Impedance spectrum-based cross-material system lithium ion battery internal short circuit diagnosis method and system under sample imbalance condition

Abstract

The invention belongs to the technical field of lithium ion battery fault diagnosis, and discloses a cross-material system lithium ion battery internal short circuit diagnosis method and system based on impedance spectroscopy under a sample imbalance condition, and the method comprises the steps: collecting and preprocessing battery impedance spectroscopy data; extracting robustness characteristics of cross-working-condition and cross-fault-degree of the preprocessed battery impedance spectrum data; and inputting the robustness characteristics into the trained diagnosis model to realize lithium ion battery internal short circuit fault diagnosis. According to the invention, accurate diagnosis of the short-circuit fault in the lithium ion battery is realized by using the electrochemical impedance spectroscopy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of lithium ion battery fault diagnosis, and is especially suitable for diagnosing model training data imbalance and cross-material system migration scenarios, and particularly relates to a sample imbalance condition-based cross-material system lithium ion battery internal short circuit diagnosis method and system based on impedance spectrum. BACKGROUND

[0002] Lithium ion batteries are widely used in new energy vehicles, energy storage systems, portable electronic devices and other fields due to their high energy density, long cycle life, no memory effect and other advantages. However, under the influence of long-term charge and discharge cycles, abuse (overcharge, overdischarge, high temperature) or manufacturing defects, lithium ion batteries are prone to internal short circuit faults. Internal short circuit not only causes battery performance to drop sharply, but also can cause thermal runaway, resulting in fire, explosion and other safety accidents. Therefore, early and accurate diagnosis of lithium ion battery internal short circuit faults is of great significance to ensure the safe operation of battery systems. As a non-destructive testing technology, electrochemical impedance spectroscopy can reflect the changes in microprocesses such as electrode / electrolyte interface reaction, ion transport and charge transfer inside the battery, providing rich feature information for battery fault diagnosis, and has become an important means of battery state monitoring and fault diagnosis. Currently, most internal short circuit fault diagnosis methods based on electrochemical impedance spectroscopy rely on machine learning models, which learn battery impedance spectrum data to establish a fault recognition model. However, the existing technology has the following shortcomings: 1. Data imbalance problem: In practical applications, the number of healthy lithium ion batteries is much larger than that of batteries with internal short circuit faults, resulting in an imbalance in the model training data set, with a large number of healthy battery impedance spectrum samples and a small number of short circuit battery impedance spectrum samples. Such imbalanced data can cause the trained diagnostic model to be biased towards identifying healthy samples, significantly reducing the accuracy of identifying short circuit fault samples, and making it difficult to meet the actual diagnostic needs. 2. Cross-material system migration problem: Existing diagnostic models are mostly trained on battery data of a specific material system. If they are to be applied to batteries of other material systems, a large amount of data needs to be collected for model training, which is time-consuming and labor-intensive, and difficult to meet the needs of rapid diagnosis in multiple scenarios.

[0003] Therefore, there is an urgent need for a method that can effectively train an internal short circuit fault diagnosis model under data imbalance conditions and realize model migration across material systems to solve the shortcomings of existing technology. SUMMARY

[0004] To address the issues of low model diagnostic accuracy and difficulty in model transfer across material systems caused by data imbalance in existing lithium-ion battery internal short-circuit fault diagnosis methods based on electrochemical impedance spectroscopy, this invention provides a cross-material system lithium-ion battery internal short-circuit diagnosis method and system based on impedance spectroscopy under sample imbalance conditions. This method achieves accurate diagnosis of lithium-ion battery internal short-circuit faults by utilizing electrochemical impedance spectroscopy (EIS).

[0005] To achieve the above objectives, the present invention provides the following solution: A method for diagnosing internal short circuits in lithium-ion batteries across material systems based on impedance spectroscopy under sample imbalance conditions, the method comprising: Collect and preprocess battery impedance spectrum data; Extract robustness characteristics of preprocessed battery impedance spectrum data across operating conditions and fault degrees; By inputting robust features into the trained diagnostic model, internal short-circuit fault diagnosis of lithium-ion batteries can be achieved.

