A multi-modal feature fusion lithium ion battery internal short circuit fault diagnosis method

Abstract

The application discloses a kind of multi-modal feature fusion's lithium ion battery internal short circuit fault diagnosis method, belong to battery fault diagnosis technical field, the method includes: the end voltage residual signal of the battery to be diagnosed and normal battery is constructed;End voltage residual signal is decomposed into intrinsic mode function, intrinsic mode function is divided into high-frequency component and low-frequency component;High-frequency component is decomposed into several finite bandwidth modal components;Low-frequency component and finite bandwidth modal component are merged to form multi-modal feature set;Multi-modal feature fusion is carried out to obtain the fusion feature sequence of unified representation fault evolution process;Global dependence of fusion feature sequence is captured using Transformer, combined with BiGRU to extract bidirectional timing dynamics, output enhanced fusion feature sequence;The local outlier factor value of enhanced fusion feature sequence is calculated, when exceeding threshold, trigger internal short circuit early warning, realize the accurate identification of internal short circuit fault, early warning and visual positioning.

Description

Technical Field

[0001] This invention relates to the field of battery fault diagnosis technology, and in particular to a method for diagnosing internal short-circuit faults in lithium-ion batteries by fusing multimodal features. Background Technology

[0002] In the field of electric vehicles and energy storage systems, lithium-ion batteries are widely used due to their high energy density and long cycle life. However, internal short-circuit faults are one of the main causes of battery thermal runaway and even fires and explosions. Internal short-circuit faults are highly insidious, with weak early characteristics that are easily masked by fluctuations in operating conditions and sensor noise.

[0003] Existing methods for internal short-circuit fault diagnosis are mainly divided into model-free diagnostic methods and model-based diagnostic methods. Model-free methods are simple to implement but are easily affected by data quality, resulting in a high false alarm rate. Model-based methods, while having better physical interpretability, are highly dependent on model accuracy and parameter identification capabilities, have high computational complexity, and struggle to balance accuracy, real-time performance, and versatility. Furthermore, existing methods lack effective early warning mechanisms. Summary of the Invention

[0004] To address the problems of existing technologies, this invention provides a method for diagnosing internal short-circuit faults in lithium-ion batteries by fusing multimodal features, the method comprising:

[0005] S1: Construct the terminal voltage residual signal between the battery to be diagnosed and the normal battery;

[0006] S2: The terminal voltage residual signal is decomposed into intrinsic mode functions using CEEMDAN, and the decomposed intrinsic mode functions are automatically divided into high-frequency components and low-frequency components through sample entropy clustering.

[0007] S3: Perform VMD secondary decomposition on the high-frequency components to decompose them into several mode components with limited bandwidth, and extract the millivolt-level transient voltage drop characteristics caused by internal short circuits.

[0008] S4: Combine low-frequency components with modal components of limited bandwidth to form a multimodal feature set;

[0009] S5: An adaptive attention weighting mechanism is introduced to fuse multimodal features, resulting in a unified fusion feature sequence that represents the fault evolution process;

[0010] S6: Construct a Transformer-BiGRU hybrid framework, using Transformer to capture the global dependencies of the fused feature sequence, and combining it with BiGRU to extract bidirectional temporal dynamics, outputting an enhanced fused feature sequence;

[0011] S7: Calculate the local outlier factor value of the enhanced fusion feature sequence, and trigger an early warning of internal short circuit when it exceeds the threshold.

[0012] Furthermore, S1 specifically includes:

[0013] S11: Real-time acquisition of terminal voltage timing data of the battery under diagnosis in the same battery pack. d (t) and the terminal voltage timing data U of the calibrated normal battery n (t);

[0014] S12: Calculate the residual signal x(t) of the terminal voltage:

[0015]

[0016] S13: Perform moving average filtering on the terminal voltage residual signal x(t).

[0017] Furthermore, S2 specifically includes:

[0018] S21: Add adaptive white noise to the preprocessed terminal voltage residual signal x(t) and take ensemble averaging to decompose the terminal voltage residual signal x(t) into K intrinsic mode functions (IMFs). k (t) and a residual component r K (t):

[0019]

[0020] In the formula, IMF k (t) represents the oscillation mode of the original terminal voltage residual signal at different time scales;

[0021] S22: Calculate each IMF k The sample entropy SampEn of (t):

[0022]

[0023] Where m is the embedding dimension, r is the similarity tolerance, and N is the IMF. k (t) Total number of discrete sampling points, B is the number of template pairs matching m discrete sampling points, and A is the number of template pairs matching m+1 discrete sampling points;

[0024] S23: Construct feature vectors from the sample entropy values ​​of each IMF, and automatically classify all IMFs into high-frequency classes C using the K-means clustering algorithm. high and low-frequency C low Reconstructing the high-frequency component x high (t) and low-frequency component x low (t):

[0025]

[0026] .

