Prediction method, data collection method, learning method, prediction device, data collection system, and learning device

Abstract

To suitably predict operating characteristics of a lithium metal-based battery cell.SOLUTION: A prediction method for predicting operating characteristics of a battery cell having a lithium metal negative electrode includes: an acquisition step of acquiring discharge capacity data in a plurality of charge / discharge cycles relating to a target battery cell; and a prediction step of predicting operating characteristics of the target battery cell by inputting one or more feature quantities obtained by referring to the discharge capacity data in the plurality of charge / discharge cycles to a prediction model. The one or more feature quantities include discharge-related feature quantities.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a prediction method, a data collection method, a learning method, a prediction device, a data collection system, and a learning device for predicting the operating characteristics of a battery cell. [Background technology]

[0002] There are known techniques for predicting the operating characteristics of a battery cell, such as the battery life, etc. For example, Patent Document 1 discloses a technique for collecting data to be used in data-driven predictive modeling for predicting the operating characteristics of a battery cell, training a machine learning model using the collected data, and predicting the operating characteristics of the battery cell using the machine learning model. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent No. 7317484 Summary of the Invention [Problem to be solved by the invention]

[0004] The technology disclosed in Patent Document 1 relates to predicting the operating characteristics of graphite-based lithium-ion batteries, but in recent years, lithium metal-based battery cells have also been attracting attention from the perspective of increasing energy density. Such lithium metal-based battery cells exhibit a more complex electrochemical profile during battery operation than graphite-based lithium-ion batteries, making it difficult to predict their operating characteristics.

[0005] One aspect of the present invention is to provide a technique that can favorably predict the operating characteristics of lithium metal-based battery cells. [Means for solving the problem]

[0006] In order to solve the above-mentioned problems, a prediction method according to one aspect of the present invention is a method for predicting operating characteristics of a battery cell having a lithium metal negative electrode, the method including: acquiring discharge capacity data for a plurality of charge / discharge cycles of a target battery cell; and predicting operating characteristics of the target battery cell by inputting one or more feature quantities obtained by referring to the discharge capacity data for the plurality of charge / discharge cycles into a prediction model, the one or more feature quantities including: Among the discharge capacity data in the plurality of charge / discharge cycles, an extreme value of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; Among the discharge capacity data in the plurality of charge / discharge cycles, the variance of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; and An intercept obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles The feature amount depends on at least one of the above.

[0007] In order to solve the above-mentioned problems, a data collection method according to one aspect of the present invention is a data collection method for collecting battery cell data to be used in data-driven predictive modeling for predicting operating characteristics of a battery cell having a lithium metal negative electrode, the data collection method including: a measurement step of sequentially or continuously measuring one or more physical properties of each battery cell in a plurality of charge / discharge cycles for each of one or more battery cells; an acquisition step of acquiring discharge capacity data for each of the battery cells based on the one or more measured physical properties; and a generation step of generating one or more feature quantities to be used in data-driven predictive modeling by referring to the discharge capacity data, wherein the one or more feature quantities include: Among the discharge capacity data in the plurality of charge / discharge cycles, an extreme value of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; Among the discharge capacity data in the plurality of charge / discharge cycles, the variance of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; and An intercept obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles The feature amount depends on at least one of the above.

[0008] In order to solve the above-mentioned problems, a learning method according to one aspect of the present invention is a learning method for training a prediction model that predicts the operating characteristics of a battery cell having a lithium metal negative electrode, the learning method including: an acquisition step of acquiring discharge capacity data in a plurality of charge / discharge cycles for each of one or a plurality of battery cells; and a learning step of training the prediction model using one or a plurality of feature amounts obtained by referring to the discharge capacity data in the plurality of charge / discharge cycles, wherein the one or a plurality of feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, an extreme value of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; Among the discharge capacity data in the plurality of charge / discharge cycles, the variance of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; and An intercept obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles The feature amount depends on at least one of the above.

[0009] In order to solve the above problem, a prediction device according to one aspect of the present invention is a prediction device that predicts operating characteristics of a battery cell having a lithium metal negative electrode, and includes an acquisition unit that acquires discharge capacity data for a plurality of charge / discharge cycles of a target battery cell, and a prediction unit that predicts operating characteristics of the target battery cell by inputting one or more feature quantities obtained by referring to the discharge capacity data for the plurality of charge / discharge cycles into a prediction model, wherein the one or more feature quantities include: Among the discharge capacity data in the plurality of charge / discharge cycles, an extreme value of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; Among the discharge capacity data in the plurality of charge / discharge cycles, the variance of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; and An intercept obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles The feature amount depends on at least one of the above.

[0010] In order to solve the above-mentioned problems, a data collection system according to one aspect of the present invention is a data collection system that collects battery cell data for use in data-driven predictive modeling for predicting operating characteristics of a battery cell having a lithium metal negative electrode, and includes a measurement unit that sequentially or continuously measures one or more physical properties of each battery cell during a plurality of charge / discharge cycles for each of one or more battery cells, an acquisition unit that acquires discharge capacity data for each of the battery cells based on the one or more measured physical properties, and a generation unit that generates one or more feature quantities to be used in data-driven predictive modeling by referring to the discharge capacity data, and the one or more feature quantities include: Among the discharge capacity data in the plurality of charge / discharge cycles, an extreme value of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; Among the discharge capacity data in the plurality of charge / discharge cycles, the variance of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; and An intercept obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles The feature amount depends on at least one of the above.

[0011] In order to solve the above problem, a learning device according to one aspect of the present invention is a learning device that learns a prediction model that predicts the operating characteristics of a battery cell having a lithium metal negative electrode, and includes an acquisition unit that acquires discharge capacity data in a plurality of charge / discharge cycles for each of one or a plurality of battery cells, and a learning unit that learns the prediction model using one or a plurality of feature amounts obtained by referring to the discharge capacity data in the plurality of charge / discharge cycles, wherein the one or a plurality of feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, an extreme value of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; Among the discharge capacity data in the plurality of charge / discharge cycles, the variance of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; and An intercept obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles The feature amount depends on at least one of the above.

[0012] The information processing device according to each aspect of the present invention may be realized by a computer. In this case, the scope of the present invention also includes a program for the information processing device that causes the computer to operate as each unit (software element) of the information processing device to realize the information processing device on the computer, and a computer-readable recording medium on which the program is recorded. Furthermore, inventions in which the "prediction model" is expressed using terms such as "learning model" or "machine learning model" are also within the scope of the present invention. [Effects of the Invention]

[0013] According to one aspect of the present invention, the operating characteristics of a lithium metal-based battery cell can be favorably predicted. [Brief explanation of the drawings]

[0014] [Figure 1] FIG. 1 is a flowchart showing the flow of a learning method according to an embodiment of the present invention. [Figure 2]FIG. 1 is a flowchart showing the flow of a prediction method according to an embodiment of the present invention. [Figure 3] 1 is a block diagram showing a configuration of an information processing system according to an embodiment of the present invention. [Figure 4] FIG. 4 is a diagram for explaining an information processing method according to an embodiment of the present invention, showing a cycle curve of a certain battery. [Figure 5] This figure is for explaining an information processing method according to an embodiment of the present invention. The upper part shows the discharge capacity-voltage curves for cycle 100 (100th cycle) and cycle 10 (10th cycle) of a specific battery included in the dataset used, and the lower part shows the ΔDQ100-10(V) curve of the battery. [Figure 6] This figure is for explaining an information processing method according to an embodiment of the present invention. The upper part shows the charge capacity-voltage curves for cycle 100 (100th cycle) and cycle 10 (10th cycle) of a specific battery included in the dataset used, and the lower part shows the discharge capacity, charge capacity, and coulombic efficiency of a representative battery as a function of cycle number. [Figure 7] FIG. 1 is a diagram for explaining an information processing method according to an embodiment of the present invention, showing the relaxation voltage as a function of relaxation time from the first cycle to the last cycle (200 cycles) of a particular battery. [Figure 8] FIG. 1 is a diagram illustrating an information processing method according to an embodiment of the present invention, in which the discharge capacity of 48 batteries is plotted as a function of the number of cycles. [Figure 9] This figure is for explaining an information processing method according to an embodiment of the present invention, and the upper part shows a representative discharge profile (normalized discharge capacity curve) of one LMB cell (cell No. 100), and the lower part shows a discharge profile (normalized discharge capacity curve) of another LMB cell (cell No. 35). [Figure 10]FIG. 1 is a diagram for explaining an information processing method according to an embodiment of the present invention, in which the upper row shows prediction results for a discharge-related feature subset, the middle row shows prediction results for a charge-related feature subset, and the lower row shows prediction results for a relaxation-related feature subset. [Figure 11] FIG. 1 is a diagram for explaining an information processing method according to an embodiment of the present invention, in which the upper part shows a heat map matrix of Pearson's correlation coefficients between variables in a dataset including feature quantities and feature quantities, and feature quantities and target values, and the lower part shows the Pearson's correlation coefficients of feature quantities with respect to the cycle life, which is the observed target value, of 40 batteries. [Figure 12] FIG. 1 is a diagram illustrating an information processing method according to an embodiment of the present invention, in which the top row shows cycle life as a function of ΔDQ100-10 (V), the middle row shows cycle life as a function of the slope of the discharge capacity from cycle 2 to cycle 100, and the bottom row shows cycle life as a function of the dispersion of the relaxation voltage from the first cycle to cycle 100. [Figure 13] FIG. 1 is a diagram for explaining an information processing method according to an embodiment of the present invention, in which the upper part shows a parity plot of ElasticNet and XGBoost for 12 features, and the lower part shows a histogram of prediction results. [Figure 14] FIG. 1 is a diagram for explaining an information processing method according to an embodiment of the present invention. The upper part shows a parity plot obtained using a set of six features, and the lower part shows the importance of each of these six features. [Figure 15] FIG. 1 is a diagram for explaining an information processing method according to an embodiment of the present invention, showing a prediction result by a prediction model according to the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0015] [Embodiment 1] Hereinafter, one embodiment of the present invention will be described in detail.

