Lithium battery micro internal short circuit fault diagnosis method based on dynamic mode decomposition and radial basis function neural network

Abstract

The invention discloses a lithium battery micro internal short circuit fault diagnosis method based on dynamic mode decomposition and a radial basis function neural network. Comprising the following steps: S1, preprocessing original voltage time sequence data collected in an operation process of a lithium battery, and constructing an input matrix suitable for dynamic mode decomposition; s2, decomposing the input matrix based on a dynamic mode decomposition method, extracting a dominant dynamic mode, screening and recombining a key mode based on Pearson correlation analysis, and generating a low-dimensional and high-sensitivity fault feature vector; s3, taking the feature vector as an input of a radial basis function neural network, taking an internal short circuit equivalent resistance value as a label, and constructing a nonlinear mapping relation between the feature space and an internal short circuit state; and S4, comparing an internal resistance prediction value output by the radial basis function neural network with a set threshold value or a reference value, calculating a prediction error and evaluating the diagnosis precision. According to the invention, misdiagnosis caused by noise and aging effects is avoided.

Description

Technical Field

[0001] This invention belongs to the field of micro-internal short circuit fault diagnosis of lithium-ion batteries, specifically involving a method for diagnosing micro-internal short circuit faults of lithium batteries based on dynamic mode decomposition and radial basis neural network. Background Technology

[0002] Lithium-ion batteries are batteries that use lithium metal or lithium alloy as the negative electrode material and a non-aqueous electrolyte solution, including mining power lithium batteries. With the continuous improvement of the electrification and intelligence of mining equipment, mining power lithium batteries, due to their advantages such as high energy density, high power output, and zero emissions, have been widely used in key mobile equipment such as underground locomotives, trackless rubber-tired transport vehicles, and auxiliary power supplies for tunneling machines, gradually replacing traditional lead-acid batteries or internal combustion power systems. However, the underground working environment in coal mines is characterized by enclosed spaces, limited ventilation, and the presence of flammable and explosive gases such as methane and coal dust, posing extremely stringent requirements for the safety of power lithium battery systems. In actual operation, mining power lithium batteries often face complex conditions such as frequent start-stop cycles, high-rate charging and discharging, mechanical vibration, and temperature fluctuations, which can easily induce micro-scale short-circuit faults inside the battery. These faults are usually caused by manufacturing defects, dendrite penetration of the separator, localized lithium plating, or long-term cycle aging, and are highly concealed and latent. In the early stages, internal short circuits only cause weak current anomalies, localized temperature rises, or voltage shifts. These signal characteristics are easily masked by normal operating noise, making them difficult to identify effectively using conventional monitoring methods. If an internal short circuit is not detected in time and continues to evolve, it will lead to localized overheating, triggering a thermal runaway chain reaction that can cause fires or even explosions. Given the confined space, difficult evacuation, and limited rescue capabilities of underground mines, the consequences of such safety accidents are particularly severe, threatening not only the lives of workers but also potentially causing significant production disruptions and property damage.

[0003] Currently, while diagnostic techniques for internal short-circuit faults in mining power lithium batteries have improved, significant shortcomings remain. These problems mainly focus on two aspects: First, existing methods largely rely on steady-state electrical parameter thresholds for judgment, making it difficult to adapt to the frequently changing load conditions and dynamic operating characteristics of batteries under mining conditions. Second, the initial fault characteristics of internal short circuits are weak, highly nonlinear, and have a low signal-to-noise ratio. The resulting voltage anomalies are easily coupled with the slow-changing characteristics generated during normal battery aging and complex operating noise. The lack of effective feature decoupling and extraction mechanisms makes it difficult to accurately distinguish between aging effects, noise interference, and true internal short-circuit fault characteristics, resulting in insufficient diagnostic sensitivity and a high rate of false alarms or missed alarms.