[0027] Furthermore, S3 specifically includes:

[0028] The high-frequency component x is obtained by solving the variational constraint problem. high (t) is decomposed into several modal components with finite bandwidth u i (t), the constraint condition is the modal components u of each finite bandwidth. i The sum of (t) equals the high-frequency component x. high (t);

[0029] By introducing a quadratic penalty factor β and a Lagrange multiplier λ, and using the alternating direction multiplier method for iterative solution, the characteristics of millivolt-level transient voltage drop caused by internal short circuit are extracted.

[0030] Furthermore, S5 specifically includes:

[0031] S51: Extract key statistical features from each component in the multimodal feature set to construct a 6-dimensional feature vector f. m :

[0032]

[0033] Where μ is the mean, σ² is the variance, Kurt is the kurtosis, Skew is the skewness, E is the energy, and SampEn is the sample entropy;

[0034] The feature vectors of all components are combined to obtain the initial fused feature set, thus completing the alignment and splicing of multimodal features;

[0035] S52: The eigenvector f of the i-th component i The input is a two-layer fully connected network. The first layer uses the tanh activation function, and the second layer outputs the raw attention score s. i :

[0036]

[0037] Among them, W b The weight matrix is ​​6×6, where d is the bias and w is the weight matrix. a This is the weight vector;

[0038] Then, the weight coefficients α are obtained by normalization using the Softmax function. i :

[0039]

[0040] S53: Based on the weighting coefficient α iFor the eigenvector f of each component i By performing weighted summation, a fusion feature sequence F that uniformly represents the fault evolution process is obtained. fusion :

[0041] .

[0042] Furthermore, S6 specifically includes:

[0043] S61: Fuse the feature sequence F fusion The sample is divided into multiple samples according to a fixed time window. Simultaneously, the internal short-circuit resistance R generated using offline calibration or simulation is used. ISC As labels, a training dataset is constructed and divided into training and validation sets proportionally;

[0044] S62: Construct a multi-scale Transformer encoder module. The input sample first passes through three parallel convolutional layers to extract multi-scale temporal patterns. The outputs are summed and then normalized to obtain a multi-scale embedding matrix. Subsequently, a segmented sparse multi-head self-attention mechanism is adopted to divide the multi-scale embedding matrix into non-overlapping segments along the time dimension. Standard multi-head self-attention is calculated in each segment, and cross-segment attention connections are introduced between adjacent segments to form a sparse interaction pattern. After passing through multiple encoding layers, the output is adaptively averaged pooled along the time dimension to obtain the global feature vector.

[0045] S63: Construct an attention-enhanced bidirectional gated recurrent unit module to perform temporal modeling on the input samples along both the forward and backward directions to obtain the hidden state at each time step.

[0046] For each time step, the gating coefficients are calculated based on the current hidden state and input features. The hidden state is then adaptively weighted to obtain the gated enhanced hidden sequence. Attention weights are calculated on the gated enhanced hidden sequence and then summed to obtain the branch output vector.

[0047] S64: Introduce a cross-attention interaction module. Use the global feature vector as the query and the branch output vector as the key and value to calculate cross-attention. Conversely, use the branch output vector as the query and the global feature vector as the key and value to calculate cross-attention in the other direction. Add the interaction result to the original global feature vector and the branch output vector and normalize them layer by layer to obtain the two enhanced branch features.

[0048] S65: The enhanced features of the two branches are concatenated and input into the gating network to obtain the fusion weights of the two branches. The enhanced features of the two branches are then weighted and summed to obtain the intermediate fusion features. Finally, a fully connected layer is used to linearly transform the features back to the original feature dimensions, and the enhanced fusion feature sequence is output.

[0049] Furthermore, S7 specifically includes:

[0050] S71: Collect multiple sets of fused feature samples of the battery pack when it is running in a healthy state, randomly sample from them to construct a reference dataset, and calculate the local outlier factor for each sample point in the reference dataset.

