Lithium battery internal short circuit diagnosis method and system based on double-state circulation equivalent circuit

Abstract

The lithium battery internal short circuit diagnosis method and system based on the double-state cycle equivalent circuit of the application relate to the lithium battery internal short circuit detection technical field. Firstly, a hybrid model of a battery second-order RC equivalent circuit embedded in an RNN is constructed, the state update is constrained to conform to a physical equation through customization of a cycle unit, and an MLP subnet is used to fit an OCV-SOC relationship; a voltage interval weighted loss function is designed to strengthen the accuracy of a sensitive area, and learning rate decay and gradient clipping are combined to optimize physical parameters and MLP parameters; a trained model is used to predict a healthy voltage and generate a residual error; a sliding window is used to extract a residual error subsequence, the DTW distance of the residual error subsequence and a zero vector is calculated, and a mean value and a maximum value are used to form a two-dimensional feature; and automatic fault identification is realized based on Gaussian mixture model unsupervised clustering. The application combines the advantages of physical mechanisms and data driving, can detect early internal short circuits with high precision without fault samples, and solves the problems of high false alarm rate of the threshold method, poor adaptability of the model method, and dependence of the data driving on labeled samples.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium battery internal short circuit detection, in particular to a lithium battery internal short circuit diagnosis method and system based on a double-state cycle equivalent circuit. BACKGROUND

[0002] As the core energy carrier of electric vehicles and energy storage systems, the safety of lithium-ion batteries is directly related to life and property safety and the sustainable development of the industry. Internal short circuit (ISC) is the main cause of thermal runaway, and early detection is crucial to avoid catastrophic accidents. Existing internal short circuit detection techniques can be divided into three categories: (1) Threshold triggered alarm based on external parameters such as voltage and temperature. Although this method is simple to deploy, it relies on fixed thresholds and is difficult to capture early weak short circuit signals, sensitive to aging and temperature interference, and has high false alarm rate, response lag, and other defects; (2) Equivalent circuit or electrochemical model combined with state estimation algorithm (such as Kalman filter). Although this method has physical interpretability, model parameter identification is complex, and it is not suitable for dynamic changes in working conditions, and the large computational overhead restricts real-time performance; (3) Machine learning directly learns fault features. Although this method has high accuracy when the features are sufficient, it relies heavily on labeled fault samples, and the model decision-making process is not interpretable, making it difficult to generalize to unknown short circuit patterns.

[0003] In the prior art, the patent for invention with publication number CN119861305A discloses a lithium battery short circuit diagnosis and detection method and system, which specifically includes a diagnosis and detection platform, a short circuit detection module, a risk assessment module, a factor investigation module, an optimization analysis module, and a storage module are communicatively connected to the diagnosis and detection platform; the short circuit detection module is used for short circuit detection and analysis of the lithium battery: generating a detection period and dividing the detection period into several detection time periods, marking the lithium battery as a detection object, and marking the detection time period as a short circuit normal period or a short circuit abnormal period through a short circuit coefficient DL; the present application can perform short circuit detection and analysis on the lithium battery, collect and calculate the lithium battery operating parameters in each detection time period to obtain a short circuit coefficient, and feedback the short circuit probability of the lithium battery through the short circuit coefficient, thereby differentiating the detection time period. However, this method does not establish an internal state model of the battery, and only relies on external parameter correlation, which cannot quantify core fault indicators such as internal short circuit resistance; its optimization mechanism is triggered after the detection period ends, and early real-time warning cannot be achieved.

