Lithium battery internal short circuit diagnosis method based on interpretable machine learning

Abstract

The invention provides a lithium battery internal short circuit diagnosis method based on interpretable machine learning, and the method comprises the steps: collecting voltage, current and time data of a battery in a constant-current charging process under different internal short circuit degrees, and building an experiment data set; preprocessing and filtering the voltage data, calculating the ratio of the voltage increment to the capacity increment, and obtaining the IC curve characteristics of the capacity increment; extracting a battery state-of-charge-open-circuit voltage relationship, and establishing a battery equivalent circuit model for a constant current stage; estimating the state of charge of the battery through an extended Kalman filtering algorithm and battery charging data; extracting characteristic parameters from the IC curve, wherein the characteristic parameters comprise a peak value, a peak position and an area between peaks; time sequence features are extracted based on a state of charge estimation result, and a multi-dimensional feature space is constructed; carrying out standardization processing on the extracted features, and constructing a classification model; and a Shapril additive feature interpretation method is introduced to realize interpretability analysis of a classification model decision process.

Description

Technical Field

[0001] The present invention relates to the field of battery fault diagnosis, and in particular to a method for diagnosing internal short - circuit of lithium - ion batteries based on interpretable machine learning. Background Art

[0002] With the increasingly prominent problems of global energy shortage and environmental pollution, electric vehicles, as new energy transportation tools replacing fuel vehicles in the transportation field, have developed rapidly. The power battery is the core component of an electric vehicle, and its safety is directly related to the safety of the whole vehicle. In recent years, safety accidents caused by thermal runaway of power batteries have occurred frequently, and internal short - circuit is one of the main reasons for thermal runaway. Therefore, it is of great significance to develop an efficient and accurate internal short - circuit diagnosis method.

[0003] For existing battery internal short - circuit diagnosis algorithms, they are roughly divided into three categories: Model - based internal short - circuit diagnosis algorithms: Such algorithms rely on establishing an accurate battery model. However, due to the strong non - linear characteristics of the battery system and its susceptibility to environmental factors, it is difficult to establish a high - precision model applicable to different working conditions. Model - based internal short - circuit diagnosis algorithms, data - driven internal short - circuit diagnosis algorithms, and signal - based internal short - circuit diagnosis algorithms. Data - driven internal short - circuit diagnosis algorithms: Such algorithms mainly rely on machine learning and deep learning technologies to construct a diagnosis model through a large amount of data training. However, such methods have high requirements for the quality of feature extraction, and the interpretability of the model is poor, making it difficult to ensure the reliability of the diagnosis results. Signal - based internal short - circuit diagnosis algorithms: Such algorithms mainly identify abnormal features through signal processing technology for fault diagnosis. However, in complex working conditions, a single signal feature is difficult to accurately reflect the development process of internal short - circuit faults. Therefore, organically combining physical models with machine learning methods, giving full play to their respective advantages, and developing an interpretable internal short - circuit diagnosis method has important research value and application prospects. Summary of the Invention

[0004] To solve the above - mentioned technical problems, the present invention provides a method for diagnosing internal short - circuit of lithium - ion batteries based on interpretable machine learning, which realizes the early detection and hierarchical diagnosis of internal short - circuit faults by integrating multi - source features and interpretable machine learning technologies, and provides a reliable basis for the safety warning of the battery system.

[0005] The specific technical solution is as follows:

[0006] The main steps of a method for diagnosing internal short - circuit of batteries based on interpretable machine learning are as follows:

[0007] Step S1: Collect voltage, current, and time data of the battery during constant - current charging under different internal short - circuit degrees, and establish an experimental data set;

[0008] Step S2: Preprocess and filter the collected voltage data, calculate the ratio of voltage increment to capacity increment, and obtain the characteristic of the capacity increment IC curve;

[0009] Step S3: Extract the relationship between the state of charge SOC and the open-circuit voltage OCV of the battery through experiments, and establish an equivalent circuit model of the battery for the constant current stage; Estimate the state of charge SOC of the battery through the extended Kalman filter algorithm and battery charging data;