[0004] The announcement number "CN119986409B" provides a method and system for diagnosing battery micro-short circuit faults. It constructs a feature point matrix by denoising through variational mode decomposition, calculating dynamic reference voltage sequences, and extracting eigenvalues ​​and correlation coefficients. Then, it obtains anomaly scores based on improved Friesian distance and compares them with a threshold to achieve fault judgment. It has the advantages of reducing interference from inconsistent batteries, amplifying fault features through improved algorithms, high sensitivity to micro-short circuits, and no need for complex electrochemical mechanism models. It is also adaptable to various battery types. However, it also has the disadvantages of being able to only qualitatively determine whether there is a fault and not quantify the severity of the fault; having a single feature dimension, making it difficult to distinguish between battery aging effects and micro-short circuit faults; and having limited anti-interference capabilities.

[0005] Patent publication number "CN117849622A" discloses a battery internal short-circuit fault detection method based on variational mode decomposition and support vector machine. It collects voltage, current, and surface temperature data by establishing a charge-discharge model, calculates internal temperature to assist in diagnosis, performs variational mode decomposition on the voltage data to obtain modal components, calculates the sample entropy, and inputs it into the support vector machine model to achieve a qualitative judgment of whether a fault exists. It has advantages such as good noise robustness of variational mode decomposition, improved basic accuracy when combined with temperature-assisted diagnosis, and suitability of support vector machine for simple classification scenarios. However, it also has drawbacks, including only outputting a binary result of normal / fault, failing to quantify the severity of the fault, making it difficult to support fault evolution prediction in engineering applications; and relying on temperature data acquisition and calculation, increasing hardware deployment costs and data processing complexity.

[0006] Patent announcement number "CN120891398B" discloses a method and system for diagnosing internal short circuits in lithium batteries based on a dual-state cyclic equivalent circuit. This method constructs a hybrid model with a second-order RC equivalent circuit embedded in an RNN, combines an MLP subnet to fit the OCV-SOC relationship, trains the model using a weighted loss function and a composite optimization strategy, predicts the healthy voltage, and generates residuals. It extracts the DTW distance between the residual subsequence and the zero vector through a sliding window, constructs two-dimensional features using the mean and maximum values, and achieves fault identification based on unsupervised clustering using a Gaussian mixture model. This method features fusion... This approach, driven by physical mechanisms and data, can detect early internal short circuits without requiring fault samples, offering advantages such as interpretability, low computational complexity, and support for rapid fault classification. However, it also suffers from limitations, including a relatively singular feature dimension, relying solely on the mean and maximum values ​​of the DTW distance to construct two-dimensional features. This fails to fully exploit the multidimensional dynamic characteristics of voltage timing signals and results in insufficient identification of minute faults. Furthermore, in terms of model building, it depends on the physical constraints of a second-order RC equivalent circuit, making model performance highly susceptible to the accuracy of physical parameter identification and limiting its adaptability under complex operating conditions (such as frequent start-stop cycles and temperature fluctuations).

[0007] Therefore, how to achieve high sensitivity and high reliability in early diagnosis of internal short-circuit faults in power lithium batteries under conditions of unavoidable environmental noise and battery aging has become an urgent problem to be solved in order to ensure the safe operation of mining lithium battery power systems. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention proposes a lithium battery micro-internal short-circuit fault diagnosis method based on dynamic mode decomposition and radial basis neural network. This method can significantly improve the identification capability of weak internal short-circuit signals under the consideration of aging effects and actual noise environments, avoid misdiagnosis caused by aging effects and noise interference, and enhance the reliability and robustness of fault warning.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] This invention provides a method for diagnosing micro-internal short-circuit faults in lithium batteries based on dynamic mode decomposition and radial basis function neural networks, comprising the following steps:

[0011] S1. Preprocess the raw voltage time series data collected during the operation of the lithium battery to construct an input matrix suitable for dynamic mode decomposition;

[0012] S2. The input matrix is ​​decomposed based on the dynamic mode decomposition method to extract the dominant dynamic modes, and key modes are screened and recombined based on Pearson correlation analysis to generate a low-dimensional, highly sensitive fault feature vector.