[0004] The invention patent with publication number CN120334752A discloses a lithium battery internal short circuit diagnosis method based on interpretable machine learning, which specifically includes collecting voltage, current and time data of the battery under different internal short circuit degrees during the constant current charging process, and establishing an experimental data set; the voltage data is preprocessed and filtered, the ratio of voltage increment to capacity increment is calculated, and the capacity increment IC curve feature is obtained; the battery state of charge-open circuit voltage relationship is extracted, and the battery equivalent circuit model for the constant current stage is established; the state of charge of the battery is estimated through the extended Kalman filtering algorithm and the battery charging data; the characteristic parameters are extracted from the IC curve, including the peak value, peak position and area between the peaks; based on the state of charge estimation result, the time sequence features are extracted, and a multi-dimensional feature space is constructed; the extracted features are standardized, and a classification model is constructed; the SHAP additivity feature interpretation method is introduced to realize the explainability analysis of the classification model decision process. This method uses IC curve features and equivalent circuit model to realize hierarchical diagnosis, but relies on constant current charging conditions, which is difficult to cover actual dynamic loads; SHAP explainability analysis only explains the feature contribution after the fact, does not embed the physical mechanism into the model structure, and the feature engineering is complex (multiple peak parameters need to be extracted). SUMMARY

[0005] The first object of the present application is to provide a lithium battery internal short circuit diagnosis method based on a double-state cyclic equivalent circuit, aiming to solve the problem of high false alarm rate, poor adaptability and weak generalization caused by the inability to capture weak signals, static parameters and the scarcity of fault samples in the prior art. To solve the above technical problems, a lithium battery internal short circuit diagnosis method based on a double-state cyclic equivalent circuit is provided, comprising the steps of:

[0006] S1, a double-state cyclic neural network equivalent circuit model is constructed, a battery second-order RC equivalent circuit model is embedded into a double-state cyclic neural network, state updating is realized through a customized cyclic neural network unit, the internal state follows a physical equation, and then a multilayer perceptron is used to fit the OCV-SOC relationship between open circuit voltage and state of charge, the activation function of the multilayer perceptron is a smoothing function;

[0007] S2, a weighted loss function is defined, a weighted mean square error loss function for terminal voltage is designed for the key area of battery voltage, and higher weight is given to the prediction error of complex voltage interval;

[0008] S3, model training, a composite optimization strategy including learning rate decay and gradient clipping is used to minimize the weighted mean square error loss function, and the equivalent circuit physical parameters and multilayer perceptron parameters are optimized jointly;

[0009] S4, voltage prediction, real-time acquisition of the actual current sequence, actual voltage sequence and battery state of charge of the lithium battery, the actual current sequence includes N time steps of actual current, N≥1, input the actual current sequence of the battery to be tested into the trained model, and output a predicted voltage sequence under the health state;

[0010] S5, residual generation, comparing the predicted voltage sequence with the actual voltage sequence, and calculating a residual sequence;

[0011] S6, fault feature extraction, dynamic time warping processing of the residual sequence under a sliding window, extracting the DTW distance between each window subsequence and a zero vector, and calculating the mean and maximum of the distance sequence to form a two-dimensional fault feature vector;

[0012] S7, fault identification, unsupervised clustering of the two-dimensional fault feature vector based on a Gaussian mixture model, automatic classification of the health state and the internal short circuit fault state, and realization of the internal short circuit early diagnosis without fault samples.

[0013] In one of the embodiments, the internal state vector of the double-state recurrent neural network is:

[0014] ;

[0015] Wherein is the polarization voltage of the two equivalent capacitors in the RC equivalent circuit at time step is the battery state of charge at time step , and the relationship with the actual current is:

[0016] ;

[0017] Wherein is the battery state of charge at the previous moment, is the sampling step, is the rated capacity, is the coulomb efficiency;

[0018] The output vector is:

[0019] ;

[0020] Wherein is the predicted value of the battery terminal voltage at time step

[0021] In one of the embodiments, the double-state recurrent neural network equivalent circuit model includes embedding the discrete state equation of the second-order RC equivalent circuit into the double-state recurrent neural network, and the state update matrix is represented as:

[0022] ​​ ;

[0023] wherein is a time constant, the battery terminal voltage prediction value satisfies:

[0024] ;

[0025] The multi-layer perceptron expression is:

[0026] ;

[0027] The nonlinear relationship between the open-circuit voltage and the state of charge is fitted, is a training parameter of the MLP subnetwork, the network structure of the multi-layer perceptron is 5 layers of hidden layers and 32 neurons per layer, and the activation function adopts .