[0010] Step S4: Extract characteristic parameters from the IC curve, including peak value, peak position, and area between peaks; Extract time series characteristics based on the state of charge SOC estimation result, construct a multi-dimensional feature space, and realize the effective fusion of IC characteristics and SOC characteristics;

[0011] Step S5: Standardize the extracted features, and use the LightGBM algorithm of the lightweight gradient boosting machine algorithm to construct a classification model;

[0012] Step S6: Introduce the Shapley Additive exPlanations (SHAP) method to realize the interpretability analysis of the decision-making process of the classification model.

[0013] The present invention has the following beneficial effects:

[0014] The present invention has a high accuracy in diagnosing internal short circuits of batteries, and using interpretable machine learning in the field of industrial fault diagnosis helps to understand how the machine learning model makes decisions and whether it conforms to known physical knowledge and expert experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a block diagram of a multi-level classification algorithm for internal short circuit diagnosis;

[0016] Figure 2 is an equivalent circuit model for the constant current stage;

[0017] Figure 3 are IC curves with different short circuit degrees;

[0018] Figure 4 are SOCs with different short circuit degrees;

[0019] Figure 5 is a diagnostic result diagram;

[0020] Figure 6 is a shap analysis diagram. DETAILED DESCRIPTION OF THE INVENTION

[0021] To make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, the present invention adopts the following technical solutions.

[0022] The present invention provides a lithium battery internal short circuit diagnosis method based on interpretable machine learning, as Figure 1 shown, the main steps are as follows:

[0023] Step S1: Collect voltage, current and time data of the battery during constant current charging under different internal short circuit degrees, and establish an experimental data set;

[0024] Step S2: Preprocess and filter the collected voltage data, calculate the ratio of voltage increment to capacity increment, and obtain the capacity increment IC curve characteristics;

[0025] Step S3: Extract the relationship between the state of charge SOC and the open circuit voltage OCV of the battery, and establish an equivalent circuit model of the battery for the constant current stage; estimate the state of charge SOC of the battery through the extended Kalman filter algorithm and battery charging data;

[0026] Step S4: Extract characteristic parameters from the IC curve, including peak value, peak position, and area between peaks; extract time series characteristics based on the state of charge SOC estimation result, construct a multi-dimensional feature space, and realize the effective fusion of IC characteristics and SOC characteristics;

[0027] Step S5: Standardize the extracted features, and use the LightGBM algorithm to construct a classification model;

[0028] Step S6: Introduce the Shapley Additive Feature Explanation method (SHAP) to realize the interpretability analysis of the classification model decision-making process.

[0029] Further, the specific steps of step S2 are as follows:

[0030] Essentially, the IC curve is obtained by calculating the derivative of the charge-discharge capacity Q with respect to the voltage V, that is, the dQ / dV-V relationship curve. Through this mathematical processing, the voltage platform that is not easily recognized in the Q-V curve can be transformed into an obvious characteristic peak, so as to effectively reveal the electrochemical characteristic changes of lithium-ion batteries under different short circuit conditions. The specific formula for extracting the IC curve is as follows:

[0031] (1)

[0032] Where, \(I(t)\) represents the charging current, \(k\) represents the sampling point number, and \(T\) represents the interval time.

[0033] Due to the noise in the IC curve of numerical calculation and the inherent measurement noise and current fluctuation problems of the current sensor during signal acquisition, the IC curve cannot effectively characterize the characteristics under different short - circuit conditions. To solve this problem, the present invention adopts an optimized processing method combining Savitzky - Golay filtering and moving average filtering. The processing flow of this combined filtering strategy is as follows: First, perform Savitzky - Golay filtering on the original data, then perform moving average processing on the filtering result, and finally obtain high - quality filtered data.

[0034] For the time - series data \(x\), the core idea of Savitzky - Golay filtering is to perform polynomial fitting within a sliding window. Specifically, select \(2m + 1\) data points to form a window, and use a \(k\) - order polynomial to fit the data within the window. Its mathematical expression is shown in Equation (2).

