maintenance rate prediction model C'ni for the corresponding number of predicted execution cycles ni, and the probability distribution E'ni of the discharge capacity maintenance rate in each cycle is obtained as output.

[0105] Furthermore, the predicted discharge capacity maintenance rate and its 95% confidence interval are obtained from the probability distribution E'ni of the converted value obtained by equation (1) of the discharge capacity maintenance rate, and the inverse transform is performed to obtain the plot shown as illustrated in Figure 9.

[0106] Figure 11 shows the configuration of cycle measurement data and discharge capacity maintenance rate data. Of these, the details of the cycle measurement data have already been explained using Figure 3 in the first embodiment described above, so the explanation is omitted here.

[0107] The discharge capacity retention rate is the discharge capacity retention rate for each cycle of each cell from which cycle measurement data was acquired. The example in Figure 11 shows that it includes the discharge capacity retention rate in cycle m1, the discharge capacity retention rate in cycle m2, etc.

[0108] Next, the configuration of the life prediction and discharge capacity maintenance rate prediction model used in the second embodiment will be described. Figure 12 is a block diagram schematically showing the configuration of the life prediction and discharge capacity maintenance rate prediction model used in the second embodiment of the present invention.

[0109] As shown in Figure 12, the lifetime prediction and discharge capacity maintenance rate prediction model in the second embodiment includes a data shaping unit C-1, a feature extraction unit C-2, a nonlinear transformation unit C-3, a regression unit C-4, a nonlinear transformation unit C'-3, and a regression unit C'-4. The lifetime prediction and discharge capacity maintenance rate prediction model receives prediction cycle measurement data up to the prediction execution cycle number ni as input, and outputs a probability distribution of lifetime and a probability distribution of discharge capacity maintenance rate for each cycle.

[0110] Below, as an example, the elements constituting the lifetime prediction and discharge capacity maintenance rate prediction models will be explained. However, since the data shaping section C-1 to the regression section C-4 have already been explained using Figure 4 in the first embodiment described above, the nonlinear transformation section C'-3 and the regression section C'-4 will be explained here.

[0111] The nonlinear transformation unit C'-3 receives the compressed data output from the feature extraction unit C-2 as input, applies a nonlinear transformation, and outputs nonlinear feature data. The subsequent regression unit C'-4 receives the nonlinear feature data as input and outputs the probability distribution of the discharge capacity maintenance rate in each cycle.

[0112] The methods used in the nonlinear transformation unit C'-3 and the regression unit C'-4 may be set individually, but it is more preferable to apply a Gaussian process regression that has both functions to C'-3 and C'-4. If set individually, it is preferable to use a kernel method in the nonlinear transformation unit C'-3. On the other hand, since the output of the regression unit C'-4 must be a probability distribution, methods other than Gaussian process regression, such as Bayesian ridge, will be used.

[0113] Furthermore, when applying Gaussian process regression to the nonlinear transformation section C'-3 and the regression section C'-4, or when applying a kernel method to the nonlinear transformation section C'-3, a nonlinear kernel such as a radial basis function kernel is used to ensure that the transformation is nonlinear.

[0114] Thus, according to the second embodiment, based on predictive cycle measurement data, it is possible to predict the probability distribution of lifetime, the predicted discharge capacity maintenance rate and 95% confidence interval for each cycle from the predicted number of cycles to the predicted lifetime, or beyond the predicted lifetime.

[0115] [Other embodiments] In the second embodiment described above, the case in which the prediction of the probability distribution of life and the prediction of the probability distribution of discharge capacity maintenance rate are realized using an integrated model (life prediction and discharge capacity maintenance rate prediction model) was explained. However, the prediction of the probability distribution of life and the prediction of the probability distribution of discharge capacity maintenance rate may also be realized using separate models (life prediction model and discharge capacity maintenance rate prediction model). Specifically, a life prediction program and a discharge capacity maintenance rate prediction program may be prepared separately, and the information processing device 500 may be configured to execute each program independently. [Industrial applicability]

[0116] The lithium-ion battery life prediction method of the present invention can be suitably used for life evaluation in lithium-ion battery development.

Claims

1. A charge-discharge test is performed on each cell of a lithium-ion battery, which includes a step of constant-current charging followed by a step of constant-current discharging, thereby acquiring learning data including cycle measurement data recording the voltage applied to each cell, the current flowing through each cell, and the current capacity, as well as life data. (a) (b) A set of trained life prediction models Cni, corresponding to the number of predicted execution cycles ni, is obtained, where for one or more predicted execution cycle numbers ni (where i is an integer of 1 or more), a training cycle measurement data Ani, obtained by extracting an initial ni cycle number from the cycle measurement data of each cell in the training data, is used as the explanatory variable, and the lifespan data of the training data is used as the objective variable. When predicted cycle measurement data obtained up to the corresponding predicted execution cycle number is input for the lithium-ion battery to be predicted, the trained model Cni outputs the probability distribution of the lifespan at that predicted execution cycle number. A method for training a life prediction model that a computer performs.