[0051] S72: Analyze the local outlier factors of all sample points in the reference dataset and determine the warning threshold of the local outlier factors;

[0052] S73: During online operation, the local outlier factor of the enhanced fusion feature sequence sample point corresponding to the current time relative to the reference dataset is calculated in real time. If it exceeds the warning threshold, an internal short-circuit warning is triggered and the warning time is recorded.

[0053] Furthermore, the calculation of the local outlier factor for each sample point in the reference dataset in step S71 specifically includes:

[0054] Determine the k nearest neighbors of each sample point, calculate the local reachability density of each sample point, which is defined as the reciprocal of the average reachability distance from the sample point to its k nearest neighbors. The local outlier factor of the sample point is the average of the ratios of the local reachability densities of all sample points in the neighborhood of the sample point to the local reachability density of the sample point itself.

[0055] The beneficial effects of the technical solution provided by the embodiments of the present invention are as follows: First, the residual terminal voltage signals of the faulty battery and the normal battery are constructed to filter out operating condition interference; then, the CEEMDAN-VMD two-layer decomposition strategy is adopted, and the intrinsic mode functions decomposed from the residual terminal voltage signals are automatically divided into high-frequency and low-frequency components through sample entropy clustering, and variational mode secondary decomposition is performed on the high-frequency components to fully extract the transient features caused by micro-short circuits; an adaptive attention weighting mechanism is introduced to fuse the features of each mode to obtain fused features that can uniformly characterize the fault evolution process; next, the fused feature sequence is input into a hybrid model composed of Transformer and bidirectional gated recurrent unit, and the Transformer is used to capture global dependencies and the BiGRU is used to extract bidirectional time-series dynamics to achieve high-precision prediction of internal short-circuit resistance; finally, online anomaly detection is performed based on the local outlier (LOF) of the sample points in the fused feature space, and an early warning is triggered when the LOF value exceeds a fixed threshold. This solves the problems of weak internal short-circuit fault features, delayed diagnosis, insufficient early warning capability, and unintuitive results in the prior art. By employing data dual decomposition, multimodal fusion, hybrid deep learning prediction, and density clustering early warning, we can achieve accurate identification, early warning, and visual location of internal short-circuit faults. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0057] Figure 1 This is a flowchart of a multimodal feature fusion method for diagnosing internal short-circuit faults in lithium-ion batteries, provided by the present invention.

[0058] Figure 2 This is a schematic diagram of a terminal voltage residual signal provided by the present invention;

[0059] Figure 3 This is a schematic diagram of multimodal feature fusion provided by the present invention;

[0060] Figure 4 This is a schematic diagram of a Transformer-BiGRU hybrid framework provided by the present invention;

[0061] Figure 5 This invention provides a local outlier distribution map under healthy conditions;

[0062] Figure 6 This is a schematic diagram of an online early warning system for local outliers provided by the present invention. Detailed Implementation

[0063] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0064] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0065] A method for diagnosing internal short-circuit faults in lithium-ion batteries by fusing multimodal features includes the following steps:

[0066] Step 1: Constructing a terminal voltage residual signal to filter out operating condition interference: By calculating the terminal voltage difference between the battery to be diagnosed and the normal battery, a residual signal is constructed to eliminate the interference of operating condition fluctuations on the original data.

[0067] Step 1: Real-time acquisition of terminal voltage timing data U of the battery to be diagnosed in the same battery pack. d (t) and the terminal voltage timing data U of the calibrated normal batteryn (t), the sampling frequency is set according to the actual monitoring requirements. In this embodiment, it is set to 1Hz to ensure that millivolt-level voltage fluctuations caused by short circuit faults can be captured.

[0068] Step 2: Calculate the terminal voltage residual signal x(t), and eliminate common-mode interference caused by operating conditions such as temperature changes, load fluctuations, and differences in state of charge through differential operations:

[0069] (1)

[0070] like Figure 2 As shown, the terminal voltage residual signal x(t) mainly reflects the weak voltage drop characteristics caused by the internal short-circuit fault. To suppress high-frequency sensor noise, x(t) is processed by moving average filtering as the input for subsequent decomposition, and the window length is set to 5.

