Battery test abnormity diagnosis method and system based on multi-modal data fusion

Abstract

The invention discloses a battery test abnormity diagnosis method and system based on multi-modal data fusion. The battery test abnormity diagnosis method comprises the following steps: synchronously acquiring multi-mode information data reflecting a battery state; preprocessing the collected multi-modal information data and then extracting high-dimensional features; inputting the high-dimensional features into a pre-trained multi-modal anomaly detection model of a double-layer model structure for detection, and outputting a comprehensive anomaly score; and comparing the comprehensive abnormity score with a set threshold value, judging whether abnormity occurs or not according to a comparison result, and triggering an alarm according to an alarm control strategy when abnormity is confirmed. According to the invention, early warning can be realized, the false alarm rate is reduced, and evaluation and processing suggestions for abnormal reasons, abnormal types and risk levels are provided.

Description

Technical Field

[0001] This invention relates to the field of battery testing technology, and in particular to a battery testing anomaly diagnosis method and system based on multimodal data fusion. Background Technology

[0002] With the rapid development of electric vehicles, energy storage systems, and other fields, lithium-ion batteries have become a key energy storage device for modern socio-economic operations. Due to the inherent risks of the lithium-ion battery chemical system and the extremely complex application environment, batteries are susceptible to thermal runaway under abuse conditions (such as overcharging, over-discharging, short circuits, high temperatures, punctures, and compression), which may lead to serious safety accidents such as fires and explosions. Therefore, batteries must undergo rigorous testing before leaving the factory to assess their performance limits and safety design boundaries. Battery testing has become an indispensable core part of the battery design, manufacturing, and certification process to evaluate its performance limits and failure modes.

[0003] Traditional battery testing anomaly diagnosis relies primarily on monitoring single parameters (such as voltage and temperature) with fixed thresholds. This approach typically triggers alarms only after an anomaly has occurred or even caused irreversible damage, failing to provide early warning. When faced with normal parameter fluctuations during battery testing, the fixed threshold method cannot distinguish between normal fluctuations and true anomaly precursors, leading to frequent false alarms. Furthermore, focusing solely on a single parameter neglects the coupling relationships and coordinated trends among multiple parameters such as voltage, current, temperature, internal resistance, and deformation, making it difficult to identify potential faults.

[0004] Furthermore, traditional battery testing anomaly diagnosis relies on personal experience, resulting in low efficiency and poor consistency. Although current testing anomaly diagnosis technologies have evolved from simple threshold alarms to data-driven intelligent diagnosis, significant shortcomings remain in accuracy, timeliness, and intelligence. Therefore, real-time and accurate anomaly diagnosis and early warning during testing are crucial for ensuring testing safety, improving testing efficiency, and enhancing battery R&D capabilities. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings and defects of the prior art and provide a battery test anomaly diagnosis method and system based on multimodal data fusion, which aims to achieve early warning, reduce false alarm rate, and provide anomaly cause, anomaly type, risk level assessment and handling suggestions.

[0006] One objective of this invention is to provide a battery testing anomaly diagnosis method based on multimodal data fusion, which synchronously collects multimodal information data reflecting the battery state, wherein the multimodal information data includes at least electrical parameters, thermal parameters, mechanical parameters, acoustic parameters, and visual parameters.

[0007] After preprocessing the collected multimodal information data, high-dimensional features are extracted. The high-dimensional features include at least time-domain, frequency-domain, time-frequency-domain, and spatiotemporal features characterizing changes in battery state.

[0008] The high-dimensional features are input into a pre-trained multimodal anomaly detection model with a two-layer model structure for detection, and a comprehensive anomaly score is output. The comprehensive anomaly score is compared with a set threshold, and an anomaly is determined based on the comparison result. If an anomaly is confirmed, an alarm is triggered according to the alarm control strategy.

[0009] Preferably, the battery state characterized by the high-dimensional features includes at least the following electrical features: voltage change rate, plateau voltage length, static voltage drift, current change rate, shelving leakage current, and DC internal resistance; thermal features: temperature change rate, maximum temperature difference, and surface temperature uniformity; mechanical features: the ratio of internal pressure to design withstand pressure, surface extrusion / puncture force, and pressure change rate of the battery casing; acoustic features: pulse energy, characteristic frequency peak, and number of acoustic emission events of the acoustic emission signal inside the battery; and visual features: leakage area, surface protrusion height, surface crack length, deformation rate of abnormal points, hot spot diffusion location, and infrared hot spot fixing base in the image.