[0016] <Learning Method> 1 is a flow chart showing the flow of the learning method according to this embodiment. measuring one or more physical properties for each of the one or more battery cells; generating one or more feature quantities by referring to the one or more physical properties; - Training a predictive model using the one or more features Here, the prediction model is a prediction model that predicts the operating characteristics of a battery cell, and is a target of data-driven predictive modeling. Therefore, the learning method according to this embodiment is as follows: A data collection method for collecting battery cell data for use in data-driven predictive modeling for predicting operating characteristics of the battery cell. The learning method according to this embodiment will be described in more detail below. Note that in this specification, the prediction model is also referred to as a learning model or a machine learning model.

[0017] (Step S11) First, in step S11, one or more physical properties of each battery cell are sequentially or continuously measured during multiple charge / discharge cycles for each of one or more battery cells, where each battery cell is, for example, a lithium metal-based battery cell, such as a lithium ion battery having a lithium metal negative electrode, but this is not intended to be a limitation of this embodiment.

[0018] Furthermore, the one or more physical properties may include at least one of the battery cell voltage, the battery cell current, the battery cell can temperature, and the battery cell internal resistance, but this does not limit the present embodiment.

[0019] The charge / discharge cycles can be performed, for example, by cycling the one or more battery cells between a first voltage and a second voltage using a battery cycling instrument (measurement device). The battery cycling instrument (measurement device) can be used to sequentially or continuously measure one or more physical properties of each battery cell over multiple charge / discharge cycles. However, these specific examples are not intended to limit the present embodiment.

[0020] (Step S12) Subsequently, in step S12, discharge capacity data for each of the battery cells is acquired (generated) based on the one or more physical properties measured in step S11. As an example, in this step, the discharge capacity data for each battery cell is obtained (generated) by referring to the one or more physical properties, and the discharge capacity vs. voltage data (DQ) for the n-th charge / discharge cycle for the battery cell is used as the discharge capacity data for the battery cell. n (V) is acquired (generated). Here, the discharge capacity vs. voltage data refers to data including the relationship between each value of the discharge capacity of the battery cell and the voltage corresponding to each value. Furthermore, n is an index for distinguishing each charge / discharge cycle from one another, and is, for example, an ordinal number indicating the ordinal number of each charge / discharge cycle.

[0021] For example, the discharge capacity vs. voltage data in the first charge / discharge cycle of the battery cell is expressed as DQ1(V), and the discharge capacity vs. voltage data in the tenth charge / discharge cycle of the battery cell is expressed as DQ 10 (V). In this step, the discharge capacity data acquisition (generation) process is performed as follows: A process of generating a voltage curve in the nth charge / discharge cycle of the battery cell, the voltage curve being determined depending on the discharge capacity or charge capacity in the nth charge / discharge cycle. can also be expressed as containing

[0022] (Step S13) Subsequently, in step S13, one or more feature quantities to be used in the data-driven predictive modeling are generated by referring to the discharge capacity data acquired (generated) in step S12. (1) Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) (m≠n) and the difference (ΔDQ n-m ) extreme value (minΔDQ n-m (V) or maxΔDQ n-m (V)), (2) Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the difference (ΔDQ n-m ) variance (varΔDQ n-m (V)), and (3) an intercept (intercept_DQ) obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles; (4) Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity (C Dch(n) ) and discharge capacity in other cycles (C Dch(m) ) ratio (CR), (5) a slope (slope_DQ) obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles; and (6) Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the difference (ΔDQ n-m ) average (meanΔDQ n-m (V) A feature quantity that depends on at least one of the above is generated.

[0023] Here, the difference (ΔDQ n-m ) can be expressed as an extreme value, The difference (ΔDQ n-m ) maximum or maximum value The difference (ΔDQ n-m ) minimum or minimum value At least one of the above is included. n-m ) is a negative value, The difference (ΔDQ n-m ) absolute value of |ΔDQ n-m | or The difference (ΔDQ n-m ) with n and m swapped (ΔDQ m-n ) are positive values, and these values ​​are the difference (ΔDQ n-m ) can also be considered to correspond to the maximum or local maximum value of the above. In this way, the terms "maximum value," "local maximum value," "minimum value," "local minimum value," etc. used in this specification do not excessively limit the invention described in this specification, and may each include matters indicated by other terms.

[0024] Also, (1') The extreme value of the difference (minΔDQ n-m (V) or maxΔDQ n-m As a feature dependent on (V), the logarithm of the minimum value of the difference (log minΔDQ m-n (V)) may be used, (2') The variance of the difference (varΔDQ n-m As a feature dependent on (V), the logarithm of the variance of the difference (log varΔDQ n-m (V)) may be used, (6') The mean difference (mean ΔDQ n-m As a feature dependent on (V), the logarithm of the mean of the difference (log mean ΔDQ n-m (V)) may also be used.

[0025] Also, (4') As the ratio (CR), the discharge capacity (C Dch(n)) and the discharge capacity in the cycle immediately before the cycle in question (C Dch(n-1) )

number

[0026] Also, (1'') The extreme value of the difference (minΔDQ n-m (V) or maxΔDQ n-m (V)) dependent features, (2'') The variance of the difference (varΔDQ n-m (V)) dependent features, and (6'') The mean difference (mean ΔDQ n-m (V)) dependent features With respect to each of the following: n is a value between 70 and 95% of the number of cycles expected as the average life (average life cycle) of the target battery cell group, m is a value between 5 and 15% of the number of cycles assumed as the average life (average life cycle) of the target battery cell group, For example, n=100, m=10, etc. may be used as n and m above.

[0027] Also, (3') As a feature dependent on the intercept (intercept_DQ), an intercept obtained by applying a linear regression process to discharge capacity data for a plurality of cycles included in 70 to 95% of the number of cycles assumed as the average life (average life cycle) of the target battery cell group may be used. As an example, the intercept (intercept_DQ) obtained by applying a linear regression process to discharge capacity data from the 91st to 100th cycles may be used. 91:100 ) may also be used.

[0028] Also, (5') As a feature dependent on the slope (slope_DQ), a slope obtained by applying a linear regression process to discharge capacity data for a plurality of cycles included in 1 to 95% of the number of cycles assumed as the average life (average life cycle) of the target battery cell group may be used. For example, the slope (slope_DQ) obtained by applying a linear regression process to discharge capacity data from the second cycle to the 100th cycle may be used. 2:100 ) may be used. Such an average lifespan can be determined in advance, for example, by conducting a random inspection of a target battery cell group.

[0029] Also, (4'') As a feature dependent on the ratio (CR), Discharge capacity in the first cycle (C Dch(1) ) and the discharge capacity at the 100th cycle (C Dch(100) ) and the ratio (CR 1:100 ), Discharge capacity in the second cycle (C Dch(2) ) and the discharge capacity in the first cycle (C Dch(1) ) and the ratio (CR 2:1 ), and Discharge capacity at 99th cycle (C Dch(99) ) and the discharge capacity at the 100th cycle (C Dch(100) ) and the ratio (CR 99:100 ) At least one ratio of the above may be used.

[0030] (Step S14) Subsequently, in step S14, a prediction model for predicting the operating characteristics of the battery cell is trained using one or more feature amounts acquired (generated) in step S13.

[0031] Here, the prediction model has one or more parameters, and is a model that outputs an index related to the operating characteristics by applying a calculation process using the one or more parameters to the one or more feature quantities.

[0032] In this step, for example, one or more feature quantities acquired (generated) in step S13 are input to the prediction model, and at least one of the one or more parameters is updated so that the index output by the prediction model becomes more appropriate. More specifically, for example, In step S11 or another step, a ground truth label for each of the one or more battery cells is obtained; At least one of the one or more parameters is updated so that the difference between the index output by the prediction model to which the one or more feature quantities acquired (generated) in step S13 are input and the correct label becomes smaller. The prediction model is trained by performing the process described above. Here, the indicator of the operating characteristics output by the prediction model may be at least one of the battery life and the logarithm of the battery life, but this is not intended to limit the present embodiment.

[0033] Note that specific examples of prediction models do not limit this embodiment, but as an example, a regression model such as ElasticNet may be used, or a model based on a decision tree such as XGBoost (Extreme Gradient Boosting) may be used.

[0034] (Effect of learning method) According to the learning method of this embodiment, which includes the steps described above, at least one of the features (1) to (6) described above is generated, and a prediction model is trained using the generated features, so that a prediction model that can appropriately predict the operating characteristics of a battery cell can be constructed.

[0035] (Data collection method aspect) As partly described above, the learning method according to this embodiment includes: A data collection method for collecting battery cell data for use in data-driven predictive modeling for predicting operating characteristics of the battery cell. More specifically, steps S11 to S13 can be considered to include the above. More specifically, steps S11 to S13 can be considered to be a data collection method according to this embodiment. According to the data collection method according to this embodiment, at least one of the feature quantities (1) to (6) described above is generated, so that data (learning data) for constructing a prediction model that can suitably predict the operating characteristics of a battery cell can be suitably collected.