[0035] (2)

[0036] Where, \(x\) is the value of the data point after fitting, \(t\) is the time - series index, are the coefficients of the \(k\) - order polynomial, and \(i\) is the order of the polynomial (\(i = 0,1,\cdots,k\)). This polynomial fitting is performed within a sliding window with a size of \(2m + 1\) data points.

[0037] The moving average filtering is used for signal processing, and its mathematical expression is shown in Equation (3).

[0038] (3)

[0039] Among them, is the data point after filtering, is the original data point, and \(N\) is the size of the sliding window.

[0040] Furthermore, the specific steps of step S3 are as follows:

[0041] Determine the functional relationship between SOC and OCV through the HPPC experiment of composite pulse power test. According to the charging characteristics of the battery in the constant - current stage, there is a pseudo - equilibrium state inside the battery, and a battery model in the constant - current stage is established.

[0042] (4)

[0043] In the above equation, represents the open - circuit voltage of the battery, \(I\) represents the charging current, represents the composite resistance, which depends on SOC and \(I\).

[0044] By simulating the internal short - circuit situation of the battery by paralleling a short - circuit branch to the equivalent - circuit model, formula (4) can be rewritten as:

[0045] (5)

[0046] Wherein, represents the short - circuit resistance, represents the current flowing through the short - circuit path. According to Kirchhoff's voltage law (KVL), the terminal voltage can be expressed as and the product of.

[0047] By combining the established formula (5) with the extended Kalman filter (EKF), the SOC of the battery under different short - circuit conditions can be estimated.

[0048] Furthermore, the specific steps of step S4 are as follows:

[0049] As the charging process progresses, the overall IC curve will form n peaks. Considering the usage habits of electric - vehicle users, the battery is usually not used until it has a very low charge or charged at a high charge. Select the middle peaks 1 and 2, and take the IC values, corresponding voltage values, and the area between the two peaks of peaks 1 and 2 for feature extraction.

[0050] Since battery short - circuit will cause battery power loss, it will take longer for the battery to charge to a certain level of power, and as the severity of the battery short - circuit increases, more time is required. Estimate the state of charge SOC of the battery under different short - circuit conditions through EKF, and extract the time from state of charge SOC 0.1 to 0.3, 0.1 to 0.5, 0.1 to 0.7, 0.1 to 0.9 as feature extraction.

[0051] Mix the features extracted from the IC curve and SOC to form a multi - dimensional feature matrix, and perform normalization processing to prepare for the next - step model input.

[0052] Furthermore, the specific steps of step S5 are as follows:

[0053] Input the data set constructed through the above steps into the LightGBM model, and quantify the degree of internal short - circuit by the magnitude of the internal short - circuit resistance. The larger the internal short - circuit resistance, the smaller the short - circuit degree. Select short - circuit resistances from 100 ohms to 200 ohms as severe short - circuit, 200 ohms to 400 ohms as moderate short - circuit, 400 to 1000 as mild short - circuit, and perform hierarchical diagnosis with the normal situation.

[0054] Furthermore, the specific steps of step S6 are as follows:

[0055] Through SHAP analysis, the interpretability analysis of the model decision-making process is realized, and the contribution degree of different features to the model prediction results is quantified, improving the interpretability and engineering practicability of the model. The SHAP value calculation formula is as follows:

[0056] (6)

[0057] In the formula, is the SHAP value of feature i, indicating the contribution degree of this feature to the model prediction; S is the feature subset that does not include feature i; |S| is the number of features in subset S; n is the total number of features; is the model prediction value when only using the feature subset S; is the model prediction value after adding feature i to subset S. This formula quantifies the importance of feature i by calculating its marginal contribution in all possible feature combinations.