[0010] Preferably, the multimodal anomaly detection model with the two-layer model structure includes:

[0011] Multiple single-modal baseline sub-models, each of which processes one type of modality information in multimodal information, are used to independently characterize the data distribution pattern of a single modality information under normal conditions based on supervised learning model / unsupervised learning model, and output the abnormal deviation score of the single modality information itself;

[0012] The cross-modal association sub-model treats each modal feature as a graph node and the inter-modal association as an undirected edge. The weight of each undirected edge represents the association strength. The graph attention network processes the feature vectors of multiple input graph nodes. For any two nodes, the mapped features are concatenated, and the original attention score is calculated through a linear layer and the LeakyReLU activation function. The original attention score of each node and all its neighboring nodes is normalized using the Softmax function to obtain the edge weight of the undirected edge. The features of each node and the features of its neighboring nodes are weighted and summed according to the edge weight to calculate the aggregated node features. The aggregated node features of the cross-modal point pairs are concatenated into a feature vector, which is then passed through a fully connected layer, multiple linear transformations, and activation to obtain a single real-valued input. The Sigmoid function maps to [0,1] and outputs the cross-modal association anomaly score.

[0013] Preferably, the comprehensive anomaly score is obtained by weighted summing of the anomaly deviation score of the individual modal information output by each single-modal baseline sub-model and the cross-modal correlation anomaly score output by the cross-modal correlation sub-model. The comprehensive anomaly score of normal samples is close to 0, and the comprehensive anomaly score of abnormal samples exceeds a set threshold.

[0014] Preferably, the alarm is triggered according to the alarm control strategy, and based on the magnitude of the comprehensive anomaly score, there are at least three levels of early warning: low-risk warning, medium-risk warning, and high-risk warning, with each risk warning corresponding to a different test and handling strategy.

[0015] Preferably, after triggering the alarm according to the alarm control strategy, the method further includes:

[0016] The diagnostic reasoning process is initiated to analyze and judge the contribution of each feature in the comprehensive anomaly score and locate the feature that causes the anomaly. The feature whose contribution to the anomaly is greater than the threshold and its changing trend are matched with the feature fingerprint of the pre-set battery fault knowledge graph, and a structured diagnostic prompt is output based on the matching result.

[0017] Preferably, the pre-set battery fault knowledge graph storage includes a variety of typical faults, and the typical faults include at least the multi-parameter abnormal fingerprint patterns corresponding to internal short circuit, lithium plating, separator rupture, and gas production in the battery.

[0018] Preferably, the structured diagnostic prompts include at least the location of the abnormality, the type of abnormality, the cause of the abnormality, risk level assessment, and treatment recommendations.

[0019] Preferably, the matching of the feature fingerprint includes at least core feature matching, numerical range matching, and trend consistency matching.

[0020] Another objective of this invention is to provide a battery test anomaly diagnosis system based on multimodal data fusion, used to implement the battery test anomaly diagnosis method, comprising:

[0021] The data acquisition module is used to simultaneously acquire multimodal information data reflecting battery status using multiple sensors;

[0022] The data processing and storage module is used for preprocessing and storing the collected multimodal information data;

[0023] The anomaly diagnosis and analysis module is used to extract high-dimensional features from the preprocessed multimodal information data, input the high-dimensional features into a pre-trained two-layer model structure multimodal anomaly detection model for detection, and output a comprehensive anomaly score; compare the comprehensive anomaly score with a set threshold, determine whether it is abnormal based on the comparison result, and trigger an alarm according to the alarm control strategy when an anomaly is confirmed.

[0024] This invention relates to a battery testing anomaly diagnosis method and system based on multimodal data fusion. By analyzing the weak precursor features of multiple parameters and their correlated changes, it can identify abnormal signals earlier than the BMS alarm threshold before the battery experiences a significant fault, thus providing early warning before the battery experiences a significant fault and avoiding irreversible damage. This solves the problem of delayed warnings in existing technologies. By fusing multimodal information such as electrical, thermal, mechanical, acoustic, and visual data and utilizing a cross-modal correlation model to achieve modal information complementarity, it can effectively distinguish between normal battery fluctuations and real anomalies, reducing false alarms and avoiding abnormal test termination, thus solving the problem of high false alarm rates in existing technologies.