[0036] <Prediction method> Next, the prediction method according to this embodiment will be described with reference to Fig. 2. Fig. 2 is a flow chart showing the flow of the prediction method according to this embodiment. The prediction method according to this embodiment can be roughly described as follows: measuring one or more physical properties of the battery cell to be predicted (also referred to as the target battery cell); generating one or more feature quantities by referring to the one or more physical properties; The one or more feature quantities are input into a trained prediction model to predict the operating characteristics of the battery cell to be predicted. Here, as an example, the prediction model may be a prediction model trained by the above-described training method. The prediction method according to this embodiment will be described in more detail below.

[0037] (Step S21) First, in step S21, discharge capacity data for a target battery cell is acquired over multiple charge / discharge cycles. The battery cell may be, for example, a lithium metal-based battery cell, such as a lithium ion battery having a lithium metal anode, but this is not intended to limit the present embodiment.

[0038] Furthermore, the one or more physical properties may include at least one of the battery cell voltage, the battery cell current, the battery cell can temperature, and the battery cell internal resistance, but this does not limit the present embodiment.

[0039] The charge / discharge cycles can be performed by cycling the one or more battery cells between a first voltage and a second voltage using a battery cycling instrument (measurement device). The battery cycling instrument (measurement device) can be used to sequentially or continuously measure one or more physical properties of the target battery cells over multiple charge / discharge cycles. However, these specific examples are not intended to limit the present embodiment.

[0040] (Step S22) Subsequently, in step S22, discharge capacity data of the target battery cell is acquired (generated) based on the one or more physical properties measured in step S21. As an example, in this step, the discharge capacity vs. voltage data (DQ) of the target battery cell in the n-th charge / discharge cycle is acquired (generated) based on the one or more physical properties measured in step S21. n (V) is acquired (generated). Here, the discharge capacity vs. voltage data refers to data including the relationship between each value of the discharge capacity of the battery cell and the voltage corresponding to each value. Furthermore, n is an index for distinguishing each charge / discharge cycle from one another, and is, for example, an ordinal number indicating the ordinal number of each charge / discharge cycle.

[0041] For example, the discharge capacity vs. voltage data in the first charge / discharge cycle of the battery cell is expressed as DQ1(V), and the discharge capacity vs. voltage data in the tenth charge / discharge cycle of the battery cell is expressed as DQ 10 (V). In this step, the discharge capacity data acquisition (generation) process is performed as follows: A process of generating a voltage curve in the nth charge / discharge cycle of the battery cell, the voltage curve being determined depending on the discharge capacity or charge capacity in the nth charge / discharge cycle. can also be expressed as containing

[0042] (Step S23) Subsequently, in step S23, the discharge capacity data acquired (generated) in step S22 is referenced to generate one or more feature amounts to be used in data-driven predictive modeling.

[0043] The types of feature quantities generated in this step are the same as the types of feature quantities (e.g., (1) to (6) described above) generated in step S13 included in the learning method according to this embodiment, so a duplicated explanation will be omitted. However, the feature quantities generated in this step are feature quantities related to the battery cell that is the target of prediction by the prediction method according to this embodiment.

[0044] (Step S24) Next, in step S24, the one or more feature quantities generated in step S23 are input into a trained prediction model to predict the operating characteristics of the target battery cell. As an example, the one or more feature quantities generated in step S23 are input into the trained prediction model to output an index related to the operating characteristics of the target battery cell. Here, as an example, the trained prediction model can be a prediction model trained by the training method according to this embodiment. Furthermore, the index of the operating characteristics output by the prediction model can be at least one of battery life and the logarithm of the battery life, but this is not intended to limit this embodiment.

[0045] Furthermore, although the specific example of the predictive model does not limit this embodiment, as mentioned in the explanation of the learning method according to this embodiment, as an example, a regression model such as ElasticNet may be used, or a model based on a decision tree such as XGBoost (Extreme Gradient Boosting) may be used.

[0046] (Effect of forecasting method) According to the prediction method of this embodiment, which includes the steps described above, at least one of the features (1) to (6) described above is generated, and the generated feature is input into a trained prediction model, thereby outputting an index related to the operating characteristics of the target battery cell, thereby making it possible to suitably predict the operating characteristics of the target battery cell.

[0047] (Information processing system 100) Next, an information processing system 100 according to this embodiment will be described with reference to FIG. 3. FIG. 3 is a block diagram showing an example of the configuration of the information processing system 100 according to this embodiment. As shown in FIG. 3, the information processing system 100 includes, as an example, an information processing device 1 and a measurement device (measurement unit) 11. Schematically speaking, the information processing system 100 includes: measuring one or more physical properties of each of the plurality of battery cells BC1, BC2, Obtaining (generating) discharge capacity data for each of the battery cells based on the physical properties; generating one or more features to be used in data-driven predictive modeling with reference to the discharge capacity data; A prediction model for predicting the operating characteristics of the battery cell is trained by referring to the one or more feature amounts; -Predict the operating characteristics of the target battery cell using a trained prediction model For this reason, the information processing system 100 may be referred to as a data collection system, and the information processing device 1 may be referred to as a learning device or a prediction device.

[0048] (Measuring device 11) The measuring device 11 is sequentially or continuously measuring one or more physical properties of each battery cell during a plurality of charge / discharge cycles for each of the one or more battery cells BC1, BC2, ...; Measurement data DD including the measurement results is provided to the information processing device 1. The battery cells to be measured by the measuring device 11 include: A battery cell (also called a learning battery cell) for collecting learning data for training a predictive model in a learning phase in which the above-described learning method is executed; and In the prediction phase, which executes the prediction method described above, a battery cell (also called a prediction target battery cell) is used to obtain prediction data to be input into the trained prediction model. The one or more battery cells may be, for example, lithium metal-based battery cells, such as lithium ion batteries having a lithium metal negative electrode, although this is not intended to limit the present embodiment.

[0049] Furthermore, the one or more physical properties may include at least one of the battery cell voltage, the battery cell current, the battery cell can temperature, and the battery cell internal resistance, but this does not limit the present embodiment.

[0050] Additionally, the charge / discharge cycles can be performed, for example, by cycling the one or more battery cells between a first voltage and a second voltage using a battery cycling device (measurement device). The measurement device 11 can include such battery cycling device and sequentially or continuously measure one or more physical properties of the target battery cells over multiple charge / discharge cycles. However, these specific examples do not limit the present embodiment.

[0051] (Information processing device 1) As shown in FIG. 3, the information processing device 1 includes a control unit 10, a storage unit 20, a communication unit 30, and an input / output unit 40.

[0052] (Communication unit 30) The communication unit 30 communicates with one or more devices external to the information processing device 1. The communication unit 30 transmits data supplied from the control unit 10 to the external device, and supplies data received from the external device to the control unit 10. As an example, the communication unit 30 supplies measurement data DD acquired from the measurement device 11 to the control unit 10, and transmits the prediction result PR derived by the control unit 10 to the external device.

[0053] (Input / output section 40) The input / output unit 40 is configured to include at least one of input / output devices such as a keyboard, a mouse, a display, a printer, and a touch panel, for example. Alternatively, the input / output unit 40 may be configured to have input / output devices such as a keyboard, a mouse, a display, a printer, and a touch panel connected to it. In this configuration, the input / output unit 40 accepts various types of information input to the information processing device 1 from the connected input devices. Furthermore, the input / output unit 40 outputs various types of information to connected output devices under the control of the control unit 10. An example of the input / output unit 40 is an interface such as a USB (Universal Serial Bus).

[0054] (Storage unit 20) The storage unit 20 stores various data referenced by the control unit 10 and various data generated by the control unit 10. As an example, the storage unit 20 stores: Measurement data DD provided by measuring device 11 Discharge capacity data DQD generated (acquired) by referring to measurement data DD Charge capacity data CQD generated (acquired) by referring to measurement data DD Voltage data VD generated (acquired) by referring to measurement data DD A feature group FG including one or more feature values ​​generated by a generating unit 13 (to be described later) A prediction model PM learned by a learning unit 14 (described later) -Prediction results obtained using the prediction model PM etc. are stored. In this embodiment, the term "prediction model PM" may include the meaning of "one or parameters that define the prediction model PM." Furthermore, specific examples of the prediction model PM do not limit this embodiment, but as an example, a regression model such as ElasticNet may be used, or a model based on a decision tree such as XGBoost (Extreme Gradient Boosting) may be used.

[0055] (Control unit 10) As shown in FIG. 3, the control unit 10 includes an acquisition unit 12, a generation unit 13, a learning unit 14, and a prediction unit 24.

[0056] (Acquisition part 12) During the learning phase, the acquisition unit 12 acquires discharge capacity data for each of one or more learning battery cells based on the one or more physical properties measured by the measurement device 11 for multiple charge / discharge cycles for each of the one or more learning battery cells.

[0057] In addition, in the prediction phase, the acquisition unit 12 acquires discharge capacity data of the battery cell to be predicted based on the one or more physical properties measured by the measurement device 11 for multiple charge / discharge cycles of the battery cell to be predicted.

[0058] More specifically, as an example, the acquisition unit 12 executes the processes described in step S12 or step S22 above. Duplicate descriptions of these processes will be omitted.