[0058] More specifically, as Figure 1 shown, the multi-level classification method for internal short circuit diagnosis proposed by the present invention mainly includes data acquisition, feature extraction, model training and diagnosis. Among them, feature extraction includes IC curve feature extraction and SOC feature extraction, and the reliability of diagnosis is improved through feature fusion. The main steps are as follows:

[0059] Step S1: Collect voltage, current and time data of the battery during the constant current charging process under different short circuit degrees (100Ω - 1000Ω) and normal conditions. The sampling frequency is set to 1Hz, the charging rates are 0.3C, 0.4C, 0.5C, and the ambient temperature is controlled at 25±2°C.

[0060] Step S2: Extract the original IC curve from the voltage, current and time data of the constant current charging. Essentially, the IC curve is obtained by calculating the derivative of the charge-discharge capacity Q with respect to the voltage V, that is, the dQ / dV-V relationship curve. Through this mathematical processing, the voltage platforms that are not easily recognized in the Q-V curve can be transformed into obvious characteristic peaks, thereby effectively revealing the electrochemical characteristic changes of the lithium-ion battery under different short circuit conditions. The IC curve extraction formula is specifically as follows:

[0061] (1)

[0062] where I(t) is the charging current, k is the sampling point number, and T represents the interval time.

[0063] Due to the noise in the IC curve of numerical calculation and the inherent measurement noise and current fluctuation problems of the current sensor during signal acquisition, the IC curve cannot effectively characterize the characteristics under different short-circuit conditions. To solve this problem, the present invention adopts an optimized processing method combining Savitzky-Golay filtering and moving average filtering. The processing flow of this combined filtering strategy is as follows: First, perform Savitzky-Golay filtering on the original data, then perform moving average processing on the filtering result, and finally obtain high-quality filtered data.

[0064] For the time series data x, the core idea of Savitzky-Golay filtering is to perform polynomial fitting within a sliding window. Specifically, select 2m + 1 data points to form a window, and use a k-order polynomial to fit the data within the window. Its mathematical expression is shown in Equation (2).

[0065] (2)

[0066] In the formula, x is the value of the data point after fitting, t is the time series index, are the coefficients of the k-order polynomial, and i is the order of the polynomial (i = 0, 1,..., k). This polynomial fitting is performed within a sliding window, and the window size is 2m + 1 data points.

[0067] Moving average filtering is used for signal processing, and its mathematical expression is shown in Equation (3).

[0068] (3)

[0069] Among them, is the data point after filtering, is the original data point, and N is the sliding window size.

[0070] Step S3: Determine the functional relationship between SOC and OCV through the HPPC experiment of over-composite pulse power test. According to the charging characteristics of the battery during the constant current stage, there is a pseudo-equilibrium state inside the battery, and a battery model during the constant current stage is established. The equivalent circuit established during the constant current stage is as Figure 2 shown.

[0071] (4)

[0072] In the above equation, represents the open circuit voltage of the battery, I represents the charging current, represents the composite resistance, which depends on SOC and I.

[0073] By simulating the internal short-circuit situation of the battery by paralleling a short-circuit branch to the equivalent circuit model, and Equation (4) can be rewritten as:

[0074] (5)

[0075] Wherein, represents the short - circuit resistance, represents the current flowing through the short - circuit path. According to Kirchhoff's voltage law (KVL), the terminal voltage can be expressed as and the product of.

[0076] By combining the established formula (5) with the extended Kalman filter (EKF), the SOC of the battery under different short - circuit conditions can be estimated.

[0077] Step S4: As the charging process progresses, the overall IC curve will form n peaks. Considering the usage habits of electric vehicle users, the battery is usually not used until it reaches a very low state of charge or charged at a high state of charge. Select the middle peaks 1 and 2, and name the IC values of peaks 1 and 2 as PIC1 and PIC2. The corresponding voltage values are named PV1 and PV2. And the area PPA between the two peaks is used for feature extraction. As Figure 3 shown, it is the state of the IC curve under different short - circuit conditions.