[0025] This application introduces a multimodal anomaly detection model with a dual design of multiple single-modal baseline sub-models and cross-modal correlation sub-models. By combining contribution analysis and battery fault knowledge graph matching, it can diagnose abnormal situations and provide systematic prompts for anomaly locations and causes, such as providing battery test anomaly causes, anomaly types, risk level assessments, and handling suggestions. This enables the automation and intelligence of anomaly diagnosis, which is beneficial to improve the accuracy of battery anomaly diagnosis under actual testing conditions, reduce reliance on human experience, improve diagnostic efficiency and consistency, and provide clear decision support for human personnel. Attached Figure Description

[0026] Figure 1 This is a flowchart of the battery testing anomaly diagnosis method based on multimodal fusion of the present invention.

[0027] Figure 2 This is a schematic diagram of the battery test anomaly diagnosis system based on multimodal fusion of the present invention.

[0028] Figure 3 This is a schematic diagram of the human-machine interface of the battery test anomaly diagnosis system based on multimodal fusion according to the present invention. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0030] See Figure 1 As shown in the exemplary embodiment of this application, the battery test anomaly diagnosis method based on multimodal fusion includes the following steps:

[0031] S1. Synchronously collect multimodal information data reflecting the battery status, wherein the multimodal information includes at least electrical parameters, thermal parameters, mechanical parameters, acoustic parameters, and visual parameters;

[0032] S2. After preprocessing the collected multimodal information data, high-dimensional features are extracted. The high-dimensional features include at least time-domain, frequency-domain, time-frequency-domain, and spatiotemporal features that characterize changes in battery state.

[0033] S3. Input the high-dimensional features into a pre-trained multimodal anomaly detection model with a two-layer model structure for detection, and output a comprehensive anomaly score; compare the comprehensive anomaly score with a set threshold, determine whether it is abnormal based on the comparison result, and trigger an alarm step according to the alarm control strategy when an anomaly is confirmed.

[0034] In step S1, during the synchronous acquisition of multimodal information data, the core software in the industrial computer selects the test type (such as electrothermal performance, mechanical reliability performance, safety reliability performance, environmental reliability performance, cycle durability performance, etc.), connects the equipment required for the test, and simultaneously arranges the required sensor array. When the test begins, the integrated sensor array is used to synchronously and in real time acquire multimodal physical quantity data reflecting the battery status.

[0035] In some embodiments, the sensor array acquires data including at least:

[0036] Electrical parameters: The electrical parameters include the battery voltage and current, which are acquired by high-precision voltage and current sensors.

[0037] Thermal parameters: The thermal parameters include the temperature of multiple points on the battery surface and the internal temperature of the battery, which are collected by thermocouples;

[0038] Mechanical parameters: The mechanical parameters include the internal pressure of the battery box collected by pressure sensors, the compressive or puncture force applied to the battery by the equipment, and the strain value collected by stress plates attached to the surface of the casing.

[0039] Acoustic parameters: The acoustic parameters include acoustic emission signals generated inside the battery, acquired by a piezoelectric acoustic emission sensor;

[0040] Visual parameters: The visual parameters include a sequence of battery deformation images captured by a high-speed visible light camera and a sequence of temperature field distribution images captured by an infrared thermal imager.

[0041] It should be noted that the above-mentioned multimodal data acquisition dimensions can be refined according to the battery test category. For example, basic electrothermal performance testing can be expanded to include dimensions such as state of charge (SOC), insulation resistance, and capacity. Data that cannot be acquired by sensors can be obtained from signals reported by the BMS. Mechanical reliability can be expanded to include dimensions such as vibration or shock acceleration and needle penetration displacement. Data that cannot be acquired by sensors can be obtained from the test equipment. Environmental reliability can be expanded to include internal humidity and insulation resistance data acquired by miniature humidity sensors, which can be customized according to test requirements.

[0042] During the test, the collected electrical, thermal, and acoustic signals are connected to the analog channel of the synchronous data acquisition card. The synchronous acquisition card triggers the output port, such as a TTL pulse, to send a synchronization signal to the visible light high-speed camera and the infrared thermal imager. The data collected by each signal is transmitted to the industrial computer through an independent interface. The equipment data required for the test is synchronously uploaded to the industrial computer, and the software layer synchronization alignment is achieved based on time.