[0071] Step 2: Extract short-circuit transient features using the CEEMD-VMD two-layer decomposition strategy: The CEEMD-VMD strategy is used to decompose the terminal voltage residual signal x(t) into intrinsic mode functions. Based on sample entropy clustering, high-frequency components and low-frequency components are distinguished. The high-frequency components are then subjected to VMD secondary decomposition to extract the transient weak features caused by micro-short circuits.

[0072] Step 1: Perform fully adaptive noise ensemble empirical mode decomposition on the preprocessed terminal voltage residual signal x(t). By adding adaptive white noise and taking the ensemble average, the terminal voltage residual signal x(t) is decomposed into K intrinsic mode functions (IMFs). k (t) and a residual component r K (t):

[0073] (2)

[0074] In equation (2), IMF k (t) represents the oscillation mode of the original terminal voltage residual signal at different time scales; the decomposition parameters are set as follows: K=10, white noise standard deviation is 0.2, and maximum number of iterations is 500.

[0075] Step 2: Calculate each IMF k The sample entropy SampEn of (t), embedding dimension m=2, similarity tolerance r=0.2×std(IMF) k The larger the sample entropy value, the more complex and frequent the sequence. The formula for calculating sample entropy is:

[0076] (3)

[0077] In equation (3), N is the IMF k(t) is the total number of discrete sampling points, B is the number of template pairs matching m discrete sampling points, and A is the number of template pairs matching m+1 discrete sampling points.

[0078] Step 3: Construct feature vectors from the sample entropy values ​​of each IMF, and automatically divide all IMFs into high-frequency classes C using the K-means clustering algorithm (K=2). high and low-frequency C low Reconstructing the high-frequency component x high (t) and low-frequency component x low (t):

[0079] (4)

[0080] (5)

[0081] Step 4: For the high-frequency component x high (t) Perform variational mode decomposition (VMD) and solve the variational constraint problem to obtain x high (t) is decomposed into several modal components with finite bandwidth u i (t):

[0082] (6)

[0083] In equation (6), the constraint condition is: L is the number of decomposition modes, L=5. i} represents the L finite bandwidth modal components to be solved, {ω i} represents the center frequency corresponding to each modal component, and L is the total number of time steps; ∂ t To obtain the partial derivative with respect to time t, δ(t) is the Dirac impulse function, and j is the imaginary unit. The convolution operation between these two functions is used to construct an analytic signal for single-sided spectrum extraction; e -jωit This is a complex exponential frequency shift term used to shift the spectrum of each mode to the baseband. This represents the square of the L2 norm, and the bandwidth of the mode. The entire expression is under the constraints. Below, by minimizing the sum of the bandwidths of all modes, the high-frequency component x is... high (t) is decomposed into several sparse and bandwidth-limited sub-signals.

[0084] By introducing a quadratic penalty factor β=2000 and a Lagrange multiplier λ, and using the alternating direction multiplier method for iterative solution, the characteristics of millivolt-level transient voltage drop caused by internal short circuit are extracted.

[0085] Step 5: Convert the low-frequency component x low (t) and all modal components u i (t) are merged to form a multimodal feature set. Each component is a time-series vector of the same length as the original signal.

[0086] Step 3: Introduce an adaptive attention weighting mechanism for multimodal feature fusion: Extract statistical features from each component in the multimodal feature set, introduce an adaptive attention mechanism to calculate weights and perform weighted summation to obtain a fusion feature sequence that uniformly represents the fault evolution process.

[0087] Step 1: For each component y in the multimodal feature set m (t) Extract key statistical features to construct a 6-dimensional feature vector f m :

[0088] (7)

[0089] In equation (7), μ is the mean, σ² is the variance, Kurt is the kurtosis, Skew is the skewness, E is the energy, and SampEn is the sample entropy. The feature vectors of all components are combined to obtain the initial fusion feature set, thus completing the alignment and splicing of multimodal features.

[0090] Step 2: Design an adaptive attention mechanism module to calculate the attention score for each component. The feature vector f of the i-th component... i The input is a two-layer fully connected network. The first layer uses the tanh activation function, and the second layer outputs the raw attention score s. i :

[0091] (8)

[0092] In equation (8), W b The weight matrix is ​​6×6, where d is the bias and w is the weight matrix. a The weight vector is then normalized using the Softmax function to obtain the weight coefficients α. i :

[0093] (9)

[0094] In equation (9), s j Let be the eigenvector of the j-th component.