[0013] S3. Using the feature vector as the input of the radial basis neural network, and using the internal short-circuit equivalent resistance value as the label, a nonlinear mapping relationship between the feature space and the internal short-circuit state is constructed.

[0014] S4. The predicted internal resistance value output by the radial basis neural network is compared with the set threshold or reference value to calculate the prediction error and evaluate the diagnostic accuracy, thereby realizing the quantitative identification and early warning of micro-internal short circuit faults in lithium-ion batteries.

[0015] As a preferred technical solution of the present invention: the preprocessing in step S1 includes normalization and resampling of the voltage time series data to eliminate dimensional differences and meet the input requirements of the dynamic mode decomposition algorithm; at the same time, the normalization method adopts the min-max normalization method to map the voltage data to the [0,1] interval to eliminate the influence of feature dimensional differences on feature extraction.

[0016] As a preferred technical solution of the present invention, step S1 is specifically as follows:

[0017] Preprocessing of raw voltage timing data:

[0018] The min-max normalization method is used to map the voltage data to the [0,1] interval to obtain the standardized voltage xi, thus eliminating the influence of feature dimension differences on feature extraction, as detailed below:

[0019] ;

[0020] Where xvi is the original voltage value at time i, and xmax and xmin are the maximum and minimum voltage values ​​of the battery during the entire cycle, respectively.

[0021] The standardized voltage xi is sampled uniformly to ensure data length consistency and meet the input requirements of the dynamic mode decomposition algorithm.

[0022] As a preferred embodiment of the present invention: the preprocessed voltage time series data in step S1 is used to construct the input matrix for dynamic mode decomposition using a sliding window method, and the window size and overlap rate of the sliding window are set. The input matrix generates a forward matrix. and backward matrix .

[0023] As a preferred technical solution of the present invention: For the preprocessed voltage time series data, a DMD input matrix is ​​constructed using a sliding window, with the window size set to 300 and the overlap rate to 70%, to generate a forward matrix. Backward matrix , where m = 2000 is the length of the downsampled data.

[0024] As a preferred technical solution of the present invention: the input matrix decomposed by the dynamic mode decomposition method in step S2 is to construct the input matrix of dynamic mode decomposition by means of a sliding window, so as to enhance the ability to characterize the dynamic changes in the battery during operation.

[0025] As a preferred technical solution of the present invention: the input matrix decomposition in step S2 is to perform singular value decomposition on the input matrix and truncate the decomposition result according to the singular value energy ratio, so as to filter out the interference of noise and aging effect while realizing data dimensionality reduction.

[0026] As a preferred technical solution of the present invention: the fault feature vector extracted in step S2 includes the dominant mode energy ratio, mode frequency characteristics, mode energy concentration degree, stable mode ratio, and signal reconstruction error.

[0027] As a preferred technical solution of the present invention: step S2 further includes performing correlation analysis on the fault feature vector and eliminating feature parameters whose correlation with the internal short-circuit equivalent resistance is lower than a preset threshold.

[0028] As a preferred technical solution of the present invention, step S2 is specifically as follows:

[0029] The input matrix is ​​decomposed based on the dynamic mode decomposition method;

[0030] Singular value decomposition (SVD) dimensionality reduction is performed as follows:

[0031] Perform singular value decomposition on the input matrix X. The decomposition results are truncated based on the cumulative energy proportion of singular values. When the cumulative energy proportion reaches a preset threshold, the corresponding small singular values ​​are discarded, and only the first few dominant singular values ​​and their corresponding low-dimensional matrix components are retained. This reduces the data dimensionality while suppressing noise interference and aging effects, and improves the retention of effective information.