[0028] In one of the embodiments, the definition of the weighted loss function comprises:

[0029] ;

[0030] wherein is a voltage-related weight function, and in a preset voltage critical interval ; is a predicted voltage in the training process; is an actual voltage.

[0031] In one of the embodiments, the composite optimization strategy further comprises an optimizer selected as Adam adaptive gradient estimation, and a learning rate decay formula thereof is:

[0032] ;

[0033] wherein is an initial learning rate, is a decay rate, is a decay step, and gradient clipping is performed simultaneously, and the formula is:

[0034] ;

[0035] wherein is an original gradient vector, is a gradient clipping threshold.

[0036] In one of the embodiments, the fault feature extraction step comprises setting a sliding window with a length of on the residual sequence , a step of , and generating a subsequence:

[0037] ;

[0038] wherein , is the total number of sample points contained in the residual sequence, the DTW distance of each sub-sequence and zero vector is calculated :

[0039] ;

[0040] wherein is the path of the sequence, is the sequence index, the mean and maximum value of the sequence are extracted to form a two-dimensional fault feature vector :

[0041] .

[0042] In one embodiment, the sliding window length corresponds to a 100-second time window at a sampling frequency of 1 Hz.

[0043] In one embodiment, the fault feature vector set is input into a Gaussian mixture model:

[0044] ;

[0045] wherein is a Gaussian distribution, is a covariance matrix, and the parameters are solved by an expectation maximization algorithm, and then the fault clusters are identified:

[0046] ;

[0047] If the sample is assigned to a cluster, it is determined to be an internal short circuit fault.

[0048] The second object of the present application is to provide a lithium battery internal short circuit diagnosis system based on a two-state cyclic equivalent circuit, aiming to solve the problems of existing systems that detection accuracy depends on fixed threshold, lack of internal state modeling ability, and closed-loop response lag. Through the collaborative architecture of physical constraint RNN and unsupervised time series feature extraction, high sensitivity, strong interpretability, and low delay of internal short circuit early detection and active protection are achieved.

[0049] To solve the above technical problems, a lithium battery internal short circuit diagnosis system based on a two-state cyclic equivalent circuit is provided, which runs the above-mentioned lithium battery internal short circuit diagnosis method based on a two-state cyclic equivalent circuit, including a data acquisition module, a mixture model module, a residual calculation module, a feature extraction module, a fault diagnosis module, and a control execution module.

[0050] In one embodiment, the data acquisition module acquires the actual current sequence of the battery in real time. With actual voltage sequence The hybrid model module incorporates a dual-state recurrent neural network equivalent circuit model, takes a current sequence as input, and outputs the predicted voltage after model training. The residual calculation module calculates the predicted voltage. With actual voltage residual sequence The feature extraction module performs sliding window segmentation on the residual sequence, calculates the DTW distance between each window subsequence and the zero vector, and generates a two-dimensional fault feature vector. The fault diagnosis module clusters feature vectors based on a Gaussian mixture model, automatically identifies fault clusters, and outputs an internal short-circuit alarm signal; the control execution module triggers battery protection strategies based on the alarm signal, including power reduction or circuit disconnection.

[0051] In one embodiment, the physical parameters of the hybrid model module The MLP parameters are jointly trained using a composite optimization strategy, which includes: the optimizer selects Adam adaptive gradient estimator, and its learning rate is denoted as:

[0052]

[0053] in The initial learning rate, The attenuation rate, To decrease the step size, gradient clipping is also performed:

[0054]

[0055] in This is the original gradient vector. This is the gradient clipping threshold.

[0056] In one embodiment, the sliding window length of the feature extraction module And preferred This corresponds to a 50-200 second time window at a sampling frequency of 1Hz.

[0057] In one embodiment, the system is integrated into an embedded processor of a battery management system (BMS) with a real-time operating cycle of ≤1 second.