[0006] Preferably, the method for extracting robustness characteristics of preprocessed battery impedance spectrum data across operating conditions and fault degrees includes: The preprocessed battery impedance spectrum data was subjected to DRT analysis using a regularized inversion algorithm to extract DRT features. Extract impedance spectrum derived features from the Nyquist plot of the impedance spectrum; By integrating DRT features with impedance spectrum-derived features, a 10-dimensional feature vector is formed.

[0007] Preferably, the method for performing DRT analysis on the preprocessed battery impedance spectrum data using a regularized inversion algorithm includes: The solution is transformed into minimizing the objective function: ; in, The impedance spectrum was measured experimentally. Angular frequency; For regularization parameters; For regularization operators; It is an ohmic resistor; For DRT functions; j The imaginary unit; Discretization: Discretization of relaxation time Discretized on a logarithmic scale , , To minimize the relaxation time, the integral is transformed into matrix multiplication: ; in, It is angular frequency. , The relaxation strength is ultimately solved through numerical optimization, using a logarithmic step size. , N It is the number of points after the curve is discretized.

[0008] Preferably, the extracted DRT features include: peak location features, minimum location features, and interval area features; Methods for extracting peak location features include: At a fixed temperature and SOC, an internal short-circuit fault causes the peak value of the maximum relaxation time of the DRT curve to shift to the lower left. The position of this peak value is extracted as a feature to distinguish the degree of fault. Specifically, the first derivative of the DRT curve is calculated. This derivative reflects the rate of change of relaxation strength with relaxation time, i.e., the slope of the curve. The slope of the curve is the critical point where the first derivative is 0 and the second derivative is less than 0. The expression is: ; Traverse the entire relaxation time interval Find all critical points that satisfy the conditions. Each critical point corresponds to a peak value on the DRT curve, denoted as . in, k = 1, 2, ..., n are the peak indexes, For the first k The relaxation time of each peak, For the corresponding relaxation strength, the relaxation time for all peak values. Sort the data and select the maximum value. The corresponding position is That is, the "peak position of maximum relaxation time"; Methods for extracting minimum value location features include: When the State of Charge (SOC) changes, the positions of the peak and minimum values ​​at high and mid frequencies on the DRT curve change. The coordinates of these peaks and minimum values ​​are extracted as compensation features for the SOC. Specifically: [The text then abruptly shifts to a different topic:] Selecting... Peak position within the range and The minimum value location near 10s is used as a feature; the selection criterion for the minimum value is: ; Methods for extracting interval area features include: As temperature increases, the area of ​​the DRT curve in the relaxation time interval of less than 10 seconds monotonically decreases. This area is extracted as a temperature-condition compensation feature. Specifically, features that meet the requirements are selected. All sampling points are denoted as a subset: ; For discrete points, the trapezoidal rule approximates the integral by connecting adjacent points with straight lines and summing the areas of the trapezoids. The formula is as follows: ; Where S is the area of ​​the "relaxation time less than 10s interval".

[0009] Preferred methods for extracting impedance spectral derived features from Nyquist plots of impedance spectra include: In the Nyquist plot, the key mathematical condition for the transition point where the low-frequency curve transitions from a "circular tail" to a "sloping extension" is: the point's - Zim When the value reaches a local minimum, in actual calculations, because the impedance data are discrete points, the "adjacent point difference method" is used for approximate judgment: traversing the low-frequency subset data, calculating the difference between adjacent points... Zim The change in value Δ- Zim =(- Zim ( k )-(- Zim ( k +1))), when the change changes from negative to positive, the corresponding point is the minimum point, denoted as P1=( Zre 1, - Zim 1) From the low-frequency subset data, select the impedance point corresponding to the point with the lowest frequency, and denote it as the lowest frequency point P2=( Zre 2, - Zim 2), Finally, the slope p for: .

[0010] Preferred methods for training diagnostic models include: A dual strategy of "variational autoencoder-based VAE sample generation + loss function optimization" is adopted, combined with deep neural network (DNN) to train the basic diagnostic model, and a well-trained diagnostic model is obtained.