[0059] (Generation part 13) In the learning phase, the generation unit 13 references the discharge capacity data on the learning battery cells acquired (generated) by the acquisition unit 12, and generates one or more feature amounts to be used in data-driven predictive modeling.

[0060] Furthermore, in the prediction phase, the generation unit 13 references the discharge capacity data regarding the battery cell to be predicted, which data has been acquired (generated) by the acquisition unit 12, and generates one or more feature amounts to be used in the data-driven predictive modeling.

[0061] The types of features generated in this step are the same as the types of features (e.g., (1) to (6) above) generated in step S13 included in the learning method according to this embodiment and in step S23 included in the prediction method according to this embodiment, so duplicated explanations will be omitted. Also, the generation unit 13 executes each process described in step S13 or step S23 above, but duplicated explanations of these processes will be omitted.

[0062] (Study Section 14) In the learning phase, the learning unit 14 uses one or more feature quantities related to the learning battery cell to learn a prediction model that predicts the operating characteristics of the battery cell. More specifically, as an example, the learning unit 14 executes the processes described in step S14 above. Duplicate descriptions of these processes will be omitted.

[0063] (Prediction Section 24) In the prediction phase, the prediction unit 24 predicts the operating characteristics of the battery cell to be predicted by inputting one or more feature quantities related to the battery cell to be predicted into a trained prediction model. Here, the trained prediction model can be a prediction model trained by the learning unit 14. More specifically, as an example, the prediction unit 24 executes each of the processes described in step S24 above. Duplicate descriptions of each of these processes will be omitted.

[0064] (Effects of the information processing system 100) According to the information processing device 1 configured as above, at least one of the feature quantities (1) to (6) described above is generated, and a prediction model is trained using the generated feature quantities, so that a prediction model that can appropriately predict the operating characteristics of a battery cell can be constructed. A data acquisition system that collects battery cell data for use in data-driven predictive modeling to predict operating characteristics of the battery cell. According to the data collection device of this embodiment, at least one of the feature quantities (1) to (6) described above is generated, and therefore it is possible to suitably collect data (learning data) for constructing a prediction model that can suitably predict the operating characteristics of a battery cell.

[0065] Furthermore, according to the information processing device 1 configured as described above, at least one of the features (1) to (6) described above is generated, and the generated feature is input into a trained prediction model, thereby outputting an index relating to the operating characteristics of the target battery cell, thereby enabling the operating characteristics of the target battery cell to be suitably predicted.

[0066] (Additional Notes Regarding Information Processing System 100) The control unit 10 of the information processing device 1 may have the functions of a monitoring unit that monitors the operating characteristics of the target battery cell and a current control unit that controls the amount of current supplied to the target battery cell based on the monitored operating characteristics. In other words, the control unit 10 of the information processing device 1 may be configured to include a monitoring unit that monitors the operating characteristics of the target battery cell and a current control unit that controls the amount of current supplied to the target battery cell based on the monitored operating characteristics.

[0067] Here, the monitoring unit may monitor the operating characteristics of the target battery cell via a measuring device 11 included in the information processing system 100. The measuring device 11 may be configured to include various detection devices for performing the monitoring. As an example, the current control unit may perform control such that, when the operating characteristics of the target battery cell exceed a predetermined allowable range, the amount of current supplied to the battery cell is reduced compared to before the operating characteristics exceeded a predetermined allowable range. The supply of current to the target battery cell may be performed via the measuring device 11 included in the information processing system 100. However, these examples do not limit the present embodiment.

[0068] With the above-described configuration, the information processing system 100 can appropriately monitor the operating characteristics of the target battery cell and appropriately control the current supplied to the target battery cell, thereby enabling the information processing system 100 to appropriately perform battery management for the target battery cell.

[0069] <Example> Below, an example using the above-described information processing system 100 will be described. Note that the matters described in the following example include matters obtained by processing by each unit included in the above-described information processing system 100. Therefore, the following example can be considered as a more specific explanation of the processing executed by each unit of the above-described information processing system 100.

[0070] (Summary of this Example) In general, because battery degradation is nonlinear, accurately predicting the cycle life of a battery is a challenging task. In the examples described below, a predictive model is constructed using machine learning, and predictions are made using the constructed predictive model. More specifically, an approach to predict the cycle life of a lithium metal-based secondary battery with a highly loaded Ni-rich NMC electrode is described. Here, the lithium metal-based secondary battery exhibits a more complex electrochemical profile during battery operation than the commonly studied LiFePO4 / graphite-based secondary battery.

[0071] In this example, various features are extracted from the discharge, charge, and relaxation processes of a battery cell, and complex battery behavior is predicted without relying on a specific degradation mechanism. As will be described later, the best predictive model that references multiple features selected from multiple feature candidates had an RMSE of 9.29 and an R 2 As a result, we achieved a correlation coefficient of 0.89, demonstrating the suitability of the prediction model for accurate prediction of battery cycle life. Logarithm of the minimum difference in discharge capacity between 100 cycles and 10 cycles (Log(|min(ΔDQ 100-10 (V))|) The importance of has become clear.

[0072] As will be discussed below, despite some inherent challenges, the predictive model exhibits a remarkable testing error of 6.6% on virgin data, demonstrating its robustness and potential for transformative advances in battery management systems. Thus, the techniques described herein contribute to the successful application of machine learning models in the area of ​​cycle life prediction for lithium metal-based secondary batteries in realistic energy-density designs.

[0073] (1: Background and Overview of the Present Example) Lithium-ion batteries (LIBs) are widely used as energy storage tools in various industries, including electric vehicles, portable electronic devices, and grid energy, due to their remarkable properties, such as energy density, low self-discharge rate, affordability, and long life. However, like many other electrochemical systems, LIBs inevitably degrade over time, resulting in a decrease in capacity and an increase in internal resistance. Therefore, accurately predicting the cycle life of LIBs can help industries optimize battery usage and replacement schedules, reducing unnecessary replacements and associated costs. Furthermore, evaluating battery quality in advance can identify potential problems and optimize battery design.

[0074] Battery degradation mechanisms are complex and degradation patterns are nonlinear, making battery life prediction difficult. Previous research has broadly divided battery life prediction methods into three categories: mechanism-based methods, model-based methods, and data-driven methods. Among these, data-driven methods using statistical data and machine learning (ML) algorithms have attracted considerable attention in recent years due to the advent of the big data era. Furthermore, several reports have been published in recent years on cycle life prediction for already mature and stable commercial LIBs.

[0075] Lithium metal batteries (LMBs) have attracted attention because of their high specific capacity (3860 mAh / g), lowest electrochemical potential (-3.04 V vs. a standard hydrogen electrode), and high energy density, which can extend the range of electric vehicles and improve performance in various energy-intensive applications. In fact, LMBs have been reported that, when combined with high-capacity Ni-rich NMC electrodes, achieve stable charge / discharge cycles for over 200 cycles and cell-level energy densities exceeding 350 Wh / kg. However, compared with conventional graphite-based lithium-ion batteries, LMBs have a lower redox potential, making them more susceptible to reductive decomposition of the electrolyte, resulting in complex lithium metal electrode degradation reactions. For example, dendritic growth of metallic lithium has been widely recognized as a critical issue for lithium metal electrodes, where needle-like structures form on the electrode surface during cycling, potentially leading to short circuits and battery failure. Recent intensive studies using various analytical techniques have revealed that the formation of isolated metallic lithium progresses during repeated cycling, resulting in significant volume expansion of the lithium metal electrode. As a result, the internal resistance increases significantly due to the lack of electrolyte. Furthermore, practical cell design conditions must also take into account chemical cross-reactions between electrodes. Therefore, these different degradation mechanisms, safety concerns, limited data, and the unique challenges posed by the complex degradation patterns of LMBs necessitate tailoring of ML-based cycle life prediction methods. Addressing these challenges requires innovative solutions to ensure the accuracy and reliability of prediction models.

[0076] Under these circumstances, the inventors worked on building a cell-level ML model for high-energy-density LMBs. In this example, 48 350Wh / kg cell-class LMBs were fabricated using high-mass-loading Ni-rich NMC electrodes. Thirty-five features were generated from the raw battery data of the first 100 cycles and classified into three groups related to charge, relaxation, and discharge. Furthermore, the linear regression model ElasticNet was used to predict the battery cycle life.

[0077] We also investigated the three feature groups separately. The correlation of the 35 features with the observed cycle life was systematically investigated by calculating the Pearson correlation coefficient. Analysis of the predictive performance of XGBoost showed that using XGBoost and selecting a feature subset containing six features resulted in a significant improvement in R. 2 The best prediction results were obtained with a mean error of 0.89 and RMSE of 8.29. The trained model was also used to predict the cycle life of eight new, unused batteries, achieving a best test error of 6.6%.

[0078] (2: Specific Contents of the Present Example) (2-1: LMB cell fabrication and battery performance testing) In this embodiment, LiNi, also known as NMC811 0.8 Mn 0.1 Co 0.1 A positive electrode (40 mm x 30 mm) consisting of O2 (mass loading: 30 mg / cm 2 ), Separator (6mm x 36mm), and Anode (42mm x 32mm) with a 50µm thick lithium layer on a 10µm thick copper (Cu) current collector A total of 57 pouch-type LMB cells (48 cells for model construction (training) and 9 cells for model testing (prediction)) were assembled. All cells were fabricated in a dry room (dew point < -50°C), and electrolyte injection was performed in a fume hood (dew point < -85°C). These battery cells are an example of the lithium metal-based battery cells described above, more specifically, an example of a lithium-ion battery with a lithium metal anode.