[0058] Implementing the embodiments of the present invention will have the following beneficial effects:

[0059] 1. The lithium battery internal short circuit diagnosis method based on the dual-state recurrent equivalent circuit in the first embodiment of the present application, by embedding a second-order RC equivalent circuit model into a dual-state recurrent neural network, a hybrid model with physical constraints is constructed, the multi-layer perceptron is used to fit the OCV-SOC nonlinear relationship, and the prediction accuracy of the key voltage region is strengthened through the weighted loss function; the physical parameters and neural network parameters are jointly trained by using the composite optimization strategy; further, the dynamic time warping (DTW) is used to extract the time sequence pattern features of the residual sequence, and the Gaussian mixture model (GMM) is used for unsupervised clustering to realize automatic fault identification. Compared with the prior art, the method solves the problems of high false negative rate of threshold discrimination method, poor working condition adaptability of model driven method, and dependence on a large number of labeled samples of data driven method, realizes high sensitivity internal short circuit detection without fault samples, and has strong interpretability and low computational complexity;

[0060] 2. The lithium battery internal short circuit diagnosis system based on the dual-state recurrent equivalent circuit in the second embodiment of the present application, the data acquisition module synchronously acquires the battery current and voltage signals in real time, the hybrid model module performs voltage prediction with physical constraints, the residual calculation module quantifies the health offset, the feature extraction module captures the time sequence anomaly by using the sliding window DTW distance, the fault diagnosis module automatically identifies the fault based on GMM clustering, and the control execution module triggers the hierarchical protection strategy to build a "detection-diagnosis-protection" closed loop. At the same time, the system ensures the efficient operation of the model parameters in the embedded processor through the composite optimization strategy, and the feature extraction module adopts a fixed window length L=100 (corresponding to a 100-second time window under 1Hz sampling) to balance the sensitivity and real-time performance. Compared with the existing system, the system solves the technical defects that the traditional BMS cannot quantize the internal short circuit resistance in real time and the closed loop response is slow, supports millisecond-level fault judgment and protection action, and provides active safety protection for electric vehicles and energy storage systems. BRIEF DESCRIPTION OF DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0062] Figure 1 The lithium battery internal short circuit diagnosis method based on the dual-state recurrent equivalent circuit according to the first embodiment of the present application is shown in the flowchart;

[0063] Figure 2 The lithium battery internal short circuit diagnosis system architecture based on the dual-state recurrent equivalent circuit according to the second embodiment of the present application is shown in the flowchart;

[0064] Figure 3The lithium battery internal short circuit diagnosis method based on the double-state cyclic equivalent circuit for the embodiment one of the application;

[0065] Figure 4 Another principle diagram of the lithium battery internal short circuit diagnosis method based on the double-state cyclic equivalent circuit for the embodiment one of the application;

[0066] Figure 5 The lithium battery internal short circuit diagnosis system implementation flow chart based on the double-state cyclic equivalent circuit for the embodiment two of the application;

[0067] Figure 6 The comparison diagram of the lithium battery internal short circuit diagnosis method based on the double-state cyclic equivalent circuit for the embodiment one of the application and the LSTM and RNN training test health voltage results;

[0068] Figure 7 Another comparison diagram of the lithium battery internal short circuit diagnosis method based on the double-state cyclic equivalent circuit for the embodiment one of the application and the LSTM and RNN training test health voltage results;

[0069] Figure 8 The comparison result diagram of the lithium battery internal short circuit diagnosis system test prediction health voltage and the first battery health voltage, 500Ω internal short circuit battery voltage and 300Ω internal short circuit battery voltage for the embodiment two of the application;

[0070] Figure 9 The comparison result diagram of the lithium battery internal short circuit diagnosis system test prediction health voltage and the second battery health voltage, 500Ω internal short circuit battery voltage and 300Ω internal short circuit battery voltage for the embodiment two of the application;

[0071] Figure 10 The comparison result diagram of the lithium battery internal short circuit diagnosis system test prediction health voltage and the third battery health voltage, 500Ω internal short circuit battery voltage and 300Ω internal short circuit battery voltage for the embodiment two of the application;

[0072] Figure 11 The diagnosis health battery, 500Ω internal short circuit battery and 300Ω internal short circuit battery result diagram of the lithium battery internal short circuit diagnosis system based on the double-state cyclic equivalent circuit for the embodiment two of the application. DETAILED DESCRIPTION

[0073] For the purposes of the present invention, the present invention will be described for a preferred embodiment with reference to the attached figures. The figures are intended to be illustrative rather than restrictive. Many suitable modifications and adaptations can be made within the scope of the present invention and accordingly the appended claims are intended to encompass all such modifications and adaptations.