[0032] Dynamic modality extraction and feature calculation are detailed below:

[0033] A low-dimensional dynamic evolution model is constructed based on the dimensionality-reduced data. The details are as follows:

[0034] ;

[0035] in Let represent the transpose matrix of the first r left singular vectors (corresponding to the spatial mode basis). Denotes the first r-order right singular vector matrix. This represents the inverse of the first r-order singular value diagonal matrix (reflecting modal energy);

[0036] Solve the dynamic evolution model eigenvalues The eigenvector W is used to calculate the corresponding dynamic modal information. The details are as follows:

[0037] ;

[0038] Multidimensional dynamic feature parameters that can characterize changes in the system's operating state are extracted from the dynamic modes, including the dominant mode energy ratio, mode frequency characteristics, mode energy concentration, stable mode ratio, and signal reconstruction error, to construct a low-dimensional fault feature vector with discriminative capabilities.

[0039] Feature validity screening and validation are as follows:

[0040] The Pearson correlation analysis method is used to evaluate the correlation between the dynamic characteristic parameters and the equivalent resistance of the internal short circuit. Invalid features with a correlation below a preset threshold of 0.3 are eliminated, and characteristic parameters that are significantly correlated with the degree of internal short circuit fault are retained to improve the accuracy and stability of subsequent fault diagnosis and prediction models.

[0041] As a preferred technical solution of the present invention: the radial basis function neural network in step S3 includes an input layer, a hidden layer and an output layer, wherein the hidden layer uses a radial basis function Gaussian function as the activation function.

[0042] As a preferred technical solution of the present invention: the radial basis neural network in step S3 is trained by supervised learning, and the accuracy and stability of the prediction of internal short-circuit equivalent resistance are improved by adjusting the network parameters.

[0043] As a preferred technical solution of the present invention, step S3 is as follows:

[0044] Sample construction and dataset partitioning are detailed below:

[0045] The dynamic modal feature vectors obtained in step S2 are paired with the corresponding internal short-circuit equivalent resistance values ​​to construct a dataset for model training and validation.

[0046] The dataset is randomly divided into training and testing datasets, and different internal short-circuit equivalent resistance levels are covered in each dataset to reduce the impact of sample distribution bias on model performance.

[0047] The radial basis function neural network is constructed as follows:

[0048] The radial basis function neural network model takes as input the dynamic mode feature vector extracted from the fault voltage and outputs as the predicted value of the internal short-circuit equivalent resistance.

[0049] In a radial basis function neural network, the nodes in the hidden layer use Euclidean distance to calculate the difference between the input and the center vector, and a Gaussian function is used as the activation function, as follows:

[0050] ;

[0051] in, The Euclidean distance between the input vector and the center vector. To expand the width;

[0052] The output of a radial basis function neural network is expressed as follows:

[0053] ;

[0054] Where x is the input feature vector, m is the number of hidden layer neurons, and w i c is the weighting value for the connection from the hidden layer to the output layer. i For the first i The center of each radial basis function Let be the width of the i-th radial basis function. Here, b is the radial basis function, and b is the output layer bias.

[0055] Model training and parameter optimization are implemented as follows:

[0056] Using the internal short-circuit equivalent resistance value as a supervised learning label, the radial basis function neural network is trained. By adjusting the model parameters such as the number of neurons in the hidden layer, the radial basis function expansion parameters, and the training termination condition, the network can achieve a balance between fitting accuracy and generalization ability. With an appropriate amount of samples, it can quickly and accurately approximate the internal short-circuit equivalent resistance prediction model.

[0057] In step S4, the accuracy of the micro-internal short circuit fault diagnosis results is evaluated by quantitatively analyzing the error between the predicted value and the reference value.

[0058] As a preferred technical solution of the present invention, step S4 is specifically as follows:

[0059] The predicted value of the internal short-circuit equivalent resistance output by the radial basis neural network in step S3 is compared with the corresponding reference resistance value. The deviation between the prediction result and the reference value is quantitatively analyzed by error calculation method to evaluate the accuracy of the model in predicting the internal short-circuit state.

[0060] Diagnostic accuracy assessment is implemented as follows:

[0061] The prediction results are comprehensively evaluated using multiple error evaluation indicators, including root mean square error, mean absolute error, mean absolute percentage error, and coefficient of determination. These indicators are used to quantitatively characterize the prediction accuracy and model stability of the internal short-circuit equivalent resistance, as detailed below.