[0078] Since battery short - circuit will cause battery power loss, it will take longer for the battery to charge to a certain state of charge, and as the severity of the battery short - circuit increases, more time is required. Estimate the SOC of the battery under different short - circuit conditions through EKF, Figure 4 shown is the SOC estimation under different short - circuit conditions. Extract the time from state of charge 0.1 to 0.3, 0.1 to 0.5, 0.1 to 0.7, 0.1 to 0.9 and name it T1, T2, T3, T4 for feature extraction.

[0079] Combine the extracted features of the IC curve and SOC into a multi - dimensional feature matrix [PIC1, PIC2, PPA, PV1, PV2, T1, T2, T3, T4] and perform normalization processing to prepare for the next model input.

[0080] Step S5: Construct a data set through the above steps. In the present invention, 0.2C, 0.3C, 0.4C are selected as the constant - current input, and the short - circuit resistance from 100 ohms to 1000 ohms is selected, with an interval of 100 ohms to simulate different degrees of short - circuit. The input multi - dimensional feature matrix is input into the LightGBM model, and the degree of internal short - circuit is quantified by the magnitude of the internal short - circuit resistance. The larger the internal short - circuit resistance, the smaller the degree of short - circuit. Select 100 ohms to 200 ohms of short - circuit resistance as severe short - circuit, 200 ohms to 400 ohms as moderate short - circuit, 400 to 1000 as mild short - circuit, and perform hierarchical diagnosis with normal conditions.

[0081] For the convenience of training and verification, three 18650 batteries are selected in the present invention for testing and verification. The diagnostic results are as Figure 5 shown. In the multi-class diagnosis of internal short circuits, the diagnostic accuracy rate of the present invention reaches 98.61%.

[0082] Step S6: Through SHAP analysis, realize the interpretability analysis of the model decision-making process, and quantify the contribution degree of different features to the model prediction result, so as to improve the interpretability and engineering practicability of the model. The SHAP value calculation formula is as follows:

[0083] (6)

[0084] In the formula, is the SHAP value of feature i, indicating the contribution degree of this feature to the model prediction; S is the feature subset that does not include feature i; |S| is the number of features in subset S; n is the total number of features; is the model prediction value when only using the feature subset S; is the model prediction value after adding feature i to subset S. This formula quantifies the importance of feature i by calculating its marginal contribution in all possible feature combinations.

[0085] As Figure 6 shown is the interpretability analysis of the model of the present invention. It can be clearly seen from Figure 6 (a) that T4 makes the greatest contribution to the diagnostic result, which also conforms to the experience that due to battery short circuit, the battery power will be lost, and it will take longer for the battery to charge to a certain power. It can be clearly seen from Figure 6 (b) that the features of the model are jointly affected by the features of the IC curve and the features of SOC.

Claims

1. A method for diagnosing internal short circuits in lithium batteries based on interpretable machine learning, characterized in that, The steps are as follows: Step S1: Collect voltage, current, and time data of the battery during constant-current charging under different internal short-circuit degrees, and establish an experimental data set. Step S2: Preprocess and filter the collected voltage data, calculate the ratio of voltage increment to capacity increment, and obtain the capacity increment IC curve characteristics. Step S3: Extract the relationship between the state of charge SOC and the open-circuit voltage OCV of the battery, and establish an equivalent circuit model of the battery for the constant-current stage; estimate the state of charge SOC of the battery through the extended Kalman filter algorithm and battery charging data. Step S4: Extract characteristic parameters from the IC curve, including peak value, peak position, and area between peaks; extract time series characteristics based on the state of charge SOC estimation result, construct a multi-dimensional feature space, and realize the fusion of IC characteristics and SOC characteristics. Step S5: Standardize the extracted features, and use the LightGBM algorithm to construct a classification model. Step S6: Introduce the Shapley additive feature explanation method to realize the interpretability analysis of the classification model decision-making process.

2. The method for diagnosing internal short circuit of a lithium battery based on interpretable machine learning according to claim 1, characterized in that, The specific steps of Step S2 are as follows: The IC curve is obtained by calculating the derivative of the charge-discharge capacity Q with respect to the voltage V, that is, the dQ / dV-V relationship curve. The specific formula for extracting the IC curve is as follows: （1） In the formula, I(t) represents the charging current, k represents the sampling point number, and T represents the interval time.