[0079] The cells were charged and discharged at 25°C using a Hokuto Denko HJ1001SD8. All cells were cycled at a constant current in the voltage range of 2 to 4.2 V. The voltage, current, and capacity of the LMB battery were continuously recorded during the cycling process, and charge / discharge curves and discharge capacity retention curves were obtained for every cycle. The complete cycle curve for a particular battery is shown in Figure 4. One cycle includes three processes: (1) a charging phase, (2) a relaxation phase after charging, and (3) a discharging phase.

[0080] (2-2: Feature configuration) The features used in this study were extracted from the raw voltage and capacity data during the entire cycle process. In addition to features generated from the discharge process, which have been commonly used in previous studies, new features generated from the charge and relaxation processes were also included for the first 100 cycles. The extracted features were classified into discharge-related features, relaxation-related features, and charge-related features, respectively.

[0081] (2-2-1: Discharge-related features) Seventeen features were generated from the discharge process. Six of these features were calculated from the discharge capacity-voltage curves (also called discharge capacity vs. voltage curves or discharge capacity vs. voltage data) as summary statistics, including the minimum value, variance, skewness, kurtosis, mean, and initial value of ΔDQ(V). The discharge capacity-voltage curves can capture the electrochemical changes of individual cells during cycling and reveal a wealth of information related to the battery degradation mechanism. The summary statistics were demonstrated to be effective in explaining changes in the shape and position of the voltage curves.

[0082] ΔDQ (V) indicates the difference in the discharge capacity-voltage curve between two cycles. As an example, using the interpolated discharge capacity that shows the change in discharge capacity in the discharge capacity-voltage curve between the 100th cycle and the 10th cycle, ΔDQ 100-10 (V) = DQ 100 (V) - DQ 10 (V) The upper part of Figure 5 shows the discharge capacity-voltage curves for cycle 100 (100th cycle) and cycle 10 (10th cycle) of a particular battery included in the data set used. The lower part of Figure 5 shows the ΔDQ of this battery. 100-10 (V) curves are shown.

[0083] In this example, the discharge capacities at cycle 2, cycle 10, and cycle 100 were extracted to quantify the energy output of the battery within the cycle. Furthermore, the discharge capacities at cycle 2, cycle 10, and cycle 100 were used to calculate the capacity retention (CR), which is a basic index of the discharge capacity at cycle 100 and cycle 10. Here, the capacity retention (CR) is, for example, the discharge capacity C at cycle n. Dch (n) Discharge capacity C at cycle n-1 Dch as a ratio to (n-1)

number

[0084] In this example, the feature quantities such as the slope and intercept of the discharge curve from the 2nd cycle to the 100th cycle, and the slope and intercept of the discharge curve from the 91st cycle to the 100th cycle were also calculated.

[0085] (2-2-2:Charging-related features) Twelve features were generated from the charging process. Six of these, as well as the discharge-related features, were calculated from the charge capacity-voltage curve (also called the charge capacity vs. voltage curve or charge capacity vs. voltage data) as summary statistics, including the minimum value, variance, skewness, kurtosis, mean value, and initial value of the difference ΔCQ between the charge capacity-voltage curves between two cycles. More specifically, the difference ΔCQ between the charge capacity-voltage curves was calculated as follows: ΔCQ = 1 / ... 100-10 (V) ΔCQ 100-10 (V) = CQ 100 (V) - CQ 10 (V) The six features were generated from the above. The top panel of Figure 6 shows the charge capacity-voltage curves for cycle 100 (100th cycle) and cycle 10 (10th cycle) of a specific battery included in the dataset used. The bottom panel of Figure 6 shows the discharge capacity, charge capacity, and coulombic efficiency of a representative battery as a function of cycle number.

[0086] In this example, the charge capacities at cycle 2, cycle 10, and cycle 100 were also included as feature quantities. The coulombic efficiencies (CE) at cycle 2, cycle 10, and cycle 100, calculated from the charge capacities at cycle 2, cycle 10, and cycle 100, were also selected as feature quantities for predicting the battery cycle life. The CE at cycle n is calculated by multiplying the measured discharge capacity C Dch (n) The measured charge capacity C of cycle n Ch As a ratio to (n),

number

[0087] Typically, in an idealized battery without side reactions, there is no loss in both the lithium transfer and electron transfer processes, and the coulombic efficiency (CE) reaches 1. The coulombic efficiency of the battery used in this example is shown in the bottom panel of Figure 6.

[0088] (2-2-3: Relaxation-related features) The relaxation process, encompassing voltage values ​​at specific time intervals and the voltage curve within a specified time frame, is known to correlate with the battery's State of Health (SoH). From the relaxation voltage-time curve shown in Figure 7, six features were generated, including the minimum, maximum, variance, skewness, kurtosis, and mean of the terminal voltage during the relaxation process from cycle 1 to cycle 100. More specifically, Figure 7 shows the relaxation voltage as a function of relaxation time from the first cycle to the last cycle (200 cycles) for a particular battery. The shading of the line is differentiated by the voltage at 600 seconds.

[0089] The 35 features constructed as described above are summarized in Table 1. [Table 1]

[0090] (2-3: Machine learning process) In this example, we used the entire dataset as the training dataset and a four-fold cross-validation strategy to generate the test dataset. In this approach, we divided the dataset into four distinct subsets, and then repeated training and testing the model four times. In each iteration, a different subset was used as the test set, and the remaining three subsets were used together as the training set. This four-fold cross-validation method allowed us to reliably evaluate the model's performance and gain a comprehensive understanding of its generalization ability. In this example, we specifically used two machine learning models (ML models): one was a linear regression model (ElasticNet) and the other was a nonlinear regression model (XGBoost).

[0091] We also evaluated the performance of the ML model using the following three statistical indicators: · Mean Absolute Error (MAE):

number

number

number

[0092] The mean absolute error (MAE) indicates the closeness of the prediction to each measurement, while the root mean square error (RMSE), which captures the variance of the error, is more sensitive to substantial deviations compared to the MAE. For these two values, a lower value indicates a better performance of the model, while a higher value indicates a poorer prediction. R 2 is a percentage metric, and in the best case scenario, R 2 approaches 100% or 1, indicating a strong match between observed and predicted values.

[0093] Furthermore, in this example, Pearson's correlation coefficient was used to express the relationship between the feature quantity and the cycle life of the battery. This coefficient is given by

number

[0094] (3:Result) In this example, a total of 48 LMB cells were fabricated and subjected to charge-discharge cycling tests. The discharge capacity of the 48 cells was plotted as a function of cycle number (see Figure 8). In Figure 8, each line corresponds to a battery, and the lines are shaded based on the discharge capacity at the end of the cycle life. The battery was defined as reaching the end of its life when the discharge capacity fell to 80% of its nominal capacity (maximum usable capacity).

[0095] The upper panel of Figure 9 shows a typical discharge profile (normalized discharge capacity curve) of one LMB cell (cell No. 100). The first discharge capacity of this cell was 0.86 mAh, with an average discharge voltage of 3.7 V. As cycling progressed, the cell maintained a capacity of over 0.68 mAh up to cycle 138. Thereafter, the discharge capacity gradually decreased, reaching 0.21 mAh at cycle 200. Most of the LMB cells examined in this study exhibited similar capacity profiles to cell No. 100, but some cell-to-cell variations were observed. For example, cell No. 35, which had a higher confining pressure than cell No. 100, exhibited capacity fluctuations after cycle 100 (see the lower panel of Figure 9). Indeed, overcharging occurred after cycle 100, which is a typical degradation mechanism for LMBs. Understanding the capacity degradation mechanism of these types of batteries is challenging because it is elusive and differs from real-world conditions. Therefore, in this study, we decided to exclude certain batteries from the dataset. As a result of the analysis, eight batteries were removed from the dataset, leaving a total of 40 batteries.

[0096] Next, individual battery cycle life predictions were performed using ElasticNet using the discharge-related features, charge-related features, and relaxation-related features. Figure 10 shows the prediction results for different feature subsets, starting from the top. More specifically, the top row of Figure 10 shows the prediction results for the discharge-related feature subset, the middle row of Figure 10 shows the prediction results for the charge-related feature subset, and the bottom row of Figure 10 shows the prediction results for the relaxation-related feature subset. The X and Y axes of Figure 10 represent the experimentally observed cycle life and the average ML-predicted cycle life after four-fold cross-validation, respectively. The top row of Figure 10 shows the prediction results using the discharge-related features as described above, with a test MSE of 9.81, RMSE of 13.54, and R 2 is 0.67, which shows the best predictive performance among the three machine learning methods. On the other hand, for the charging-related features (middle panel of Fig. 10) and relaxation-related features (bottom panel of Fig. 10), the parity plot is scattered, the MAE and RMSE are large, and R 2 However, even when using features related to discharge, as mentioned above, R 2 is equal to 0.67, which is not a satisfactory prediction performance. Therefore, in order to improve the prediction performance, we aimed to optimize the model from the aspects of feature values ​​and ML methods.