[0095] Step 3: Based on the calculated weighting coefficient α i For the eigenvector f of each component i Perform a weighted summation, such as Figure 3 As shown, a unified multimodal fusion feature vector F is obtained. fusion :

[0096] (10)

[0097] The fused feature is a single-channel time-series signal that can uniformly characterize the complete dynamic process of internal short-circuit faults, while effectively suppressing interference from irrelevant modes, providing high-quality input for subsequent prediction models. In this implementation scheme, the fused feature vector has a dimension of 6.

[0098] Step 4: Construct a Transformer-BiGRU hybrid framework: Use Transformer to capture the global dependencies of fused features, combine with BiGRU to extract bidirectional temporal dynamics, and output the enhanced fused feature sequence.

[0099] Step 1: Merge the feature sequence F fusion (t) is divided into multiple samples according to a fixed time window, with a window length of L=100 and a step size of S=20. Each sample Z i The dimension is L×d, where d=6 is the feature dimension. Meanwhile, the internal short-circuit resistance R is generated using offline calibration or simulation. ISC As labels, a training dataset is constructed, with the training set and validation set divided in a 7:3 ratio.

[0100] Step 2: Construct the Transformer encoder module. Input sample Z i ∈R (L×d) First, the system passes through three parallel 1D convolutional layers with kernel sizes of k=3, 5, and 7, a stride of 1, and d output channels. model =64, edge padding maintains the sequence length. The outputs of the three convolutions are summed and then normalized to obtain the multi-scale embedding matrix E. ms ∈R^ (L×64) :

[0101] (11)

[0102] We employ h=8 parallel attention heads, each with dimension d. k =16. After L=3 identical coding layers, the output global feature representation Y is obtained. trans The hidden layer dimension d of a feedforward neural network (FFN) ff =128, using the ReLU activation function.

[0103] Multi-scale embedding matrix E ms Divide into N uniformly along the time dimension s =4 non-overlapping segments, each segment having a length L s =25. Standard multi-head self-attention is computed within each segment; simultaneously, cross-segment attention connections are introduced every other segment (step size 2) between adjacent segments, forming a sparse interaction pattern. The attention calculation formula is:

[0104] (12)

[0105] In equation (12), Q is the Query matrix, obtained by linear projection of the input sequence, used to initiate a matching query; K is the Key matrix, obtained by linear projection of the input sequence, used for similarity matching with the Query; and V is the Value matrix, obtained by linear projection of the input sequence, and finally aggregated based on similarity weighted features. The dot product of Query and Key, T represents the matrix transpose operation, M is a sparse mask matrix, and computation is only allowed within segments and at specified cross-segment positions. The number of heads is h=8, and each head has dimension d. k =8.

[0106] After multi-head attention, L is stacked through feedforward neural networks, residual connections, and layer normalization. tr = Two identical encoding layers, outputting global feature Y trans ∈R (L×64) .

[0107] For global feature Y trans Adaptive average pooling is performed along the time dimension to obtain a global feature vector Y with fixed dimensions. pool ∈R 64 :

[0108] (13)

[0109] In equation (13), L is the total number of time steps.

[0110] Step 3: Construct an attention-enhanced bidirectional gated recurrent unit, introducing a self-attention gating mechanism in bidirectional temporal modeling to enhance attention to fault transient moments.

[0111] Sample Z i ∈R (L×6) As input to BiGRU. The number of hidden units in both the forward GRU and backward GRU is set to 32. For the forward GRU, the update process at time step t is as follows: First, calculate the update gate z. t and reset door r t Then, the candidate hidden states are calculated, and finally the hidden state h at the current time is obtained. t The specific formula is:

[0112] (14)

[0113] (15)

[0114] (16)

[0115] (17)

[0116] In equations (14), (15), (16), and (17), z t For updating the gate; r t To reset the door; For candidate hidden state; h t The hidden state at the current moment. σ is the Sigmoid activation function; W z W r W h Input feature x to the corresponding gate and candidate state t The weight matrix; U z U r U h For the corresponding gate and candidate states, the hidden state h from the previous time step t−1 The weight matrix; b z ,b r ,b h This is the bias vector.