[0043] In step S2, the raw data collected in step S1 is preprocessed, including denoising, filtering and normalizing all data component signals after alignment, and mapping signals with different physical dimensions to a unified scale, such as [0, 1] or [-1, 1], so as to correct data deviation, eliminate invalid interference, unify data scale and ensure data quality.

[0044] Among them, the extraction of high-dimensional features involves extracting time-domain, frequency-domain, time-frequency-domain, and spatiotemporal features that can characterize state changes from the preprocessed modal data, which together form a high-dimensional feature vector.

[0045] The characteristics representing state changes include electrical features such as voltage change rate, plateau voltage length, static voltage drift, current change rate, shelving leakage current, and DC internal resistance; thermal features such as temperature change rate, maximum temperature difference, and surface temperature uniformity; mechanical features such as the ratio of internal pressure to design withstand pressure, surface extrusion / puncture force, and pressure change rate of the battery casing; acoustic features such as pulse energy, characteristic frequency peak, and number of acoustic emission events of the acoustic emission signal inside the battery; and visual features such as the area of ​​the leakage region, surface protrusion height, surface crack length, deformation rate of abnormal points, hot spot diffusion location, and infrared hot spot fixing base in the image. Specifically, the corresponding feature vectors can be extracted according to the test category.

[0046] In this embodiment of the application, in step S3, when performing multimodal anomaly detection based on the multimodal anomaly detection model with a two-layer model structure, the high-dimensional feature vector obtained in step S2 is input into the pre-trained multimodal anomaly detection model for detection, and then the model outputs the detection result.

[0047] The multimodal anomaly detection model described in this application adopts a two-layer analysis architecture. The first layer consists of multiple single-modal baseline sub-models, that is, a baseline model is established for each type of modal data in electrical, thermal, mechanical, acoustic, and visual modes. Each single-modal baseline sub-model independently characterizes the data distribution pattern of each test item in the normal state of each modality, and is used to detect the abnormal deviation score S of a single modality itself. i The second layer is a cross-modal correlation sub-model. Based on the graph attention network, it detects whether the coupling relationship between different physical quantities of battery state changes is appropriate and outputs a cross-modal correlation anomaly score.

[0048] Among these, the multiple single-modal baseline sub-models are selected based on the requirements of battery testing. For example, for models with clear judgment conditions or parameter thresholds, a support vector machine (SVM) supervised learning model can be selected, such as one for processing electrical and thermal modal information. The SVM supervised learning model uses a linear kernel function, and the loss function is minimized by min... w,b,ξ ½||w|| 2 +C∑ n i=1 ξi, w are the hyperplane normal vectors, b is the bias term, ξi is the slack variable, C>0 is the penalty coefficient, used to control the penalty for misclassified samples, and the constraint condition is: y i (w·x i (b)≥1-ξ i (ξ i ≥0).

[0049] Specifically, the process is achieved through the following steps:

[0050] Determine the judgment conditions that each single module must meet, such as a voltage surge amplitude greater than 0.2V and a static leakage current greater than 50μA being considered abnormal. Extract the D-dimensional feature vectors corresponding to the judgment conditions from the raw data of each single mode, and use them as the input feature sequence X, X={x1,x2,x3……,x…} for the SVM. D This includes features such as voltage change rate (dV / dt) and static leakage current; each feature X is converted into a standardized feature by selecting a normalization formula according to its data type. The input is used to optimize the model. For data with an approximate normal distribution, Z-score standardization is used, and for others, Min-Max normalization is used. If the physical threshold of any one of the original features in the feature vector X meets the anomaly detection criteria, it is labeled as an anomaly with y=1. If all features meet the normal range (e.g., x1=0.1v, x2=30μA), they are labeled as normal with y=0. The labeled dataset is divided into training and validation sets in a 7:3 ratio, with 70% allocated to the training set for establishing mapping relationships. , 30% is allocated to the validation set for model optimization. , ;

[0051] The penalty parameter C is tuned using the training set, and then the optimal model is validated or tested using the validation set to verify the model's accuracy, ensuring that the anomaly identification accuracy is ≥95%. If this is not achieved, the penalty parameter C is further optimized until it meets the standard, allowing the optimal SVM model to score new samples. The SVM's decision values ​​are fitted using the sigmoid function, outputting the probabilities of belonging to normal samples (y=0) and abnormal samples (y=1). The anomaly probability is taken as the single-modal anomaly score, and finally, the single-modal anomaly deviation score S is output. i .