[0011] Preferably, the VAE model architecture employs a 5-layer encoder and a symmetric decoder. The first 3 layers of the encoder are fully connected layers, and the fourth layer outputs the mean of the latent distribution. μ with standard deviation σ The fifth layer implements sample sampling through reparameterization; the decoder maps the latent vectors into 10-dimensional feature vectors through three fully connected layers; among them, and These are the weight vectors and bias vectors for each layer; specifically: Encoder structure: The first layer, in which d in The input feature dimension has a value of 10; d 1 represents the first layer's output dimension, where the ReLU activation function introduces a non-linear transformation: ; The second floor, d 2 represents the second output dimension, which further compresses features and enhances non-linear representation: ; The third layer, d 3 represents the third layer of output dimension, forming an abstract feature representation: ; The fourth layer also outputs the mean of the latent distribution. μ and standard deviation σ : ; ; in d z Given the potential space dimension, the activation function is defined as: ; The fifth layer uses reparameterization techniques to extract parameters from the latent distribution. Mid-sampled latent vector z : ; in, I It is the identity matrix; Decoder structure: The first layer of the decoder is symmetrical to the third layer of the encoder, further restoring the feature dimensions: ; The second layer of the decoder is symmetrical to the second layer of the encoder, further recovering the feature dimensions: ; The decoder directly outputs 10-dimensional reconstructed features from the third layer: ; Using a composite loss function: ;in, l _ recon To reconstruct the loss, l _ KL For KL divergence loss, l _ class This represents the probability that a virtual sample is identified as one of multiple classes. α , β, γThe weight hyperparameters are used to determine the optimal combination using a Bayesian optimization method.

[0012] The present invention also provides a cross-material system lithium-ion battery internal short circuit diagnostic system based on impedance spectroscopy under sample imbalance conditions. The system is used to implement the aforementioned method and includes: an acquisition module, an extraction module, and a diagnostic module. The acquisition module is used to acquire and preprocess battery impedance spectrum data; The extraction module is used to extract the robustness characteristics of the preprocessed battery impedance spectrum data across operating conditions and fault degrees. The diagnostic module is used to input robustness features into the trained diagnostic model to achieve internal short-circuit fault diagnosis of lithium-ion batteries.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: Robust feature extraction strategy: By combining DRT analysis with impedance spectrum derived features, a 10-dimensional feature vector is extracted across temperature, SOC, and fault degree. This solves the problem of low feature recognition caused by the high dimension of the original impedance spectrum and overlapping operating conditions, laying the foundation for improving diagnostic accuracy. Data Imbalance Solution: Use a VAE model with physical constraints to generate minority class samples, and combine it with composite loss function optimization to ensure the rationality of virtual samples, achieve dataset balance, avoid model bias, and improve the ability to identify healthy and faulty samples in a balanced manner. Low data dependency cross-system transfer: Based on model fine-tuning, transfer learning maintains the basic training model structure and fine-tunes parameters using only small sample data from the target domain. It eliminates the need to rebuild the model, significantly reducing the data requirements and development costs for cross-material system applications and improving the versatility of the method. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of an impedance spectroscopy-based method for diagnosing internal short circuits in lithium-ion batteries under sample imbalance conditions, according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the features of an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 like Figure 1 As shown, this invention provides a method for diagnosing internal short circuits in lithium-ion batteries across material systems based on impedance spectroscopy under conditions of sample imbalance, comprising the following steps: (I) Impedance Spectrum Data Acquisition and Preprocessing 1. Multi-scenario data collection (1) Laboratory data acquisition Lithium-ion batteries of the target type (including types with abundant data and types with scarce data) were selected as experimental subjects, covering different battery states (State of charge (SOC), internal short-circuit fault state) and different operating conditions (temperature) parameters: Fault state simulation: The external parallel resistor method is used to simulate internal short circuit faults of different severity. Equivalent short circuit resistance values ​​(such as 5Ω, 10Ω, 20Ω) are set to correspond to micro short circuit, moderate short circuit and severe short circuit respectively, to ensure the diversity of fault samples. Operating conditions coverage: Control the experimental temperature (e.g., 15℃, 30℃, 45℃) and battery SOC (e.g., 0%, 20%, 40%, 60%, 80%, 100%). Under each operating condition, first place the battery in a constant temperature chamber for 2-4 hours until the state stabilizes, and then test the impedance spectrum through an electrochemical workstation. The test parameters are set as follows: excitation current amplitude 1A, frequency range 0.01Hz-1000Hz, and 5 test points are collected every ten harmonics.