[0117] For each time step t, based on the current hidden state h t ∈R 64 and input features x t ∈R^ 6 Calculate the gating coefficient g t :

[0118] (18)

[0119] (19)

[0120] In equations (18) and (19), W g ∈R (64×(64+6)) W h ∈R (64×64) b g ∈R 64 σ is a learnable parameter; σ is the Sigmoid function; ⊙ represents element-wise multiplication. The gating coefficient g t The closer the value is to 1, the greater the contribution of the feature at that moment to fault diagnosis, and the original hidden state is preserved; the closer the value is to 0, the more noise interference is suppressed. This yields the gated enhanced hidden sequence. ∈R^(L×64).

[0121] The adaptive weights are calculated and weighted summation is performed to obtain the branch output vector H. attn :

[0122] (20)

[0123] (twenty one)

[0124] In equations (20) and (21), v∈R64 W a ∈R^ (64×64) b a ∈R 64 Here, L is a learnable parameter, and L is the total number of time steps.

[0125] Step 4: Introduce a cross-attention interaction module to achieve bidirectional information exchange between the two branches, enhancing the complementarity of feature representations. The output Y of the multi-scale Transformer branch... pool As query Q, the output H of the BiGRU branch attn As the key K and value V, calculate the cross-attention:

[0126] (twenty two)

[0127] In equation (22), W Q W K W V ∈R (64×64) Let d be a linear projection matrix. k =64. Similarly, with H attn For querying, Y pool For keys and values, calculate C gru2tr .

[0128] The interaction result is added to the original branch output and then normalized to obtain the enhanced branch feature:

[0129] (twenty three)

[0130] (twenty four)

[0131] Step 5: Construct a dynamically gated fusion network to adaptively fuse the enhanced features of the two branches, replacing fixed concatenation or average fusion. The output vector of the concatenated two branches is:

[0132] (25)

[0133] Gated network calculation of fusion weights: The weights are obtained by passing the data through a two-layer fully connected network. The first layer has 16 neurons, and the second layer outputs a 2D raw score. After Softmax normalization, the weight coefficients α and β are obtained.

[0134] (26)

[0135] In equation (26), W1 ∈ R (16×128) b1∈R 16 W2∈R (2×16) b2∈R 2 .

[0136] Weighted fusion yields intermediate features:

[0137] (27)

[0138] The fused feature vectors are linearly transformed back to the original feature dimension d=6 through a fully connected layer to obtain the enhanced fused feature sequence F. e :

[0139] (28)

[0140] In equation (28), W out ∈R (6×64) b out ∈R 6 F e ∈R 6 .

[0141] Step 5: Online anomaly detection and early warning based on local outlier factors: A reference set is constructed offline using local outlier factors and an early warning threshold is determined. The local outlier factor value of the fused features is calculated online in real time. When the threshold is exceeded, an early warning of internal short circuit is triggered.

[0142] Step 1: Collect multiple sets of fused feature samples of the battery pack when it is in a healthy state, and randomly sample from them to construct a reference dataset D. ref For each sample point f i Calculate its k=20th nearest neighbor and the local reachability density lrd k (f i ) is defined as:

[0143] (29)

[0144] Local outlier (LOF) k (f i The value represents the average ratio of the local reachability density of each point in the neighborhood to the density of that point itself.

[0145] (30)

[0146] In equations (28) and (30), For sample f i The set of k nearest neighbors, O is f i Any neighboring sample point within the neighborhood.

[0147] Step 2: Determine the LOF warning threshold based on normal experimental data. For the reference dataset D... ref Local outlier (LOF) of all sample points v Offline calculation and statistical analysis showed that, under healthy conditions, the LOF value remained stable around 1, with a maximum value not exceeding 2.5. Figure 5 As shown. Therefore, the detection threshold T is... h Set to 2.5.

[0148] Step 3: During online runtime, calculate the fused feature sample point F corresponding to the current time t in real time. fusion (t) relative to the reference dataset D ref Local outlier LOF v The calculation reuses the k-value and reference dataset set offline. For example... Figure 6 As shown, when LOF v Exceeding a fixed threshold T h When the short circuit occurs, the system immediately triggers an internal short circuit warning and records the warning time.