[0052] For modal information processing without explicit decision conditions or parameter thresholds, an autoencoder unsupervised learning model can be chosen, such as for audio and visual modal information. The autoencoder in this unsupervised learning model uses mean squared error (MSE) as the loss function: Loss = 1 / N*D∑ N i=1 ∑ D j=1 (xi,jx^i,j) 2 N is the number of samples, D is the feature dimension, where x i,j Let x^i,j be the j-th input feature of the i-th sample, and x^i,j be the j-th reconstructed feature of the i-th sample. A smaller loss indicates more accurate reconstruction. The complete processing steps are as follows:

[0053] An encoder is used to compress the feature vector D into a low-dimensional vector D1 to extract core features. A decoder is then used to restore the D1-dimensional latent vector to D-dimensional features, aiming to make it as close as possible to the input features. The loss function uses mean squared error (MSE) as the loss, and the optimizer is Adam to quickly find the parameters that minimize the loss. Training stops when the MSE loss on the validation set stabilizes. The trained autoencoder is used to reconstruct the validation set, and the reconstruction error MSE of each sample is calculated. The 95th percentile of the reconstruction error of normal samples is taken as the anomaly threshold T. For each newly input unimodal feature vector Xnew, its reconstruction error MSE is calculated and compared with the threshold T to obtain the unimodal anomaly deviation score Si.

[0054] In this embodiment, the cross-modal correlation sub-model uses a GAT (Graph Attention Network) capable of capturing complex relationships. Each modality v is considered a node in a graph, V={v1,v2,v3,v4,v5}, where v1 is an electrical signal, v2 is a thermal signal, v3 is a force signal, v4 is an acoustic signal, and v5 is a visual signal. The correlation strength β between each set of modal features is considered as the weight of the edge, β={e12,e13,…,e45}, which includes 10 edges: electro-thermal, electro-force, electro-acoustic, electro-visual, thermal-force, thermal-acoustic, thermal-visual, force-acoustic, force-visual, and acoustic-visual. The initial value of β is random. Edges with different levels of sensitivity can be defined according to the characteristics of the battery test project. For example, in thermal runaway tests, electro-thermal and thermal-visual edges are sensitive edges with higher weights, while force-visual and acoustic-visual edges are relatively weakly correlated edges with lower weights. Subsequently, the GAT learns and updates to better match the physical correlation patterns in normal samples.

[0055] The complete processing steps of this cross-modal association sub-model are as follows: The five modalities mentioned above are treated as five nodes in a graph, with the feature vector of each node derived from the output of the corresponding single-modal baseline sub-model; ten sets of undirected edges are defined between modalities, with the weight of each edge representing the association strength; initial weights are set, and subsequent weights are dynamically updated by GAT; the feature vectors and labels of the five modalities, along with normal / abnormal samples, are input into the cross-modal association sub-model. During detection, the input to the cross-modal association sub-model is the feature matrix H∈R of the five nodes. 5×mn Five nodes, each with D-dimensional features, are transformed by a shared linear transformation matrix W∈R. D ×F (where F is the mapped dimension) maps the low-dimensional features of each node to a high-dimensional space, enhancing its expressive power, denoted as hi′, where hi′ = W·hi, and hi The original features of node vi, hi′∈R F For any two nodes vi and vj, concatenate their mapped features hi′ and hj′, and calculate the original attention score e using a linear layer a and a LeakyReLU activation function. ij =LeakyReLU(a·[hi′∥hj′]), where (hi′, hj′) represents feature concatenation with a dimension of 2F; the Softmax function is used to normalize the original scores of node vi and all its neighboring nodes to obtain the final edge weight (attention weight) α. ij α ij =Softmax j (e ij )=exp(e ij ) / ∑ k∈N(i) exp(e ik N(i) represents all neighboring nodes of node vi, and e ij Let ∑ represent the original attention score of vi to vj.k∈N(i) exp(e ik The sum of the original scores of node i and all its neighboring nodes k is calculated; the features of each node i are weighted by attention weight α with the features of its neighboring nodes j. ij Summing yields the aggregated node features h. i ′′, h i ′′=σ(∑ j∈N(i) α ij ·h j σ is the ReLU activation function; finally, the cross-modal point pair (i,j) is determined, and the aggregated features (hi′′, hj′′) are concatenated to form the associated feature vector X, which is fed into a fully connected layer of length 2F, and then subjected to multiple linear transformations and activations to obtain a single real number Z, which is substituted into the Sigmoid function to map to [0,1]. cross = 1 / 1+e -z Output cross-modal correlation anomaly score S cross S cross A larger value indicates a more disordered association.