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] in, This is the actual short-circuit resistance value. For predicted values, is the average of the true values, and n is the number of samples in the test set.

[0067] The proposed method for diagnosing micro-internal short-circuit faults in lithium batteries, based on dynamic mode decomposition and radial basis function neural networks, enables quantitative diagnosis, directly outputs predicted values ​​of the equivalent resistance of the internal short circuit, accurately characterizes the severity of the fault, and meets engineering prediction requirements. Multidimensional dynamic feature extraction combined with Pearson correlation analysis effectively isolates aging and noise interference, resulting in stronger robustness in noisy environments.

[0068] The lithium battery micro-internal short circuit fault diagnosis method proposed in this invention, based on dynamic mode decomposition and radial basis neural network, can perform quantitative diagnosis and distinguish the degree of fault, making it more applicable in the engineering field. It does not require temperature data assistance and can complete the diagnosis using only voltage time series data, reducing hardware costs and data processing pressure. The lightweight radial basis neural network has low computational cost and fast convergence, making it suitable for complex operating conditions.

[0069] The lithium battery micro-internal short-circuit fault diagnosis method proposed in this invention, based on dynamic mode decomposition and radial basis function neural network, achieves more comprehensive and in-depth feature extraction, overcoming the limitations of two-dimensional features. It extracts multi-dimensional dynamic features such as the dominant mode energy ratio and frequency characteristics through dynamic mode decomposition, and combines this with Pearson correlation analysis to screen effective features. This method better characterizes the essence of the fault than a single DTW distance feature, resulting in higher identification accuracy for minute faults. Purely data-driven with strong adaptability, it eliminates dependence on physical models: It does not require embedding equivalent circuits or relying on physical parameters for identification; diagnosis can be completed solely through voltage and timing data. Its adaptability to complex mining conditions (frequent start-stop, high-rate charging and discharging) far surpasses that of hybrid models that rely on physical constraints.

[0070] Beneficial effects:

[0071] This invention employs dynamic mode decomposition and Pearson correlation analysis to extract and filter features from lithium-ion battery charge-discharge voltage time-series data. It transforms the high-dimensional, non-stationary voltage signal, susceptible to aging effects and noise interference, into low-dimensional, physically meaningful key dynamic features. This effectively distinguishes between voltage changes caused by normal battery aging and operating noise and abnormal dynamic behaviors corresponding to micro-internal short-circuit faults. Furthermore, based on a radial basis function neural network (RBN) with global approximation capabilities, simple structure, and fast learning speed, an internal short-circuit equivalent resistance prediction model is established, enabling early and accurate identification and quantitative assessment of internal short-circuit faults. This method does not rely on complex equivalent circuit models or electrochemical mechanism models and maintains good robustness and diagnostic accuracy even under noisy operating data conditions. The lightweight model effectively reduces system implementation complexity, avoids misdiagnosis caused by aging effects and noise interference, and improves the reliability and economy of lithium-ion battery micro-internal short-circuit fault diagnosis. It is suitable for mining lithium battery applications with high safety and cost requirements. Attached Figure Description

[0072] Figure 1 This is a flowchart of the present invention;

[0073] Figure 2 This is a diagram showing the experimental results of feature extraction in this invention;

[0074] Figure 3 The figure shows the experimental results of the lithium battery micro-internal short circuit fault diagnosis method of the present invention;

[0075] Figure 4 The figure shows the experimental results of the lithium battery micro-internal short circuit fault diagnosis method of the present invention under a noisy environment. Detailed Implementation

[0076] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention:

[0077] like Figure 1-2 As shown, this invention provides a method for diagnosing micro-internal short-circuit faults in lithium batteries based on dynamic mode decomposition and radial basis function neural networks, comprising the following steps:

[0078] S1. Preprocess the raw voltage time series data collected during the operation of the lithium battery to construct an input matrix suitable for dynamic mode decomposition;

[0079] S2. The input matrix is ​​decomposed based on the dynamic mode decomposition method to extract the dominant dynamic modes, and key modes are screened and recombined based on Pearson correlation analysis to generate a low-dimensional, highly sensitive fault feature vector.