[0074] It is to be understood that where the terms specific elements are preceded by the term "a" or "an" they can be considered to mean "one or more" unless otherwise stated.

[0075] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used in the description of the invention herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.

[0076] Embodiment One

[0077] Reference is made to Figure 1 , 3 , 4, the embodiment one of the present invention provides a lithium battery internal short circuit diagnosis method based on double state cycle equivalent circuit, the embodiment includes:

[0078] Firstly, the physical constraint modeling is established, the second-order RC equivalent circuit discrete state equation is embedded into the double state RNN, and the state update matrix strictly follows the physical equation:

[0079]

[0080] Wherein is the time constant, is the sampling step, is the rated capacity, is the coulomb efficiency. The 5-layer MLP subnetwork (32 neurons in each layer, activation) is used to fit the OCV-SOC nonlinear relationship:

[0081]

[0082] Then, the joint optimization training is carried out, and the voltage interval weighted loss function is designed:

[0083]

[0084] Assign high weight in voltage critical interval To strengthen the accuracy of nonlinear segment. Adopt composite optimization strategy: Adam optimizer combined with learning rate decay is the decay rate, is the decay step) and gradient clipping

[0085]

[0086] Joint optimization of physical parameters and MLP parameters .

[0087] Then extract the time sequence features, input the test battery current sequence , output the predicted voltage after model training , generate residual sequence:

[0088]

[0089] Divide the residual sequence by sliding window (length , step ), calculate the DTW distance of each subsequence and zero vector :

[0090]

[0091] Extract the mean and maximum of the sequence to form a two-dimensional fault feature vector :

[0092]

[0093] Finally, unsupervised diagnosis is performed, and the feature vector set is input into the Gaussian mixture model:

[0094]

[0095] Solve the parameters by expectation maximization algorithm , identify the fault cluster:

[0096]

[0097] If the sample is assigned to cluster, it is determined as internal short circuit fault and the protection mechanism is triggered.

[0098] ​The embodiment one of the present application is implemented, a second-order RC equivalent circuit model is embedded into a double-state recurrent neural network, a hybrid model with physical constraints is constructed, a multi-layer perceptron is used to fit the OCV-SOC nonlinear relationship, and the prediction accuracy of the key voltage area is strengthened through a weighted loss function; a composite optimization strategy is used to jointly train the physical parameters and the neural network parameters; further, a dynamic time warping (DTW) is used to extract the time sequence pattern features of the residual sequence, and an unsupervised clustering is realized through a Gaussian mixture model (GMM) to realize automatic fault identification. Compared with the prior art, the present method solves the problems of high false negative rate of threshold discrimination method, poor working condition adaptability of model driven method, and dependence on a large number of labeled samples of data driven method, realizes high sensitivity internal short circuit detection without fault samples, and has strong interpretability and low computational complexity.

[0099] Embodiment two

[0100] The lithium battery internal short circuit diagnosis system based on the double-state recurrent equivalent circuit in the present embodiment two is different from the subject matter protected by the lithium battery internal short circuit diagnosis method based on the double-state recurrent equivalent circuit in the embodiment one, and the specific differences are as follows: the lithium battery internal short circuit diagnosis system based on the double-state recurrent equivalent circuit comprises a data acquisition module, a hybrid model module, a residual calculation module, a feature extraction module, a fault diagnosis module and a control execution module.