[0149] It is worth noting that this invention first constructs the terminal voltage residual signal between the faulty battery and the normal battery to filter out operating condition interference; then, it employs a CEEMD-VMD two-layer decomposition strategy, automatically dividing the intrinsic mode functions decomposed from the terminal voltage residual signal into high-frequency and low-frequency components through sample entropy clustering, and performing variational mode secondary decomposition on the high-frequency components to fully extract the transient features caused by micro-short circuits; an adaptive attention weighting mechanism is introduced to fuse the features of each mode, obtaining a fused feature that can uniformly characterize the fault evolution process; next, the fused feature sequence is input into a hybrid model composed of a Transformer and a bidirectional gated recurrent unit, using the Transformer to capture global dependencies and the BiGRU to extract bidirectional temporal dynamics, achieving high-precision prediction of internal short-circuit resistance; finally, online anomaly detection is performed based on the local outlier (LOF) of sample points in the fused feature space, triggering an early warning when the LOF value exceeds a fixed threshold. This addresses the problems of weak internal short-circuit fault features, delayed diagnosis, insufficient early warning capabilities, and unintuitive results in existing technologies. By employing data dual decomposition, multimodal fusion, hybrid deep learning prediction, and density clustering early warning, we can achieve accurate identification, early warning, and visual location of internal short-circuit faults.

[0150] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0151] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for diagnosing internal short-circuit faults in lithium-ion batteries using multimodal feature fusion, characterized in that, The method includes: S1: Construct the terminal voltage residual signal between the battery to be diagnosed and the normal battery; S2: The terminal voltage residual signal is decomposed into intrinsic mode functions using CEEMDAN, and the decomposed intrinsic mode functions are automatically divided into high-frequency components and low-frequency components through sample entropy clustering. S3: Perform VMD secondary decomposition on the high-frequency components to decompose them into several mode components with limited bandwidth, and extract the millivolt-level transient voltage drop characteristics caused by internal short circuits. S4: Combine low-frequency components with modal components of limited bandwidth to form a multimodal feature set; S5: An adaptive attention weighting mechanism is introduced to fuse multimodal features, resulting in a unified fusion feature sequence that represents the fault evolution process; S6: Construct a Transformer-BiGRU hybrid framework, using Transformer to capture the global dependencies of the fused feature sequence, and combining it with BiGRU to extract bidirectional temporal dynamics, outputting an enhanced fused feature sequence; S7: Calculate the local outlier factor value of the enhanced fusion feature sequence, and trigger an early warning of internal short circuit when it exceeds the threshold.

2. The multimodal feature fusion-based method for diagnosing internal short-circuit faults in lithium-ion batteries according to claim 1, characterized in that, S1 specifically includes: S11: Real-time acquisition of terminal voltage timing data of the battery under diagnosis in the same battery pack. d (t) and the terminal voltage timing data U of the calibrated normal battery n (t); S12: Calculate the residual signal x(t) of the terminal voltage: S13: Perform moving average filtering on the terminal voltage residual signal x(t).

3. The method for diagnosing internal short-circuit faults in lithium-ion batteries based on multimodal feature fusion according to claim 2, characterized in that, S2 specifically includes: S21: Add adaptive white noise to the preprocessed terminal voltage residual signal x(t) and take ensemble averaging to decompose the terminal voltage residual signal x(t) into K intrinsic mode functions (IMFs). k (t) and a residual component r K (t): In the formula, IMF k (t) represents the oscillation mode of the original terminal voltage residual signal at different time scales; S22: Calculate each IMF k The sample entropy SampEn of (t): Where m is the embedding dimension, r is the similarity tolerance, and N is the IMF. k (t) Total number of discrete sampling points, B is the number of template pairs matching m discrete sampling points, and A is the number of template pairs matching m+1 discrete sampling points; S23: Construct feature vectors from the sample entropy values ​​of each IMF, and automatically classify all IMFs into high-frequency classes C using the K-means clustering algorithm. high and low-frequency C low Reconstructing the high-frequency component x high (t) and low-frequency component x low (t): 。 4. The method for diagnosing internal short-circuit faults in lithium-ion batteries based on multimodal feature fusion according to claim 3, characterized in that, S3 specifically includes: The high-frequency component x is obtained by solving the variational constraint problem. high (t) is decomposed into several modal components with finite bandwidth u i (t), the constraint condition is the modal components u of each finite bandwidth. i The sum of (t) equals the high-frequency component x. high (t); By introducing a quadratic penalty factor β and a Lagrange multiplier λ, and using the alternating direction multiplier method for iterative solution, the characteristics of millivolt-level transient voltage drop caused by internal short circuit are extracted.