[0080] S3. Using the feature vector as the input of the radial basis neural network, and using the internal short-circuit equivalent resistance value as the label, a nonlinear mapping relationship between the feature space and the internal short-circuit state is constructed.

[0081] S4. The predicted internal resistance value output by the radial basis neural network is compared with the set threshold or reference value to calculate the prediction error and evaluate the diagnostic accuracy, thereby realizing the quantitative identification and early warning of micro-internal short circuit faults in lithium-ion batteries.

[0082] Step S1 is as follows:

[0083] Preprocessing of raw voltage timing data:

[0084] The min-max normalization method is used to map the voltage data to the [0,1] interval to obtain the standardized voltage xi, thus eliminating the influence of feature dimension differences on feature extraction, as detailed below:

[0085] ;

[0086] Where xvi is the original voltage value at time i, and xmax and xmin are the maximum and minimum voltage values ​​of the battery during the entire cycle, respectively.

[0087] The standardized voltage xi is sampled uniformly to ensure the consistency of data length and meet the input requirements of the dynamic mode decomposition algorithm.

[0088] For the preprocessed voltage time series data, a DMD input matrix is ​​constructed using a sliding window, with a window size of 300 and an overlap rate of 70%. A forward matrix is ​​then generated. Backward matrix Where m = 2000 is the length of the downsampled data;

[0089] Step S2 is as follows:

[0090] The input matrix is ​​decomposed based on the dynamic mode decomposition method;

[0091] Singular value decomposition (SVD) dimensionality reduction is performed as follows:

[0092] Perform singular value decomposition on the input matrix X. The decomposition results are truncated based on the cumulative energy ratio of singular values. When the cumulative energy ratio reaches a preset threshold, the corresponding small singular values ​​are discarded, and only the first few dominant singular values ​​and their corresponding low-dimensional matrix components are retained. This reduces the data dimensionality while suppressing noise interference and aging effects, and improves the retention of effective information.

[0093] Dynamic modality extraction and feature calculation are detailed below:

[0094] A low-dimensional dynamic evolution model is constructed based on the dimensionality-reduced data. The details are as follows:

[0095] ;

[0096] in Let represent the transpose matrix of the first r left singular vectors (corresponding to the spatial mode basis). Denotes the first r-order right singular vector matrix. This represents the inverse of the first r-order singular value diagonal matrix (reflecting modal energy);

[0097] Solve the dynamic evolution model eigenvalues The eigenvector W is used to calculate the corresponding dynamic modal information. The details are as follows:

[0098] ;

[0099] Multidimensional dynamic feature parameters that can characterize changes in the system's operating state are extracted from the dynamic modes, including the dominant mode energy ratio, mode frequency characteristics, mode energy concentration, stable mode ratio, and signal reconstruction error, to construct a low-dimensional fault feature vector with discriminative capabilities.

[0100] Feature validity screening and validation are as follows:

[0101] The Pearson correlation analysis method is used to evaluate the correlation between the dynamic characteristic parameters and the internal short-circuit equivalent resistance. Invalid features with a correlation below a preset threshold of 0.3 are eliminated, and characteristic parameters that are significantly correlated with the degree of internal short-circuit fault are retained to improve the accuracy and stability of subsequent fault diagnosis and prediction models.

[0102] Step S3 is as follows:

[0103] Sample construction and dataset partitioning are detailed below:

[0104] The dynamic modal feature vectors obtained in step S2 are paired with the corresponding internal short-circuit equivalent resistance values ​​to construct a dataset for model training and validation.

[0105] The dataset is randomly divided into training and testing datasets, and different internal short-circuit equivalent resistance levels are covered in each dataset to reduce the impact of sample distribution bias on model performance.