[0101] In an optional embodiment, first, the data acquisition module synchronously acquires the current and voltage signals with a high-precision ADC; then, the hybrid model module generates a predicted voltage on an embedded processor by running an RNN equivalent circuit; the residual module calculates the difference between the predicted value and the actual value in real time; the feature extraction module compresses the residual sequence into two-dimensional features by using a sliding window and a DTW algorithm; the fault diagnosis module automatically identifies faults by GMM clustering; and finally, the control module executes a protection action.

[0102] The embodiment two of the present application is implemented, and has the following beneficial effects:

[0103] The data acquisition module synchronously acquires the battery current and voltage signals in real time, the hybrid model module performs physical constraint voltage prediction, the residual calculation module quantifies health deviation, the feature extraction module captures time sequence anomalies by using a sliding window DTW distance, the fault diagnosis module automatically identifies faults based on GMM clustering, and the control execution module triggers a hierarchical protection strategy to construct a “detection-diagnosis-protection” closed loop. At the same time, the system ensures efficient operation of the model parameters in the embedded processor through a composite optimization strategy, and the feature extraction module adopts a fixed window length L=100 (corresponding to a 100-second time window under 1Hz sampling) to balance sensitivity and real-time performance. Compared with the existing system, the present system solves the technical defects that the traditional BMS cannot quantize the internal short circuit resistance in real time and the closed loop response is slow, supports millisecond-level fault judgment and protection action, and provides active safety protection for electric vehicles and energy storage systems.

[0104] Example Three

[0105] The system exhibits significant technical advantages under strictly designed test conditions. Experiments were conducted in a constant 25°C environment, and full discharge tests were conducted on three groups of batteries from full charge state (SOC = 1) to complete cutoff (SOC = 0), where the short circuit condition was simulated by connecting a resistor in parallel across the battery. The experimental results confirm that, as shown in Figure 6 、 7 , the health voltage prediction error (RMSE = 8.7 mV) of the physically constrained RNN is reduced by 63% compared to LSTM and RNN, and the error in the critical voltage interval (3.0-3.5 V) is <5 mV, proving that the weighted loss function effectively enhances the accuracy of the nonlinear section. The dynamic characteristics of the fault detection of the three groups of batteries are shown in Figure 8 、 9 , 10: after 500 Ω short circuit and 300 Ω short circuit, the residual sequence deviates significantly from the zero baseline (>100 mV), and the mean value of the DTW feature jumps from 0.12 in the healthy state to 2.87, achieving fault transient capture; Figure 11 The GMM clustering results show that the features of healthy batteries and short-circuit batteries are completely separated, and the fault cluster center norm (healthy cluster = 12.3), verifying the reliability of unsupervised diagnosis.

Claims

1. A lithium battery internal short circuit diagnosis method based on a two-state cyclic equivalent circuit, characterized by, Comprise: S1, constructing a double-state recurrent neural network equivalent circuit model: embedding a battery second-order RC equivalent circuit model into a double-state recurrent neural network, updating the state through a customized recurrent neural network unit, and the internal state follows a physical equation, then fitting the open-circuit voltage and state of charge OCV-SOC relationship using a multi-layer perceptron, the activation function of the multi-layer perceptron is a smoothing function; S2, define a weighted loss function: design a weighted mean square error loss function for the end voltage of the battery voltage key area, and give higher weight to the prediction error of the complex voltage interval; S3, model training: using a composite optimization strategy including learning rate decay and gradient clipping to minimize the weighted mean square error loss function, and jointly optimizing the equivalent circuit physical parameters and multi-layer perceptron parameters; S4, voltage prediction: real-time acquisition of the actual current sequence, actual voltage sequence and battery state of charge of the lithium battery, the actual current sequence includes N actual currents at N time steps, N≥1; input the actual current sequence of the battery to be tested into the trained model, and output the predicted voltage sequence under the healthy state; S5, residual generation: compare the predicted voltage sequence with the actual voltage sequence to calculate the residual sequence; S6, fault feature extraction: dynamic time warping processing of the residual sequence under the sliding window, extracting the DTW distance between each window subsequence and the zero vector, and calculating the mean and maximum of the distance sequence to form a two-dimensional fault feature vector; S7, fault identification: unsupervised clustering of the two-dimensional fault feature vector based on Gaussian mixture model, automatic classification of healthy state and internal short circuit fault state, and realization of internal short circuit early diagnosis without fault samples; Wherein, the double-state recurrent neural network equivalent circuit model comprises: The discrete state equation of the second-order RC equivalent circuit is embedded into the double-state recurrent neural network, and the state update matrix is represented as: ； wherein is a time constant, the battery terminal voltage prediction satisfies: ； The multi-layer perceptron expression is: ； wherein to fit the non-linear relationship between open circuit voltage and state of charge, are the training parameters of the MLP subnetwork, and the activation function is ; The definition of the weighted loss function comprises: ； wherein is a voltage-dependent weight function, within a pre-set voltage critical interval is a predicted voltage during the training process; is an actual voltage.