3. The method for diagnosing internal short circuit of a lithium battery based on interpretable machine learning according to claim 2, wherein In Step S2, the Savitzky-Golay filter and the moving average filter are combined to process the data. The processing flow is as follows: For the time series data x, perform Savitzky-Golay filtering, and perform polynomial fitting within the sliding window. Specifically, select 2m + 1 data points to form a window, and use a k-order polynomial to fit the data within the window. Its mathematical expression is shown in Equation (2): （2） where x is the value of the data point after fitting, t is the time series index, are the coefficients of the k-th order polynomial, i is the order of the polynomial (i = 0, 1, ..., k), and the polynomial fitting is performed within a sliding window with a window size of 2m + 1 data points; The moving average filter is used for signal processing, and its mathematical expression is shown in Equation (3): （3） Among them, is the filtered data point, is the original data point, and N is the sliding window size.

4. A method for diagnosing internal short circuit of a lithium battery based on interpretable machine learning according to claim 1, characterized in that, The specific steps of Step S3 are as follows: Determine the functional relationship between SOC and OCV through the hybrid pulse power characterization HPPC experiment. According to the battery charging characteristics in the constant-current stage, there is a pseudo-equilibrium state inside the battery, and a battery model for the constant-current stage is established: （4） In the above equation, represents the open-circuit voltage of the battery, and I represents the charging current, represents the composite resistance, which depends on SOC and I; Simulate the internal short-circuit situation of the battery by connecting a short-circuit branch in parallel to the equivalent circuit model, and Equation (4) is rewritten as: （5） Among them, represents the short-circuit resistance, represents the current flowing through the short-circuit path; Combined with the extended Kalman filter through the established Equation (5), the SOC of the battery under different short-circuit conditions can be estimated.

5. A method for diagnosing internal short circuit of a lithium battery based on interpretable machine learning according to claim 1, characterized in that, The specific steps of Step S4 are as follows: The IC curve forms n wave peaks as a whole. Select any two wave peaks in the middle, and take the IC values of the two wave peaks, the corresponding voltage values, and the area between the two peaks for feature extraction.

6. The method for diagnosing internal short circuit of a lithium battery based on interpretable machine learning according to claim 5, wherein, Estimate the state of charge SOC of the battery under different short-circuit conditions through the Kalman filter, and extract the time from the state of charge SOC0.1 to 0.3, 0.1 to 0.5, 0.1 to 0.7, and 0.1 to 0.9 as feature extraction.

7. The method for diagnosing internal short circuit of a lithium battery based on interpretable machine learning according to claim 6, wherein Mix the extracted features of the IC curve and SOC to form a multi-dimensional feature matrix, and perform normalization processing.

8. A method for diagnosing internal short circuit of a lithium battery based on interpretable machine learning according to claim 1, characterized in that, The specific steps of Step S5 are as follows: The constructed dataset is input into the LightGBM model and quantified by the magnitude of the internal short-circuit resistance within the degree of internal short circuit.

9. The method for diagnosing internal short circuit of a lithium battery based on interpretable machine learning according to claim 1, wherein The specific steps of step S6 are as follows: The calculation formula for Shapley additive feature values is as follows: （6） In the formula, is the SHAP value of feature i, representing the contribution of this feature to the model prediction; S is the feature subset that does not include feature i; |S| is the number of features in subset S; n is the total number of features; is the model prediction value when only using feature subset S; is the model prediction value after adding feature i to subset S. This formula quantifies the importance of feature i by calculating its marginal contribution in all possible feature combinations.

10. A method for diagnosing internal short circuit of a lithium battery based on interpretable machine learning according to claim 8, characterized in that, Define a short-circuit resistance of 100 ohms to 200 ohms as a severe short circuit, 200 ohms to 400 ohms as a moderate short circuit, 400 to 1000 as a slight short circuit, and perform hierarchical diagnosis with the normal situation.

Images