[0106] The radial basis function neural network is constructed as follows:

[0107] The radial basis function neural network model takes as input the dynamic mode feature vector extracted from the fault voltage and outputs as the predicted value of the internal short-circuit equivalent resistance.

[0108] In a radial basis function neural network, the nodes in the hidden layer use Euclidean distance to calculate the difference between the input and the center vector, and a Gaussian function is used as the activation function, as follows:

[0109] ;

[0110] in, The Euclidean distance between the input vector and the center vector. To expand the width;

[0111] The output of a radial basis function neural network is expressed as follows:

[0112] ;

[0113] Where x is the input feature vector, m is the number of hidden layer neurons, and w i c is the weighting value for the connection from the hidden layer to the output layer. i For the first i The center of each radial basis function Let be the width of the i-th radial basis function. Here, b is the radial basis function, and b is the output layer bias.

[0114] Model training and parameter optimization are implemented as follows:

[0115] Using the internal short-circuit equivalent resistance value as a supervised learning label, the radial basis function neural network is trained. By adjusting the model parameters such as the number of neurons in the hidden layer, the radial basis function expansion parameters, and the training termination condition, the network can achieve a balance between fitting accuracy and generalization ability. With an appropriate amount of samples, it can quickly and accurately approximate the internal short-circuit equivalent resistance prediction model.

[0116] Step S4 is as follows:

[0117] The predicted value of the internal short-circuit equivalent resistance output by the radial basis neural network in step S3 is compared with the corresponding reference resistance value. The deviation between the prediction result and the reference value is quantitatively analyzed by error calculation method to evaluate the accuracy of the model in predicting the internal short-circuit state.

[0118] Diagnostic accuracy assessment is implemented as follows:

[0119] The prediction results are comprehensively evaluated using multiple error evaluation indicators, including root mean square error, mean absolute error, mean absolute percentage error, and coefficient of determination. These indicators are used to quantitatively characterize the prediction accuracy and model stability of the internal short-circuit equivalent resistance, as detailed below.

[0120] (6);

[0121] (7);

[0122] (8);

[0123] (9);

[0124] in, This is the actual short-circuit resistance value. For predicted values, is the average of the true values, and n is the number of samples in the test set.

[0125] This invention provides a method for diagnosing micro-internal short-circuit faults in lithium batteries based on dynamic mode decomposition and radial basis function neural networks. By acquiring and processing voltage time-series data during battery charging and discharging, dynamic mode decomposition and Pearson correlation analysis are used to extract physically meaningful low-dimensional fault dynamic features from high-dimensional, non-stationary voltage signals affected by aging effects and noise interference. This allows for the identification of abnormal dynamic behaviors corresponding to voltage changes and micro-internal short-circuit faults caused by normal battery aging and operating noise. The features are then input into a radial basis function neural network to identify internal short-circuit faults and quantitatively predict equivalent resistance. This method does not rely on complex equivalent circuit models or electrochemical mechanism models and maintains good diagnostic accuracy and robustness even under actual conditions with noise and aging effects. It avoids misdiagnosis caused by aging effects and noise interference, making it suitable for mining lithium battery applications with high safety and cost requirements.

[0126] This invention proposes a method for diagnosing micro-internal short-circuit faults in lithium batteries based on dynamic mode decomposition and radial basis function neural networks. It mainly consists of three parts: dynamic mode decomposition feature extraction, radial basis function neural network modeling and training, and lithium battery internal short-circuit resistance prediction and diagnosis.

[0127] Dynamic mode decomposition can effectively extract features from lithium battery charge and discharge voltage time series data. Moreover, there are obvious differences in features between fault data and aging data. The method of this invention can effectively distinguish between internal short-circuit fault state and aging state.

[0128] Case 1: For example Figure 3 As shown, the experimental results of the lithium battery micro-internal short circuit fault diagnosis method based on dynamic mode decomposition and radial basis neural network are presented under internal short circuit conditions. The experimental results show that the predicted short circuit resistance is in high agreement with the actual results. The method of this invention achieves quantitative diagnosis of internal short circuit.