[0019] (2) Data collection in actual scenarios Battery impedance spectrum data are collected from actual operating scenarios or retired batteries to construct a real-world scenario dataset.

[0020] Data statistics: After removing outlier data (such as impedance spectrum abrupt changes caused by noise interference), the data of batteries with the same material system in the two datasets are merged together to form multiple datasets divided by material system. Each dataset contains laboratory data and data from actual scenarios.

[0021] 2. Data Preprocessing Noise Removal: Wavelet transform algorithm is used to denoise the original impedance spectrum data (real part Zre, imaginary part Zim) to eliminate noise caused by electromagnetic interference and equipment errors during the test. Data normalization: The real and imaginary parts of the denoised impedance spectrum are mapped to the [0,1] interval using the formula: ; in m The original data, m min , m max These are the minimum and maximum values ​​for the data in this dimension, respectively, to avoid the impact of differences in data volume on model training; Outlier filtering: Outliers in the preprocessed data are filtered and removed based on the 3σ principle to ensure the quality of the dataset.

[0022] (II) Robust Fault Feature Extraction To address the issues of high dimensionality in the original impedance spectrum and low feature recognition due to overlapping operating conditions, a combined approach of "Distribution of relaxation time (DRT) analysis + impedance spectrum derived features" is adopted to extract robust features across operating conditions and fault degrees. 1. DRT Analysis and Feature Extraction The core expression for the DRT curve is g(τ), where τ is the relaxation time (unit: s) and g(τ) is the impedance strength (unit: Ω) at the corresponding relaxation time.

[0023] The essence of DRT is to analyze the impedance spectrum. ( (where angular frequency is the frequency) represents different relaxation times. The corresponding impedance contribution integral: ; in, Ohmic resistance (such as electrolyte and current collector resistance); DRT function (core unknown, reflecting relaxation time) (corresponding impedance relaxation strength). j Imaginary unit. DRT analysis is performed on the preprocessed impedance spectrum data using a regularized inversion algorithm. This method does not require a pre-defined equivalent circuit structure and directly reflects changes in the battery's internal physicochemical processes (such as charge transfer and ion diffusion). The solution process is as follows: The solution is transformed into minimizing the objective function: ; Among them, the impedance spectrum measured in the experiment: ; Regularization parameter (balancing the accuracy of data fitting with the smoothness of the solution); Regularization operators (usually first- or second-order difference operators, constraining) (rate of change).

[0024] Discretization: Discretization of relaxation time Discretized on a logarithmic scale The integral is transformed into matrix multiplication: ; in It is angular frequency. ( (with a logarithmic step size), and finally solved through numerical optimization. , N It is the number of points after the curve is discretized.

[0025] The DRT curve is processed to extract the following features: Peak position characteristics: At a fixed temperature and SOC, an internal short-circuit fault causes the peak value of the maximum relaxation time of the DRT curve to shift to the lower left. This peak position is extracted as a feature to distinguish the degree of fault. The specific process is as follows: First, the first derivative of the DRT curve is calculated. This derivative reflects the rate of change of relaxation strength with relaxation time (i.e., the slope of the curve). The critical point where the slope of the curve changes from positive to negative, i.e., the first derivative is 0 and the second derivative is less than 0, is expressed as: ; Traverse the entire relaxation time interval Find all critical points that satisfy the above conditions. Each critical point corresponds to a peak value on the DRT curve, denoted as . (in( k = 1, 2, ..., n) represents the peak index. For the first k The relaxation time of each peak, (This corresponds to the relaxation strength). Relaxation time for all peak values. Sort the data and select the maximum value. The corresponding position is This refers to the "peak position of maximum relaxation time".