[0056] This cross-modal correlation sub-model can detect misalignments in the coupling relationships between different physical quantities, such as whether temperature changes match the current heat generation rate or whether deformation is synchronized with internal pressure changes. Normal samples in this cross-modal correlation sub-model can be trained using multiple tests with normal data, while abnormal samples can be trained by accumulating multiple tests with abnormal data or by artificially disrupting the correlation of a certain edge, controlling the change of one edge and observing the changes of other edges. Test samples are labeled (0 = normal, 1 = abnormal).

[0057] As mentioned above, the multimodal anomaly detection model, by fusing the output detection results of the aforementioned two-layer model, provides a calculated comprehensive anomaly score S. total The comprehensive anomaly score S total The anomaly score S is the output of each single-modal baseline sub-model. i The correlation anomaly score S output by the cross-modal correlation sub-model cross Weighted summation, normal sample S total S is close to 0, indicating anomaly. total Exceeding the set threshold T total .

[0058] In this embodiment of the application, after obtaining the comprehensive anomaly score, when making dynamic early warnings based on the comprehensive anomaly score, the comprehensive anomaly score S is compared. total Exceeding the set threshold T total Whether or not, dynamic alarms are implemented.

[0059] Specifically, in implementation, the correlation anomaly score S output by the cross-modal correlation sub-model is defined. crossIn the comprehensive abnormal score S total The weights δ and the anomaly scores S output by each single-modal baseline sub-model are: i In the comprehensive abnormal score S total The weight γ in the algorithm outputs multiple sets of score weight combinations. The scores weight combinations are multiplied by the corresponding abnormal scores and weighted to calculate the comprehensive abnormal score. For each set of score weight combinations (γ1~γ5, δ), the effectiveness of the set of parameters is finally evaluated by the accuracy of the test abnormality judgment to determine the optimal combination, which can at least guarantee the correct distinction of 95% of normal and abnormal sample data.

[0060] In this embodiment of the application, during battery test anomaly diagnosis, the comprehensive anomaly score S is... total Score and set threshold T total The system compares the data and, once a threshold is exceeded, immediately triggers an early warning, thus issuing an alarm before the battery shows obvious physical damage. The alarm has at least three levels, as illustrated below:

[0061] Level 1 Warning: T total total ≤T total +0.2, a pop-up message in the human-computer interaction system interface (red icon) indicates low risk, but the battery status (such as abnormal thermal-electrical-visual correlation and acoustic emission) is slightly abnormal. Closely monitor the battery status.

[0062] Level 2 Warning: T total +0.2 total ≤T total +0.4, a pop-up window on the human-computer interaction system interface (red icon) indicates medium risk, and the battery status (such as abnormal thermal-electrical-visual correlation and abnormal acoustic emission) is abnormal. The battery status check is paused.

[0063] Level 3 Warning: S total >T total +0.4, a pop-up window on the human-computer interaction system interface (red icon) indicates high risk, abnormal battery status (such as abnormal thermal-electrical-visual correlation and abnormal acoustic emission), and the test should be stopped immediately.

[0064] Specifically, the T in this application total The threshold is based on the true labels of the test or validation set, combined with model S. total The output distribution determines the initial threshold. For example, in anomaly detection tasks, the model will output a comprehensive anomaly score S. total If the model classifies something as "normal" but the true label is "abnormal," or vice versa, the T-squared value needs to be adjusted based on the actual results. total Threshold.

[0065] ​​In this embodiment of the application, after the alarm is triggered, the system further includes steps of anomaly tracing and intelligent diagnosis, that is, when the warning is triggered, the system starts the diagnostic reasoning process:

[0066] First, a contribution analysis is performed, analyzing the overall anomaly score S. total Contribution ξ of each feature dimension i Size, specific, analysis of contribution ξ i When the size is small, the greater the degree to which the analytical feature deviates from the normal state, and the larger its corresponding weight in the model, the greater its impact on the overall anomaly score S. total Contribution ξ i The larger.