[0129] Case 2: For example Figure 4 As shown, the experimental results of the lithium battery micro-internal short circuit fault diagnosis method based on dynamic mode decomposition and radial basis neural network are presented under noisy conditions. The experimental results show that the method of the present invention can achieve accurate diagnosis of internal short circuits in lithium batteries under noisy conditions.

[0130] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for diagnosing micro-internal short-circuit faults in lithium batteries based on dynamic mode decomposition and radial basis function neural networks, characterized in that, Includes the following steps: S1. Preprocess the raw voltage time series data collected during the operation of the lithium battery to construct an input matrix suitable for dynamic mode decomposition; S2. The input matrix is ​​decomposed based on the dynamic mode decomposition method to extract the dominant dynamic modes, and key modes are screened and recombined based on Pearson correlation analysis to generate a low-dimensional, highly sensitive fault feature vector. S3. Using the feature vector as the input of the radial basis neural network, and using the internal short-circuit equivalent resistance value as the label, a nonlinear mapping relationship between the feature space and the internal short-circuit state is constructed. S4. The predicted internal resistance value output by the radial basis neural network is compared with the set threshold or reference value to calculate the prediction error and evaluate the diagnostic accuracy, thereby realizing the quantitative identification and early warning of micro-internal short circuit faults in lithium-ion batteries.

2. The lithium battery micro-internal short-circuit fault diagnosis method based on dynamic mode decomposition and radial basis function neural network according to claim 1, characterized in that, The preprocessing in step S1 includes normalization and resampling of the voltage time series data. The normalization method uses the min-max normalization method to map the voltage data to the [0,1] interval.

3. The lithium battery micro-internal short-circuit fault diagnosis method based on dynamic mode decomposition and radial basis function neural network according to claim 1, characterized in that, The input matrix decomposed in step S2 is constructed by using a sliding window method to construct the input matrix for dynamic mode decomposition.

4. The lithium battery micro-internal short-circuit fault diagnosis method based on dynamic mode decomposition and radial basis function neural network according to claim 1, characterized in that, The preprocessed voltage time series data in step S1 is used to construct the input matrix for dynamic mode decomposition using a sliding window. The window size and overlap rate of the sliding window are set to generate the forward matrix. and backward matrix .

5. The lithium battery micro-internal short-circuit fault diagnosis method based on dynamic mode decomposition and radial basis neural network according to claim 1, characterized in that, The input matrix decomposition in step S2 involves performing singular value decomposition on the input matrix and truncating the decomposition result based on the singular value energy ratio.

6. The lithium battery micro-internal short-circuit fault diagnosis method based on dynamic mode decomposition and radial basis function neural network according to claim 1, characterized in that, The fault feature vector extracted in step S2 includes the dominant mode energy ratio, mode frequency characteristics, mode energy concentration, stable mode ratio, and signal reconstruction error.

7. The lithium battery micro-internal short-circuit fault diagnosis method based on dynamic mode decomposition and radial basis function neural network according to claim 1, characterized in that, Step S2 also includes performing correlation analysis on the fault feature vector and removing feature parameters whose correlation with the internal short-circuit equivalent resistance is lower than a preset threshold.

8. The lithium battery micro-internal short-circuit fault diagnosis method based on dynamic mode decomposition and radial basis function neural network according to claim 1, characterized in that, The radial basis function neural network in step S3 includes an input layer, a hidden layer, and an output layer, wherein the hidden layer uses a radial basis function Gaussian function as the activation function.

9. The lithium battery micro-internal short-circuit fault diagnosis method based on dynamic mode decomposition and radial basis function neural network according to claim 1, characterized in that, The radial basis function neural network in step S3 is trained using supervised learning, and the accuracy and stability of the internal short-circuit equivalent resistance prediction are improved by adjusting the network parameters.

Images