[0026] Minimum Location Features: When SOC changes, the positions of the peaks and minima at high and mid frequencies on the DRT curve change significantly. Therefore, these peak and minimum coordinates are extracted as compensation features for SOC. The specific extraction process is similar to that described above, selecting... Peak position within the range and The minimum value location near 10s is used as a feature; the selection criterion for the minimum value is: ; Interval area characteristic: As temperature rises, the area of ​​the DRT curve in the interval with a relaxation time of less than 10s decreases monotonically. This interval area is extracted as a temperature condition compensation feature. The specific process is as follows: First, select the intervals that meet the requirements... All sampling points are denoted as a subset: For discrete points, the trapezoidal rule approximates the integral by connecting adjacent points with straight lines and summing the areas of the trapezoids. The formula is: ; Where S is the area of ​​the "relaxation time less than 10s interval" (unit: Ω·s).

[0027] 2. Impedance spectroscopy-derived feature extraction Derived features were extracted from the Nyquist plot of the impedance spectrum, including the slopes of the imaginary and real parts of the impedance in the low-frequency range (0.01Hz-0.1Hz). p Reflecting ion diffusion capability, the slope decreases during internal short circuits. In the high-frequency range (100Hz-1000Hz), it reflects the resistance of the electrolyte and current collector; changes are not significant during faults but can serve as a basic reference. The specific process is as follows: Definition of a minimum point: In a Nyquist plot, a minimum point is a characteristic point where the low-frequency curve transitions from a "circular tail" to a "sloping extension." Its core mathematical condition is: the minimum point... Zim The value reaches a local minimum. In actual calculations, since the impedance data are discrete points, the "adjacent point difference method" is used for approximate judgment: traversing the low-frequency subset data, calculating the difference between adjacent points - Zim The change Δ Zim =(- Zim ( k )-(- Zim ( k +1))). When the change changes from negative to positive, the corresponding point is the minimum point, denoted as P1=( Zre 1, - Zim 1). From the low-frequency subset data, select the impedance point corresponding to the point with the lowest frequency, denoted as the lowest frequency point P2 = ( Zre 2, - Zim 2). Finally, the slope p for: ; 3. Feature Vector Construction The features from the above steps are concatenated sequentially to obtain a feature vector, which looks like this: ,like Figure 2 As shown, a 10-dimensional feature vector is formed, which serves as the input to the diagnostic model, significantly reducing the data dimensionality and improving feature recognition.

[0028] (III) Handling Data Imbalance and Training of Basic Diagnostic Models To address the imbalance between healthy samples (minority class) and faulty samples (majority class) in the dataset, a dual strategy of "sample generation based on variational autoencoder (VAE) + loss function optimization" is adopted, combined with deep neural network (DNN) training of the basic diagnostic model.

[0029] 1. Minority class sample generation based on VAE Innovative use of VAE modeling to specifically generate minority class samples (healthy or faulty) ensures the physical plausibility and diversity of virtual samples: VAE model architecture: It adopts a 5-layer encoder and a symmetric decoder. The first 3 layers of the encoder are fully connected layers (ReLU activation function), and the 4th layer outputs the mean of the latent distribution. μ with standard deviation σ Layer 5 uses reparameterization techniques ( z = μ + σ × ε , ε Sample sampling is performed using standard normally distributed noise; the decoder maps the latent vectors to 10-dimensional feature vectors (consistent with the true feature dimension) through three fully connected layers. The specific model architecture is as follows: First, the encoder structure: The first layer is as follows, among which d in The input feature dimension has a value of 10; d 1 represents the output dimension of the first layer, where the ReLU activation function introduces a nonlinear transformation. Wherein, and These are the weight vector and bias vector for each layer, respectively.

[0030] ; The second layer is as follows: d 2 represents the output dimension of the second layer, which further compresses features and enhances non-linear representation.

[0031] ; The third layer is as follows: d3 represents the third layer output dimension, forming an abstract feature representation.

[0032] ; The fourth layer also outputs the mean of the latent distribution. μ and standard deviation σ (To ensure) σ >0, activated using Softplus): ; ; in d z Given the potential space dimension, the activation function is defined as: ; The fifth layer uses reparameterization techniques to extract parameters from the latent distribution. Mid-sampled latent vector z : ; in I It is an identity matrix.

[0033] Next is the decoder structure: The first layer of the decoder is symmetrical to the third layer of the encoder, further restoring the feature dimensions as follows: ; The second layer of the decoder is symmetrical to the second layer of the encoder, further restoring the feature dimensions, as follows: ; The decoder directly outputs 10-dimensional reconstructed features from the third layer: .