[0067] For the output of the Support Vector Machine (SVM) supervised learning model, the decision function coefficients of the SVM and the weight w of each feature are used. k This reflects its impact on anomaly detection. If the actual value x of a certain feature... k The deviation from the normal mean is Δx k =∣x k -μ k | (μk is the mean of normal samples), then this feature affects a certain outlier score S. i The local contribution is w k ·Δx k Normalize the local contributions of all features within the same modality to obtain the feature's contribution in S. i The proportion α i , k (e.g., the proportion of voltage change rate characteristics in S1);

[0068] For the output of an autoencoder unsupervised learning model, the reconstruction error can be used to measure the degree of feature anomaly; the reconstruction error of a certain feature is the loss. k = (x k -x ^ k ) 2 (x) ^ The larger the value of k (where k is the autoencoder reconstruction value), the more significant the deviation from the normal mode. The loss of all features within the same modality is calculated. k Normalization yields the feature in S i The proportion α i , k (e.g., the proportion of acoustic emission pulse energy characteristics in S4). Combined with single-mode fractional S... i In S total The weight γ in i A certain feature for S total The direct contribution is ξ single,i,k =γ i·S i ·α i,k Where ξ single,i,k This represents the direct contribution of the k-th feature in the i-th mode to a single mode (such as the direct contribution of the maximum temperature difference feature in the thermal mode).

[0069] In this application, the cross-modal correlation score S cross To reflect whether the coupling relationship between modes is disordered, it is necessary to quantify the characteristics and analyze the correlation between S. total The indirect contribution is calculated by extracting the edge weights β related to the modality to which the feature belongs in the GAT model (e.g., features of visual modality v5 need to be associated with β15 (electro-visual), β25 (thermal-visual), etc.), and calculating the impact of the feature on cross-modal association dysregulation. If the feature abnormality causes the association strength between its modality and other modalities to deviate from the normal range (e.g., deformation and pressure changes in visual capture are not synchronized), then its position in S is allocated according to the edge weight β. cross The proportion β in i,j,k Combined with S cross In S total The weight δ in the equation yields the indirect contribution ξ. cross,i,k =δ·S cross·βi,j,k ξ cross,i,k This represents the percentage contribution of the k-th feature of the i-th mode to the association edge between the i-th and j-th modes. The total contribution ξ of this feature is obtained by summing the direct contribution from a single mode and the indirect contribution across modes. i =ξ single,i,k +ξ cross,i,k ξ for all features i Normalization can clarify the role of each feature in S. total The proportion of.

[0070] Using the above method, the contribution ranking of each feature can be obtained. For example, if the ξ of the temperature change rate... i The highest value indicates that single-mode thermal anomaly is the dominant factor. If the voltage-temperature correlation coefficient ξ... cross,i,k The emphasis highlights that cross-modal coupling mismatch is the core issue. This contribution analysis method preserves the independent influence of features in single modes while also reflecting their role in cross-modal correlation, ensuring that contribution calculation is consistent with S... total The weighted logic is consistent, providing an interpretable quantitative basis for subsequent anomaly tracing.

[0071] In this application, through the aforementioned contribution analysis, the main parameters or combinations of parameters causing the anomaly can be accurately located, such as determining that the temperature anomaly is the main cause or the voltage-temperature correlation anomaly is the main cause.

[0072] In this embodiment, after the aforementioned contribution analysis, the currently identified high contribution features and their changing trends (abnormal patterns) are matched with the feature fingerprints of the pre-set battery fault knowledge graph (such as core feature matching, numerical range matching, and trend consistency matching). Finally, a structured diagnostic prompt is output, including but not limited to the abnormality type, abnormality cause, risk level assessment, and handling suggestions, providing clear decision support for human intervention.

[0073] The pre-set battery fault knowledge graph storage includes various typical faults, including but not limited to multi-parameter abnormal fingerprint patterns and fault evolution patterns corresponding to internal short circuits, lithium plating, separator rupture, and gas generation.

[0074] Specifically, when constructing a knowledge graph, one can first label the corresponding abnormal feature fingerprints for each typical fault based on historical fault sample data. When the fault occurs, the feature combination, numerical range and trend of the multimodal features are used as the link between the fault and the sample. A matching degree of >85% is taken as the associated fault of this sample.