5. The method for diagnosing internal short-circuit faults in lithium-ion batteries based on multimodal feature fusion according to claim 4, characterized in that, S5 specifically includes: S51: Extract key statistical features from each component in the multimodal feature set to construct a 6-dimensional feature vector f. m : Where μ is the mean, σ² is the variance, Kurt is the kurtosis, Skew is the skewness, E is the energy, and SampEn is the sample entropy; The feature vectors of all components are combined to obtain the initial fused feature set, thus completing the alignment and splicing of multimodal features; S52: The eigenvector f of the i-th component i The input is a two-layer fully connected network. The first layer uses the tanh activation function, and the second layer outputs the raw attention score s. i : Among them, W b The weight matrix is ​​6×6, where d is the bias and w is the weight matrix. a This is the weight vector; Then, the weight coefficients α are obtained by normalization using the Softmax function. i : S53: Based on the weighting coefficient α i For the eigenvector f of each component i By performing weighted summation, a fusion feature sequence F that uniformly represents the fault evolution process is obtained. fusion : 。 6. The method for diagnosing internal short-circuit faults in lithium-ion batteries based on multimodal feature fusion according to claim 5, characterized in that, S6 specifically includes: S61: Fuse the feature sequence F fusion The sample is divided into multiple samples according to a fixed time window. Simultaneously, the internal short-circuit resistance R generated using offline calibration or simulation is used. ISC As labels, a training dataset is constructed and divided into training and validation sets proportionally; S62: Construct a multi-scale Transformer encoder module. The input sample first passes through three parallel convolutional layers to extract multi-scale temporal patterns. The outputs are summed and then normalized to obtain a multi-scale embedding matrix. Subsequently, a segmented sparse multi-head self-attention mechanism is adopted to divide the multi-scale embedding matrix into non-overlapping segments along the time dimension. Standard multi-head self-attention is calculated in each segment, and cross-segment attention connections are introduced between adjacent segments to form a sparse interaction pattern. After passing through multiple encoding layers, the output is adaptively averaged pooled along the time dimension to obtain the global feature vector. S63: Construct an attention-enhanced bidirectional gated recurrent unit module to perform temporal modeling on the input samples along both the forward and backward directions to obtain the hidden state at each time step. For each time step, the gating coefficients are calculated based on the current hidden state and input features. The hidden state is then adaptively weighted to obtain the gated enhanced hidden sequence. Attention weights are calculated on the gated enhanced hidden sequence and then summed to obtain the branch output vector. S64: Introduce a cross-attention interaction module. Use the global feature vector as the query and the branch output vector as the key and value to calculate cross-attention. Conversely, use the branch output vector as the query and the global feature vector as the key and value to calculate cross-attention in the other direction. Add the interaction result to the original global feature vector and the branch output vector and normalize them layer by layer to obtain the two enhanced branch features. S65: The enhanced features of the two branches are concatenated and input into the gating network to obtain the fusion weights of the two branches. The enhanced features of the two branches are then weighted and summed to obtain the intermediate fusion features. Finally, a fully connected layer is used to linearly transform the features back to the original feature dimensions, and the enhanced fusion feature sequence is output.

7. The method for diagnosing internal short-circuit faults in lithium-ion batteries based on multimodal feature fusion according to claim 6, characterized in that, Specifically, S7 includes: S71: Collect multiple sets of fused feature samples of the battery pack when it is running in a healthy state, randomly sample from them to construct a reference dataset, and calculate the local outlier factor for each sample point in the reference dataset. S72: Analyze the local outlier factors of all sample points in the reference dataset and determine the warning threshold of the local outlier factors; S73: During online operation, the local outlier factor of the enhanced fusion feature sequence sample point corresponding to the current time relative to the reference dataset is calculated in real time. If it exceeds the warning threshold, an internal short-circuit warning is triggered and the warning time is recorded.

8. The method for diagnosing internal short-circuit faults in lithium-ion batteries based on multimodal feature fusion according to claim 7, characterized in that, The calculation of the local outlier factor for each sample point in the reference dataset in step S71 specifically includes: Determine the k nearest neighbors of each sample point, calculate the local reachability density of each sample point, which is defined as the reciprocal of the average reachability distance from the sample point to its k nearest neighbors. The local outlier factor of the sample point is the average of the ratios of the local reachability densities of all sample points in the neighborhood of the sample point to the local reachability density of the sample point itself.

Images