2. A method for learning a lifetime prediction model according to claim 1, wherein the lifetime data included in the learning data is estimated using a combination of the change in discharge capacity up to the measured charge-discharge cycle when the charge-discharge test is not performed until the cycle lifetime is reached, and a mathematical model or a time-series data analysis method.

3. The method for learning a lifetime prediction model according to claim 2, wherein the mathematical model includes either a square root law or a power law, and the time series data analysis method includes either an autoregressive integrated moving average (ARIMA) model, a Kalman filter, or a Gaussian process regression.

4. A life prediction method for predicting the life of a lithium-ion battery using a trained life prediction model Cni trained by the life prediction model training method described in claim 2 or 3, wherein the life prediction model is trained by the trained life prediction model Cni Step (c) is to perform a charge-discharge test on the lithium-ion battery cell to be predicted, which includes a constant-current charging step followed by a constant-current discharging step, for a predicted number of cycles nx (where x is any integer of 1 or more), thereby obtaining predictive cycle measurement data that records the voltage applied to the cell, the current flowing through the cell, and the current capacity. Step (d) involves inputting the prediction cycle measurement data obtained up to the predicted execution cycle number nx into a trained lifetime prediction model Cni (where i = x) for the predicted execution cycle number ni (where i = x), and obtaining the probability distribution of lifetime at the predicted execution cycle number nx as output. A computer-based method for predicting the lifespan of lithium-ion batteries.

5. The aforementioned lifetime prediction model, A data formatting unit receives the learning cycle measurement data Ani from which the initial ni cycle count has been extracted for each cell, generates fixed-length data by arranging the values ​​of each measurement item for each cycle in an increasing direction to increase the number of columns, and then formats the generated fixed-length data into matrix-format fixed-length data by arranging the generated fixed-length data in an increasing direction to increase the number of rows by the number of cells. A feature extraction unit compresses the aforementioned matrix-formatted fixed-length data into compressed data, A nonlinear transformation unit converts the compressed data into nonlinear feature data mapped to a high-dimensional space, A method for predicting the lifespan of a lithium-ion battery according to claim 4, characterized in that a regression unit that takes the aforementioned nonlinear feature data as input and outputs a probability distribution of lifespan is connected in this order.

6. The lithium-ion battery life prediction method according to claim 5, characterized in that a dimensionality reduction method is used as the data compression method in the feature extraction unit.

7. The method for predicting the lifespan of a lithium-ion battery according to claim 6, characterized in that the dimensionality reduction method is principal component analysis.

8. A charge-discharge test is performed on each cell of a lithium-ion battery, which includes a step of constant-current charging followed by a step of constant-current discharging, thereby acquiring learning data including cycle measurement data and life data, which record the voltage applied to each cell, the current flowing through each cell, and the current capacity. (b) A set of trained life prediction models Cni, corresponding to the number of predicted execution cycles ni, is obtained, where for one or more predicted execution cycle numbers ni (where i is an integer of 1 or more), a training cycle measurement data Ani, obtained by extracting an initial ni cycle number from the cycle measurement data of each cell in the training data, is used as the explanatory variable, and the lifespan data of the training data is used as the objective variable. When predicted cycle measurement data obtained up to the corresponding predicted execution cycle number is input for the lithium-ion battery to be predicted, the trained model Cni outputs the probability distribution of the lifespan at that predicted execution cycle number. A training program for a lifespan prediction model to be executed by a computer.

9. A charge-discharge test is performed on each cell of a lithium-ion battery, which includes a step of constant-current charging followed by a step of constant-current discharging, thereby acquiring learning data including cycle measurement data and life data, which record the voltage applied to each cell, the current flowing through each cell, and the current capacity. (b) A set of trained life prediction models Cni, corresponding to the number of predicted execution cycles ni, is obtained, where for one or more predicted execution cycle numbers ni (where i is an integer of 1 or more), a training cycle measurement data Ani, obtained by extracting an initial ni cycle number from the cycle measurement data of each cell in the training data, is used as the explanatory variable, and the lifespan data of the training data is used as the objective variable. When predicted cycle measurement data obtained up to the corresponding predicted execution cycle number is input for the lithium-ion battery to be predicted, the trained model Cni outputs the probability distribution of the lifespan at that predicted execution cycle number. An information processing device that performs the following actions.

Images