[0097] (Feature Aspects) For feature-related aspects, we mapped a heatmap matrix of Pearson correlation coefficients between variables in the dataset, including feature-to-feature and feature-to-target values. The heatmap matrix is ​​shown in the top panel of Figure 11. As shown in the top panel of Figure 11, the heatmap matrix is ​​a square matrix, with rows and columns representing feature values ​​and observed cycle life, and each cell containing the correlation coefficient between the corresponding variables. The correlation coefficient between two feature values ​​is an indicator of the magnitude and direction of the linear connection between those features, providing insight into how changes in one feature value correspond to changes in another. The most significant correlation in this example is the relationship between the observed cycle life and feature values ​​in the rightmost column of the correlation matrix. The coefficient values ​​provide a clear understanding of the correlation between them. Features with high coefficients are considered to be important predictors of cycle life.

[0098] The bottom of Figure 11 shows the Pearson correlation coefficient of the feature values ​​for the cycle life, which is the observation target value for 40 batteries. From the bottom of Figure 11, we can see that the coefficient of the discharge-related feature values ​​is relatively higher than the charge-related feature values ​​and the relaxation-related feature values. For example, the feature value with the highest correlation is ΔDQ 100-10 The logarithm of the smallest value of (V) has a negative Pearson correlation coefficient (r=-0.9) with the observed cycle life. The next most correlated feature is ΔDQ 100-10 On the other hand, some features, such as the variance of the discharge capacity at cycle 100, the coulombic efficiency at cycle 100, and the relaxation end voltage, only have a weak correlation with the observed cycle life.

[0099] In this example, several feature quantities of the battery cells included in the data set were selected, and the observed cycle life of the cells as a function of these feature quantities is shown in Figure 12. The top row of Figure 12 shows the ΔDQ 100-10The middle panel of FIG. 12 shows the cycle life as a function of the slope of the discharge capacity from cycle 2 to cycle 100, and the bottom panel of FIG. 12 shows the cycle life as a function of the dispersion of the relaxation voltage from the first cycle to cycle 100.

[0100] As shown in the upper part of Figure 12, the feature log(|min(ΔDQ 100-10 (V))|) shows a strong correlation, as is clear from the absolute value of Pearson's correlation coefficient (r = -0.90) exceeding |0.8|. The battery cycle life clearly has a linear relationship with the feature. Here, in this example, ΔDQ 100-10 It is worth emphasizing that not only was (V) extracted as a summary statistic, but that the base 10 logarithm of that value was used. In this example, when dealing with logarithmic functions, ΔDQ 100-10 As (V) approaches zero, the difference between the discharge capacity-voltage curves decreases, meaning the logarithm becomes more negative. As visually shown in the top row of Figure 12, the more negative the logarithm, the longer the battery's cycle life and the less noticeable the capacity degradation.

[0101] As shown in the middle panel of Figure 12, the correlation between the slope of discharge capacity from cycle 2 to cycle 100 and cycle life shows a moderate correlation (r = 0.53). The plot shows a gradual but significant upward trend. For batteries with short life spans, degradation tends to proceed relatively quickly, as evidenced by the steeper slope. For batteries with long cycle life spans, the slope is less negative compared to batteries with short life spans, showing a more gradual decline in degradation trajectory. However, analysis of the dataset identified several uncertainties regarding batteries with long cycle life spans, which may be explained by the moderate correlation between cycle life and the feature.

[0102] As shown in the bottom panel of Figure 12, the plot of cycle life as a function of the variance in relaxation voltage from the first cycle to cycle 100 shows data points that are more widely scattered than the two features mentioned above. This means that the observed cycle life is weakly correlated with the variance in relaxation voltage from the first cycle to the 100th cycle. Therefore, this particular feature is considered a poor predictor of battery cycle life. After investigation, we found that nine discharge-related features and three relaxation-related features showed strong or moderate correlations with battery cycle life. Meanwhile, all charge-related features showed weak correlations with battery cycle life. These 12 features were selected for use in the study described below.

[0103] (ML methodology aspects) Next, in terms of ML techniques, we implemented the ElasticNet and XGBoost ML algorithms separately for the 12 selected features to utilize their prediction performance. Figure 13 shows the prediction results using the 12 features by XGBoost and ElasticNet. The top panel of Figure 13 shows the parity plot of ElasticNet and XGBoost for the 12 features. The bottom panel of Figure 13 shows a histogram of the prediction results.

[0104] From the parity plot shown in the top panel of Figure 13, it is clear that the plots for ElasticNet are more scattered than those for XGBoost. Furthermore, from the prediction results shown in the bottom panel of Figure 13, it can be seen that the RMSE and MAE of XGBoost are slightly reduced, with 9.49 and 7.8, respectively, compared to ElasticNet, which has 12.06 and 8.8. Meanwhile, the R 2 The mean mean ,values ​​are 0.86 and 0.72, indicating that the nonlinear ML model ,XGBoost has better predictive performance for battery cycle life than ,ElasticNet.,Therefore, in further research in this example, we decided to build an ML model ,based on XGBoost.

[0105] In this example, exhaustive feature selection (EFS) was performed on the 12 features to eliminate potential feature overfitting. EFS is an approach that unbiasedly evaluates the optimal feature subset using specific evaluation metrics. This method ensures the evaluation of all possible combinations, ensuring comprehensive analysis without excessive computational cost. EFS generated a total of 4,095 feature combinations from the 12 features, and each combination was evaluated using XGBoost with 4-fold cross-validation to identify the optimal predictive performance for various numbers of features (n). The results are shown in Table 2. [Table 2] From Table 2, we can see that within the range of n = 3 (number of features = 3) to n = 10 (number of features = 10), there is little variability in the scores. However, outside this range, relatively large variability was observed for both n < 3 and n > 10. When n equals 6, the model provided the most accurate predictions, with R 2 The value of these six features was found to be 0.89. Log(|min(ΔDQ 100-10 (V))|) Log(|var (ΔDQ 100-10 (V))|) Intercept_DQ 91:100 ·CR 10:100 ·Slope_DQ 2:100 Mean ΔDQ 100-10 (V) The values ​​of these six features for each battery are shown in Table 3. [Table 3] The parity plot obtained using the set of six features above is shown in the top panel of Figure 14. As shown in the top panel of Figure 14, RMSE = 8.29 and MAE = 6.45. We also analyzed the importance of each feature for a subset of these six features. The bottom panel of Figure 14 shows the importance of each of these six features. Different features have different relative importance to the model. For this feature subgroup, ΔDQ 100-10 The logarithm of the minimum value of ΔDQ 100-10 The most important contribution is the discharge voltage-related feature, such as the logarithm of the variance of ΔDQ 100-10 It can be seen that the logarithm of the minimum value of is the most important feature.

[0106] On the other hand, the slope of the linear fit of the discharge capacity from the 2nd to 100th cycles and the average relaxation voltage contribute equally to the model, but are of the lowest relative importance. Furthermore, the capacity retention (CR) from the 10th to 100th cycles is more important as a feature than the other two. However, according to the Pearson correlation coefficient, the capacity retention (CR) has a weaker correlation with the observed cycle life than the other two. This indicates that the capacity retention (CR) has a high ability to reduce the error in the prediction results, but its linear relationship with the cycle life is weaker.

[0107] As described in this example, by using the exhaustive feature selection method and the advanced ML method XGBoost, a satisfactory ML model for LMB cycle life was realized.

[0108] Finally, to test whether the prediction model constructed in this example can accurately predict cycle life, eight new NMC811 / Li metal batteries were fed into the prediction model as unseen data.

[0109] Here, one cell was omitted due to its unstable capacity profile. The prediction results are shown in Figure 15. As shown in Figure 15, the MAE and RMSE of the test data are relatively small, with the test error (MAPE: Mean Absolute Percentage Error) equal to 6.6%. The achieved RMSE and MAE values ​​indicate that the prediction model according to this embodiment provides predictions with reasonable accuracy. The relatively low MAPE reinforces the accuracy of the prediction model, especially considering the deviation in percentage. As these indicators indicate, the performance of the model according to this embodiment is suitable for accurate cycle life prediction.

[0110] (summary) The use of machine learning modeling holds great potential for the diagnosis and prediction of LIBs, offering possibilities in various aspects, including their development, manufacturing, and optimization. In this study, we focused on using machine learning techniques to model the complex degradation mechanism of NMC811 / Li metal batteries and predict their cycle life through various feature quantities generated from various cycle processes. Forty-eight degradation data sets for NMC811 / Li metal batteries were recorded, and the features generated from the data were classified into three groups: discharge-related features, charge-related features, and relaxation-related features. First, we used the linear regression model ElasticNet for the different feature groups, but the prediction performance still has room for improvement.

[0111] In this study, we extracted 35 features with a strong or moderate correlation with cycle life and applied the nonlinear regression model XGBoost to battery cycle life prediction. Compared with the results of ElasticNet, XGBoost demonstrated superior performance in battery life prediction, with an RMSE of approximately 9.49 and an MAE of 7.8. After exhaustive feature selection, we found that six of the 12 features provided the best prediction results, reducing the RMSE to 8.29 and the MAE to 6.45, while increasing the R² to 0.89. Finally, by testing on unknown data, our best model achieved a test error of 6.6%. This demonstrates the suitability of the machine learning model in this study for LMB cycle life prediction.