[0034] A coarse neural network model is trained using an imbalanced dataset. The input is a generated dummy sample, and the output is the probability that the dummy sample belongs to the majority class. This probability is used as a term in the loss function during VAE training. Reducing this value makes the dummy sample more closely resemble the minority class. This addresses the problem of some existing sample generation methods lacking direction, resulting in lower sample quality.

[0035] Constraint design: Set physical constraints (each feature value of the virtual sample must be between the maximum and minimum values ​​of the features of the real sample); Loss function optimization: using a composite loss function ; in, l _ reconThe reconstruction loss is (minimizing the feature difference between virtual samples and real samples). l _ KL KL divergence loss (makes the latent distribution closer to the standard normal distribution, improving sample diversity). l _ class This represents the probability that a virtual sample is identified as one of multiple classes. α , β, γ For weight hyperparameters ( α + β + γ =1), and the optimal combination is determined using a Bayesian optimization method. The loss for each part is defined as follows: ; in, Indicates the first i The first virtual sample j One characteristic, This represents the corresponding features of the real sample.

[0036] ; in, d For the dimensions of latent features, and They represent the first k The standard deviation and mean of each latent feature.

[0037] Balanced dataset construction: The generated virtual healthy samples (making the ratio of the total number of healthy samples to the number of faulty samples close to 1:1) are merged with the original training set to obtain a balanced training set.

[0038] 2. DNN basic diagnostic model training (based on sufficient data) Model Architecture: A 5-layer DNN model is constructed. The input layer consists of 10-dimensional feature vectors, the first 4 layers are fully connected layers, and the output layer uses a sigmoid activation function (outputting a probability value between 0 and 1; ≥0.5 indicates an internal short-circuit fault, and <0.5 indicates a healthy condition). The specific architecture is as follows: A preliminary nonlinear transformation is performed on the input features, and LeakyReLU is introduced to solve the "dead neuron" problem of ReLU: ; The second layer further extracts higher-order features and accelerates training convergence: ; The third layer extracts core features, further focusing on specific dimensions: ; Core Feature Refinement at Layer 4: ; Layer 5 performs binary classification, using the sigmoid function for activation: ; ; Training parameter settings: Using the binary cross-entropy loss function: ; in, n The number of samples; y i For the first i The actual label of each sample p i For the first i The predicted probability of each sample. The optimizer is Adam, and the hyperparameters are optimized using a Bayesian optimization method; Model validation: The balanced dataset was divided into a training set and a validation set in a 7:3 ratio using stratified random sampling (to maintain a consistent ratio of healthy to faulty samples). Training was stopped when the F1 score on the validation set was ≥0.95, thus obtaining a basic diagnostic model for battery types with sufficient data (such as LFP).

[0039] (iv) Optimization of cross-material system transfer learning To achieve model transfer across different material systems and reduce the data requirements for the target system, a transfer learning strategy of "model fine-tuning" is adopted: 1. Target domain data preparation Collect small sample data (e.g., 30% of the sufficient data volume, including healthy and faulty samples) of the target material system (such as NCM, which is different from the well-datad LFP), and complete the preprocessing and feature extraction according to the methods in "Step (I) (II)" to obtain the target domain small sample dataset; 2. Pre-trained model transfer The source system basic model trained in "Step (3)" is used as a pre-trained model to ensure that the model structure remains unchanged. All parameters are used as initial values ​​and then unfrozen for retraining and fine-tuning. 3. Model fine-tuning training Using a small sample dataset of the target domain as input, the same loss function and optimizer as the base model are employed. Learning rate, number of training epochs, and other parameters are used as hyperparameters for Bayesian optimization. Training stops when the F1 score of the target domain validation set (split in a 7:3 ratio) is ≥0.90, thus obtaining a reliable internal short-circuit diagnostic model for the target material system.

[0040] (v) Fault diagnosis application The following procedure should be followed to diagnose the internal short circuit fault in the lithium-ion battery to be diagnosed: 1. Test the electrochemical impedance spectroscopy of the battery to be diagnosed (test parameters are consistent with "Step (I)"); 2. Preprocess the impedance spectrum data (denoise removal, normalization) and extract 10-dimensional feature vectors; 3. Select the corresponding basic diagnostic model or transfer diagnostic model based on the material system of the battery to be diagnosed.