2. The lithium battery internal short circuit diagnosis method based on double-state recurrent equivalent circuit according to claim 1, characterized in that: The internal state vector of the double-state recurrent neural network is: ； wherein is the polarization voltage of the two equivalent capacitances in the RC equivalent circuit of the time step , is the state of charge of the battery at the time step , is the relationship between the actual current ； wherein SoCprev is the battery state of charge at the previous time, SoCprev is the battery state of charge at the previous time, SoCprev is the battery state of charge at the previous time, SoCprev is the battery state of charge at the previous time, The output vector is: ； wherein is the battery terminal voltage prediction value at time step .

3. The lithium battery internal short diagnosis method based on a two-state cyclic equivalent circuit according to claim 1, characterized by, The composite optimization strategy further comprises: The optimizer selects Adam adaptive gradient estimation, and the learning rate decay formula is: ； wherein is the initial learning rate, is the decay rate, is the decay step, while gradient clipping is performed, with the formula: ； wherein is the original gradient vector, is the gradient clipping threshold.

4. The lithium battery internal short diagnosis method based on a two-state cyclic equivalent circuit according to claim 3, characterized by, The fault feature extraction comprises: In the residual sequence A sliding window of length is set above, with a step size , generating a subsequence: ； wherein , is the number of all sample points contained in the residual sequence, the DTW distance of each sub-sequence to the zero vector is calculated : ； wherein is the alignment path, is the sequence index, the mean and maximum of the sequence are extracted to form a two-dimensional fault feature vector : 。 5. The lithium battery internal short circuit diagnosis method based on double-state recurrent equivalent circuit according to claim 4, characterized in that: the sliding window length .

6. The lithium battery internal short diagnosis method based on a two-state cyclic equivalent circuit according to claim 5, characterized by, The fault identification comprises: a set of fault feature vectors in a gaussian mixture model: ； where is a Gaussian distribution, is a covariance matrix, and parameters are solved by an expectation-maximization algorithm , and subsequently identify failure clusters: ； If the sample is assigned to a cluster, then an internal short circuit fault is determined.

7. A lithium battery internal short circuit diagnosis system based on a dual state cyclic equivalent circuit, for the lithium battery internal short circuit diagnosis method based on a dual state cyclic equivalent circuit according to any one of claims 1 to 6, characterized by, Comprise: a data acquisition module, which acquires an actual current sequence of the battery in real time and an actual voltage sequence ; A hybrid model module, wherein the dual-state recurrent neural network equivalent circuit model is built-in, an input current sequence, and an output predicted voltage after model training is completed ; a residual calculation module that calculates a residual voltage from an actual voltage ;​ The feature extraction module performs sliding window segmentation on the residual sequence, calculates DTW distance between each window sub-sequence and a zero vector, and generates a two-dimensional fault feature vector ; A fault diagnosis module that clusters the feature vector based on a Gaussian mixture model, automatically identifies fault clusters and outputs an internal short circuit alarm signal; A control execution module that triggers a battery protection strategy including power reduction or circuit shutdown according to the alarm signal; The system is integrated into the embedded processor of the battery management system.

Images