[0112] (Additional information about ElasticNet and XGBoost) In the ElasticNet linear regression model, the relationship between observed and predicted battery cycle life is defined by the following equation:

number

number

[0113] On the other hand, the XGBoost technology is an adaptable and efficient tree boosting system designed for scalability, flexibility, and portability. XGBoost applies machine learning algorithms within the gradient boosting framework. In contrast to multiple regression, XGBoost excels at managing nonlinear relationships. The mathematical representation of a tree, denoted as f(x), is as follows:

number

[0114] The objective function is

number

number

[0115] The objective function can also be expressed as follows:

number

number

number

number

number

number

number

[0116] To prevent overfitting, the hyperparameters of the ElasticNet and XGBoost machine learning methods may be set using a five-fold cross-validation procedure. For each fold, four subsets are used to train the model, and the remaining subset is reserved for validation. The model is trained and evaluated five times, with each subset used once as the validation set. This iterative process allows for a comprehensive evaluation of the model's performance across different hyperparameter configurations and facilitates the selection of optimal hyperparameters that generalize well to unseen data.

[0117] (Notes on discharge-related features) As partially described above, in the above example, six feature quantities including the minimum value, variance, skewness, kurtosis, mean value, and initial value of ΔDQ were calculated as summary statistics from the discharge capacity-voltage curve as discharge-related feature quantities. More specifically, the following feature quantity group including these six feature quantities was calculated: Log(|min(ΔDQ 100-10 (V))|) Log(|mean(ΔDQ 100-10 (V))|) Log(|var(ΔDQ 100-10 (V))|) Log(|var(ΔDQ 100-10 (V) [0]|) Log(|skew(ΔDQ 100-10 (V))|) Log(|Kur(ΔDQ 100-10 (V))|) Log(|skew((V))|) Log(|Kur((V))|) The feature values ​​were normalized using logarithms as described above. The three feature values ​​were discharge capacities 2, 10, and 100, which represent the exact discharge capacities at the second, tenth, and 100th cycles. The slopes and intercepts of the linear fits to the capacity fade curves for cycles 2 to 100 and 91 to 100 were generated by using the discharge capacity / cycle life curve to calculate the slopes of the discharge capacities for cycles 2 to 100 and 91 to 100 (connecting two points to form a single line). The intercept feature was extracted by the intercept of each line on the Y-axis. The difference in maximum discharge capacity between the 100th and second cycles was calculated by subtracting the discharge capacity at the second cycle from the discharge capacity at the 100th cycle. The capacity retention ratios of 1:100, 2:1, and 99:100 were calculated using the following equations:

number

[0118] (Notes regarding charging-related features) As partially described above, in the above example, six feature quantities including the minimum value, variance, skewness, kurtosis, mean value, and initial value of ΔCQ were calculated as summary statistics from the charge capacity-voltage curve as charge-related feature quantities. Log(|min(ΔCQ 100-10 (V))|) Log(|mean(ΔCQ100-10 (V))|) Log(|var(ΔCQ 100-10 (V))|) Log(|var(ΔCQ 100-10 (V) [0]|) ·Log(|skew(ΔCQ 100-10 (V))|) ·Log(|Kur(ΔCQ 100-10 (V))|) The charge capacities for cycles 2, 10, and 100 are also similar to the discharge-related features. The coulombic efficiencies for cycles 2, 10, and 100 were calculated using the following formula:

number

[0119] (Notes on mitigation-related features) Relaxation-related features are also called voltage-related features. Voltage-related features include the minimum, maximum, variance, skewness, kurtosis, and mean of the terminal voltage from the relaxation voltage-time curve from cycle 1 to cycle 100. Here, skewness indicates the asymmetry of the distribution of values ​​in the dataset. Kurtosis indicates the peakedness or flatness of the distribution of values ​​in the dataset.

[0120] [Software implementation example] The functions of the information processing device 1 (hereinafter referred to as the "device") can be realized by a program that causes a computer to function as the device, and a program that causes a computer to function as each control block of the device (particularly each part included in the control unit 10).

[0121] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., a memory) as hardware for executing the program. The control device and storage device execute the program, thereby realizing the functions described in each of the above embodiments.

[0122] The program may be non-transitory and may be recorded on one or more computer-readable recording media. The recording media may or may not be included in the device. In the latter case, the program may be supplied to the device via any wired or wireless transmission medium.

[0123] Furthermore, some or all of the functions of the control blocks can be realized by logic circuits. For example, an integrated circuit in which a logic circuit that functions as each of the control blocks is formed is also included in the scope of the present invention. In addition, the functions of the control blocks can also be realized by, for example, a quantum computer.

[0124] (Additional notes) This specification describes at least the following configurations.

[0125] (Configuration A1) 1. A method for predicting the operating characteristics of a battery cell having a lithium metal negative electrode, comprising: an acquisition step of acquiring discharge capacity data for a plurality of charge / discharge cycles of the target battery cell; a prediction step of predicting the operating characteristics of the target battery cell by inputting one or more feature amounts obtained by referring to the discharge capacity data in the plurality of charge / discharge cycles into a prediction model; Including, The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the extreme value of the difference (minΔDQ 100-10 (V)), Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the variance of the difference (varΔDQ 100-10 (V)), and An intercept (intercept_DQ) obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles 91:100 ) A prediction method including a feature that depends on at least one of the above.

[0126] (Configuration A2) The one or more feature amounts include: As a feature dependent on the extreme value of the difference, the logarithm of the minimum value of the difference (log minΔDQ 100-10 (V)), As a feature dependent on the variance of the difference, the logarithm of the variance of the difference (log varΔDQ 100-10 (V)) The prediction device according to configuration A1 (Configuration A3) The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity (C Dch(n) ) and discharge capacity in other cycles (C Dch(n-1) ) ratio (CR), A slope (slope_DQ) obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles 2:100 ), and Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the mean difference (mean ΔDQ 100-10 (V) It also includes features that depend on at least one of the The prediction method according to configuration A1 or A2.

[0127] (Configuration A4) The one or more feature amounts include: As a feature dependent on the average of the differences, the logarithm of the average of the differences (log mean ΔDQ 100-10 (V)) The prediction method according to configuration A3.

[0128] (Configuration A5) The operating characteristics include at least one of a battery life and a logarithm of the battery life. The prediction method according to any one of A1 to A4.

[0129] (Configuration B1) 1. A data collection method for collecting battery cell data for use in data-driven predictive modeling for predicting operating characteristics of a battery cell having a lithium metal negative electrode, comprising: a measuring step of sequentially or continuously measuring one or more physical properties of each battery cell during a plurality of charge / discharge cycles for each of the one or more battery cells; acquiring discharge capacity data for each of the battery cells based on the one or more measured physical properties; a generating step of generating one or more feature quantities to be used in data-driven predictive modeling by referring to the discharge capacity data; Including, The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the extreme value of the difference (minΔDQ 100-10 (V)), Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the variance of the difference (varΔDQ 100-10 (V)), and An intercept (intercept_DQ) obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles 91:100 ) A data collection method including features that depend on at least one of the above.

[0130] (Configuration B2) The one or more feature amounts include: As a feature dependent on the extreme value of the difference, the logarithm of the minimum value of the difference (log minΔDQ 100-10 (V)), As a feature dependent on the variance of the difference, the logarithm of the variance of the difference (log varΔDQ 100-10 (V)) The data collection method described in configuration B1.

[0131] (Configuration B3) The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity (C Dch(n) ) and discharge capacity in other cycles (C Dch(n-1) ) ratio (CR), A slope (slope_DQ) obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles 2:100 ), and Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the mean difference (mean ΔDQ 100-10 (V) It also includes features that depend on at least one of the The data collection method according to configuration B1 or B2.

[0132] (Configuration B4) The one or more feature amounts include: As a feature dependent on the average of the differences, the logarithm of the average of the differences (log mean ΔDQ 100-10 (V)) The data collection method described in configuration B3.

[0133] (Configuration B5) The operating characteristics include at least one of a battery life and a logarithm of the battery life. The data collection method according to any one of configurations B1 to B4.

[0134] (Configuration B6) The method further includes a learning step of learning a prediction model that predicts the operating characteristics of the battery cell by referring to the one or more feature amounts. The data collection method according to any one of configurations B1 to B5.

[0135] (Configuration C1) 1. A method for training a predictive model for predicting operating characteristics of a battery cell having a lithium metal anode, comprising: an acquiring step of acquiring discharge capacity data for each of one or more battery cells in a plurality of charge / discharge cycles; a learning step of learning the prediction model using one or more feature amounts obtained by referring to discharge capacity data in the plurality of charge / discharge cycles; Including, The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the extreme value of the difference (minΔDQ 100-10 (V)), Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the variance of the difference (varΔDQ 100-10 (V)), and An intercept (intercept_DQ) obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles 91:100 ) A learning method that includes features that depend on at least one of the above.

[0136] (Configuration C2) The one or more feature amounts include: As a feature dependent on the extreme value of the difference, the logarithm of the minimum value of the difference (log minΔDQ 100-10 (V)), As a feature dependent on the variance of the difference, the logarithm of the variance of the difference (log varΔDQ 100-10 (V)) The learning method described in configuration C1.

[0137] (Configuration C3) The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity (C Dch(n) ) and discharge capacity in other cycles (C Dch(n-1) ) ratio (CR), A slope (slope_DQ) obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles 2:100 ), and Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the mean difference (mean ΔDQ 100-10 (V) It also includes features that depend on at least one of the The learning method according to structure C1 or C2.

[0138] (Configuration C4) The one or more feature amounts include: As a feature dependent on the average of the differences, the logarithm of the average of the differences (log mean ΔDQ 100-10 (V)) The learning method described in configuration C3.

[0139] (Configuration C5) The operating characteristics include at least one of a battery life and a logarithm of the battery life. A learning method according to any one of the configurations C1 to C4.