[0041] This invention offers high diagnostic accuracy: In source material systems (such as LFP), the basic diagnostic model achieves a test set accuracy of ≥94.5%, precision of ≥97%, and recall of ≥92%, solving the problem of early fault omissions in existing models; it also boasts good cross-system adaptability: In target material systems (such as NCM), the transfer diagnostic model achieves a diagnostic accuracy of ≥90% using only 30% of the target domain data from the source material system, improving accuracy by more than 15% compared to models trained directly with small samples from the target domain; and it is highly practical: the method covers different material systems and operating conditions, and the testing process is simple (requiring only a conventional electrochemical workstation), making it directly applicable to lithium-ion battery safety monitoring in scenarios such as new energy vehicles and energy storage systems, and possessing potential for industrialization.

[0042] Example 2 The present invention also provides a cross-material system lithium-ion battery internal short circuit diagnostic system based on impedance spectroscopy under sample imbalance conditions. The system is used to implement the method described in Example 1. The system includes: an acquisition module, an extraction module, and a diagnostic module. The acquisition module is used to acquire and preprocess battery impedance spectrum data; The extraction module is used to extract the robustness characteristics of the preprocessed battery impedance spectrum data across operating conditions and fault degrees. The diagnostic module is used to input robustness features into the trained diagnostic model to achieve internal short-circuit fault diagnosis in lithium-ion batteries.

[0043] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for diagnosing internal short circuits in lithium-ion batteries across material systems based on impedance spectroscopy under conditions of sample imbalance, characterized in that, The method includes: Collect and preprocess battery impedance spectrum data; Extract robustness characteristics of preprocessed battery impedance spectrum data across operating conditions and fault degrees; By inputting robust features into the trained diagnostic model, internal short-circuit fault diagnosis of lithium-ion batteries can be achieved.

2. The method according to claim 1, characterized in that, Methods for extracting robustness characteristics of preprocessed battery impedance spectrum data across operating conditions and fault degrees include: The preprocessed battery impedance spectrum data was subjected to DRT analysis using a regularized inversion algorithm to extract DRT features. Extract impedance spectrum derived features from the Nyquist plot of the impedance spectrum; By integrating DRT features with impedance spectrum-derived features, a 10-dimensional feature vector is formed.

3. The method according to claim 2, characterized in that, Methods for performing DRT analysis on preprocessed battery impedance spectrum data using regularized inversion algorithms include: The solution is transformed into minimizing the objective function: ； in, The impedance spectrum was measured experimentally. Angular frequency; For regularization parameters; For regularization operators; It is an ohmic resistor; For DRT functions; j The imaginary unit; Discretization: Discretization of relaxation time Discretized on a logarithmic scale , , To minimize the relaxation time, the integral is transformed into matrix multiplication: ； in, It is angular frequency. , The relaxation strength is ultimately solved through numerical optimization, using a logarithmic step size. , N It is the number of points after the curve is discretized.

4. The method according to claim 3, characterized in that, The extracted DRT features include: peak location features, minimum location features, and interval area features; Methods for extracting peak location features include: At a fixed temperature and SOC, an internal short-circuit fault causes the peak value of the maximum relaxation time of the DRT curve to shift to the lower left. The position of this peak value is extracted as a feature to distinguish the degree of fault. Specifically, the first derivative of the DRT curve is calculated. This derivative reflects the rate of change of relaxation strength with relaxation time, i.e., the slope of the curve. The slope of the curve is the critical point where the first derivative is 0 and the second derivative is less than 0. The expression is: ； Traverse the entire relaxation time interval Find all critical points that satisfy the conditions. Each critical point corresponds to a peak value on the DRT curve, denoted as . in, k = 1, 2, ..., n are the peak indexes, For the first k The relaxation time of each peak, For the corresponding relaxation strength, the relaxation time for all peak values. Sort the data and select the maximum value. The corresponding position is That is, "the peak position of the maximum relaxation time"; Methods for extracting minimum value location features include: When the State of Charge (SOC) changes, the positions of the peak and minimum values ​​at high and mid frequencies on the DRT curve change. The coordinates of these peaks and minimum values ​​are extracted as compensation features for the SOC. Specifically: [The text then abruptly shifts to a different topic:] Selecting... Peak position within the range and The minimum value location near 10s is used as a feature; the selection criterion for the minimum value is: ； Methods for extracting interval area features include: As temperature increases, the area of ​​the DRT curve in the relaxation time interval of less than 10 seconds monotonically decreases. This area is extracted as a temperature-condition compensation feature. Specifically, features that meet the requirements are selected. All sampling points are denoted as a subset: ； For discrete points, the trapezoidal rule approximates the integral by connecting adjacent points with straight lines and summing the areas of the trapezoids. The formula is as follows: ； Where S is the area of ​​the interval with a relaxation time of less than 10s.