[0140] (Configuration C6) and further comprising predicting an operating characteristic of a target battery cell using the predictive model. A learning method according to any one of the structures C1 to C5.

[0141] (Configuration D1) A prediction device for predicting the operating characteristics of a battery cell having a lithium metal negative electrode, comprising: an acquisition unit that acquires discharge capacity data for a plurality of charge / discharge cycles of a target battery cell; a prediction unit that predicts the operating characteristics of the target battery cell by inputting one or more feature amounts obtained by referring to the discharge capacity data in the plurality of charge / discharge cycles into a prediction model; Equipped with The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the extreme value of the difference (minΔDQ 100-10 (V)), Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the variance of the difference (varΔDQ 100-10 (V)), and An intercept (intercept_DQ) obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles 91:100 ) A prediction device including a feature that depends on at least one of the above.

[0142] (Configuration D2) a monitoring unit that monitors the operating characteristics of the target battery cell; a current control unit that controls the amount of current supplied to the target battery cell based on the monitored operating characteristics; The prediction device according to configuration D1, further comprising:

[0143] (Configuration E1) 1. A data collection system for collecting battery cell data for use in data-driven predictive modeling for predicting operating characteristics of a battery cell having a lithium metal anode, comprising: a measurement unit that sequentially or continuously measures one or more physical properties of each battery cell during a plurality of charge / discharge cycles for each of the one or more battery cells; an acquisition unit that acquires discharge capacity data of each of the battery cells based on the one or more measured physical properties; a generation unit that references the discharge capacity data and generates one or more feature quantities to be used in data-driven predictive modeling; Including, The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the extreme value of the difference (minΔDQ 100-10 (V)), Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the variance of the difference (varΔDQ 100-10 (V)), and An intercept (intercept_DQ) obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles 91:100 ) A data collection system including features that depend on at least one of the above.

[0144] (Configuration F1) A learning device that learns a prediction model that predicts operating characteristics of a battery cell having a lithium metal negative electrode, an acquisition unit that acquires discharge capacity data for a plurality of charge / discharge cycles for each of one or a plurality of battery cells; a learning unit that learns the prediction model using one or more feature amounts obtained by referring to discharge capacity data in the plurality of charge / discharge cycles; Including, The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the extreme value of the difference (minΔDQ 100-10 (V)), Among the discharge capacity data in the plurality of charge / discharge cycles, the discharge capacity vs. voltage data (DQ n (V)) and discharge capacity vs. voltage data (DQ m (V)) and the variance of the difference (varΔDQ 100-10 (V)), and An intercept (intercept_DQ) obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles 91:100 ) A learning device including features that depend on at least one of the above.

[0145] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention. [Explanation of symbols]

[0146] 1. Information processing device 11 Measuring device 12...Acquisition part 13...Generation section 14 Learning Department 24 Prediction section

Claims

1. 1. A method for predicting the operating characteristics of a battery cell having a lithium metal negative electrode, comprising: an acquisition step of acquiring discharge capacity data for a plurality of charge / discharge cycles of the target battery cell; a prediction step of predicting the operating characteristics of the target battery cell by inputting one or more feature amounts obtained by referring to the discharge capacity data in the plurality of charge / discharge cycles into a prediction model; Including, The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, an extreme value of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; Among the discharge capacity data in the plurality of charge / discharge cycles, the variance of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; and An intercept obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles A prediction method including a feature that depends on at least one of the above.

2. The one or more feature amounts include: the feature dependent on the extreme value of the difference includes a logarithm of the minimum value of the difference; The logarithm of the variance of the difference is included as a feature dependent on the variance of the difference. The prediction method of claim 1 .

3. The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, a ratio of a discharge capacity in a certain cycle to a discharge capacity in another cycle; a slope obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles; and Among the discharge capacity data in the plurality of charge / discharge cycles, the average of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle It also includes features that depend on at least one of the The prediction method according to claim 2 .

4. The one or more feature amounts include: The logarithm of the average of the differences is included as a feature dependent on the average of the differences. The prediction method according to claim 3 .

5. The operating characteristics include at least one of a battery life and a logarithm of the battery life. The prediction method according to any one of claims 1 to 4.

6. 1. A data collection method for collecting battery cell data for use in data-driven predictive modeling for predicting operating characteristics of a battery cell having a lithium metal negative electrode, comprising: measuring one or more physical properties of each battery cell sequentially or continuously during a plurality of charge / discharge cycles for each of the one or more battery cells; acquiring discharge capacity data for each of the battery cells based on the measured one or more physical properties; a generating step of generating one or more feature quantities to be used in data-driven predictive modeling by referring to the discharge capacity data; Including, The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, an extreme value of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; Among the discharge capacity data in the plurality of charge / discharge cycles, the variance of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; and An intercept obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles A data collection method including features that depend on at least one of the above.

7. The one or more feature amounts include: the feature dependent on the extreme value of the difference includes a logarithm of the minimum value of the difference; The logarithm of the variance of the difference is included as a feature dependent on the variance of the difference. The data collection method according to claim 6.

8. The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, a ratio of a discharge capacity in a certain cycle to a discharge capacity in another cycle; a slope obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles; and Among the discharge capacity data in the plurality of charge / discharge cycles, the average of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle It also includes features that depend on at least one of the The data collection method according to claim 7.

9. The one or more feature amounts include: The logarithm of the average of the differences is included as a feature dependent on the average of the differences. The data collection method according to claim 8.

10. The operating characteristics include at least one of a battery life and a logarithm of the battery life.

10. The data collection method according to any one of claims 6 to 9.

11. The method further includes a learning step of learning a prediction model that predicts the operating characteristics of the battery cell by referring to the one or more feature amounts.

10. The data collection method according to any one of claims 6 to 9.

12. 1. A method for training a predictive model for predicting operating characteristics of a battery cell having a lithium metal anode, comprising: an acquiring step of acquiring discharge capacity data for each of one or more battery cells in a plurality of charge / discharge cycles; a learning step of learning the prediction model using one or more feature amounts obtained by referring to discharge capacity data in the plurality of charge / discharge cycles; Including, The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, an extreme value of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; Among the discharge capacity data in the plurality of charge / discharge cycles, the variance of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; and An intercept obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles A learning method that includes features that depend on at least one of the above.

13. The one or more feature amounts include: the feature dependent on the extreme value of the difference includes a logarithm of the minimum value of the difference; The logarithm of the variance of the difference is included as a feature dependent on the variance of the difference. The learning method according to claim 12.

14. The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, a ratio of a discharge capacity in a certain cycle to a discharge capacity in another cycle; a slope obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles; and Among the discharge capacity data in the plurality of charge / discharge cycles, the average of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle It also includes features that depend on at least one of the The learning method according to claim 13.

15. The one or more feature amounts include: The logarithm of the average of the differences is included as a feature dependent on the average of the differences. The learning method according to claim 14.

16. The operating characteristics include at least one of a battery life and a logarithm of the battery life.

16. A learning method according to any one of claims 12 to 15.

17. and further comprising predicting an operating characteristic of a target battery cell using the predictive model.

16. A learning method according to any one of claims 12 to 15.

18. A prediction device for predicting the operating characteristics of a battery cell having a lithium metal negative electrode, comprising: an acquisition unit that acquires discharge capacity data for a plurality of charge / discharge cycles of a target battery cell; a prediction unit that predicts the operating characteristics of the target battery cell by inputting one or more feature amounts obtained by referring to the discharge capacity data in the plurality of charge / discharge cycles into a prediction model; Equipped with The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, an extreme value of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; Among the discharge capacity data in the plurality of charge / discharge cycles, the variance of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; and An intercept obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles A prediction device including a feature that depends on at least one of the above.

19. a monitoring unit that monitors the operating characteristics of the target battery cell; a current control unit that controls the amount of current supplied to the target battery cell based on the monitored operating characteristics; The prediction device of claim 18 further comprising:

20. 1. A data collection system for collecting battery cell data for use in data-driven predictive modeling for predicting operating characteristics of a battery cell having a lithium metal anode, comprising: a measurement unit that sequentially or continuously measures one or more physical properties of each battery cell during a plurality of charge / discharge cycles for each of the one or more battery cells; an acquisition unit that acquires discharge capacity data of each of the battery cells based on the one or more measured physical properties; a generation unit that references the discharge capacity data and generates one or more feature quantities to be used in data-driven predictive modeling; Including, The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, an extreme value of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; Among the discharge capacity data in the plurality of charge / discharge cycles, the variance of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; and An intercept obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles A data collection system including features that depend on at least one of the above.

21. A learning device that learns a prediction model that predicts operating characteristics of a battery cell having a lithium metal negative electrode, an acquisition unit that acquires discharge capacity data for a plurality of charge / discharge cycles for each of one or a plurality of battery cells; a learning unit that learns the prediction model using one or more feature amounts obtained by referring to discharge capacity data in the plurality of charge / discharge cycles; Including, The one or more feature amounts include: Among the discharge capacity data in the plurality of charge / discharge cycles, an extreme value of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; Among the discharge capacity data in the plurality of charge / discharge cycles, the variance of the difference between the discharge capacity vs. voltage data in a certain cycle and the discharge capacity vs. voltage data in another cycle; and An intercept obtained by applying a linear regression process to at least a portion of the discharge capacity data in the plurality of charge / discharge cycles A learning device including features that depend on at least one of the above.

Images