5. The method according to claim 2, characterized in that, Methods for extracting impedance spectrum-derived features from Nyquist plots of impedance spectra include: In the Nyquist plot, the key mathematical condition for the transition point where the low-frequency curve changes from a "circular tail" to a "sloping extension" is: the point's - Zim When the value reaches a local minimum, in actual calculations, because the impedance data are discrete points, the "adjacent point difference method" is used for approximate judgment: traversing the low-frequency subset data, calculating the difference between adjacent points... Zim The change in value Δ- Zim =(- Zim ( k )-(- Zim ( k +1))), when the change changes from negative to positive, the corresponding point is the minimum point, denoted as P1=( Zre 1,- Zim 1) From the low-frequency subset data, select the impedance point corresponding to the point with the lowest frequency, and denote it as the lowest frequency point P2=( Zre 2, - Zim 2), Finally, the slope p for: 。 6. The method according to claim 1, characterized in that, Methods for training diagnostic models include: A dual strategy of "variational autoencoder-based VAE sample generation + loss function optimization" is adopted, combined with deep neural network (DNN) to train the basic diagnostic model, and a well-trained diagnostic model is obtained.

7. The method according to claim 6, characterized in that, The VAE model architecture employs a 5-layer encoder and a symmetric decoder. The first 3 layers of the encoder are fully connected layers, and the fourth layer outputs the mean of the latent distribution. μ with standard deviation σ The fifth layer achieves sample sampling through reparameterization; The decoder maps the latent vectors into 10-dimensional feature vectors through three fully connected layers; where, and These are the weight vectors and bias vectors for each layer; specifically: Encoder structure: The first layer, in which d in The input feature dimension has a value of 10; d 1 represents the first layer's output dimension, where the ReLU activation function introduces a non-linear transformation: ； The second floor, d 2 represents the second output dimension, which further compresses features and enhances non-linear representation: ； The third layer, d 3 represents the third layer of output dimension, forming an abstract feature representation: ； The fourth layer also outputs the mean of the latent distribution. μ and standard deviation σ : ； ； in d z Given the potential space dimension, the activation function is defined as: ； The fifth layer uses reparameterization techniques to extract parameters from the latent distribution. Mid-sampled latent vector z : ； in, I It is the identity matrix; Decoder structure: The first layer of the decoder is symmetrical to the third layer of the encoder, further restoring the feature dimensions: ； The second layer of the decoder is symmetrical to the second layer of the encoder, further recovering the feature dimensions: ； The decoder directly outputs 10-dimensional reconstructed features from the third layer: ； Using a composite loss function: ;in, l _ recon To reconstruct the loss, l _ KL For KL divergence loss, l _ class This represents the probability that a virtual sample is identified as one of multiple classes. α , β, γ The weight hyperparameters are used to determine the optimal combination using a Bayesian optimization method.

8. A cross-material system lithium-ion battery internal short-circuit diagnostic system based on impedance spectroscopy under sample imbalance conditions, the system being used to implement the method described in any one of claims 1-7, characterized in that, The system includes: a data acquisition module, an extraction module, and a diagnostic module; The acquisition module is used to acquire and preprocess battery impedance spectrum data; The extraction module is used to extract the robustness characteristics of the preprocessed battery impedance spectrum data across operating conditions and fault degrees. The diagnostic module is used to input robustness features into the trained diagnostic model to achieve internal short-circuit fault diagnosis of lithium-ion batteries.

Images