[0075] See Figure 2 As shown, this embodiment of the invention also provides a battery test anomaly diagnosis system based on multimodal data fusion, used to implement the battery test anomaly diagnosis method, including:

[0076] The system includes a data acquisition module for synchronously acquiring multimodal information data reflecting battery status using multiple sensors; a data processing and storage module for preprocessing and storing the acquired multimodal information data; an anomaly diagnosis and analysis module for extracting high-dimensional features from the preprocessed multimodal information data, inputting these high-dimensional features into a pre-trained two-layer multimodal anomaly detection model, and outputting a comprehensive anomaly score; comparing the comprehensive anomaly score with a set threshold, determining whether an anomaly exists based on the comparison result, and triggering an alarm according to an alarm control strategy when an anomaly is confirmed; and a human-machine interface for displaying alarm information.

[0077] The data acquisition module consists of various sensors and data acquisition hardware, responsible for synchronously acquiring multimodal data. The data processing and storage module consists of a synchronous data acquisition card and an industrial control computer with a preprocessing module, responsible for data preprocessing (such as data alignment, noise reduction, filtering, and normalization), storage, and historical database management. The historical database management is used for both data tracing and as a database for subsequent model training. The data alignment, using a synchronous acquisition card (such as a PCIe high-precision synchronous DAQ), requires multiple steps depending on the scenario. Analog signals output from electrical, thermal, and acoustic sensors can be directly connected to the high-frequency analog channel of the synchronous acquisition card. Sampling is driven by the card's unified clock, ensuring accurate and reliable timestamps for these signals. The synchronous acquisition card uses a hardware trigger signal (such as a TTL pulse) to send a trigger pulse. Upon receiving the pulse, the camera and thermal imager immediately expose and acquire a frame, ensuring strict alignment between the image frame's exposure time and the sampling time of the aforementioned signals. Each acquired signal is independently transmitted to the computer. The computer maps the acoustic emission signal, visible light frame, and infrared frame onto the same time axis based on the timestamp of each device (accurate to the microsecond level), and synchronously uploads the equipment data required for the test to the industrial computer to achieve high-frequency synchronous output.

[0078] In the computer, data denoising, filtering, and normalization are performed on each signal component. It's worth noting that noise in battery testing may vary with the charging and discharging stages (e.g., higher current noise during fast charging). Adaptive filtering (such as variable window moving average or time-varying threshold wavelet denoising) is required at different stages to dynamically adjust noise characteristics, while preserving characteristic segments during filtering. For electrical and thermal signal normalization, Min-Max normalization can be used: x = (x...) max -x min ) / (xx) min (Based on voltage / current extreme value calculation from the training set), for infrared and visible high-speed cameras, pixel value normalization can be used, and for force signals, x′ = x / x can be used. ref x ref For reference standards (such as the pressure resistance limit of the housing, the maximum output force of the extrusion equipment, the force value corresponding to the yield strength of the material) or Z score Normalization is achieved by x′=(x-μ) / σ (based on the mean μ and standard deviation σ of the training set, suitable for force signals with approximately normal distribution, such as expansion force). For acoustic emission signals, peak normalization can be used, x′= x / max|x|. The specific normalization method should be appropriately adjusted and optimized according to the characteristics of the data and multimodal correlation verification to ensure the robustness of the normalization statistics.

[0079] The aforementioned anomaly diagnosis and analysis module is the core software module of the system, including:

[0080] The feature extraction unit extracts complex features using NVIDIA Jetson + PyTorch / TensorFlow, and uses GPU to accelerate the extraction. Different GPUs are used to process different modal data, such as image and time series data, which need to be loaded quickly from the hard drive.

[0081] The model training and update unit uses dual GPUs with large video memory to support batch loading of multimodal data. The dual-card parallelism accelerates the cross-modal fusion layer. It uses SVM / autoencoder and graph attention network to train the model, which is used to build and optimize the multimodal anomaly detection model in step S3 for real-time detection and early warning.

[0082] The diagnostic inference unit uses a GPU server and converts different model training frameworks to ONNX format via ONNX Runtime (a multi-model compatible platform engine) before unified deployment. It uses Python diagnostic logic and integrates rule engines such as PyDrools to provide anomaly causes and handling suggestions. The FastAPI (API interface) application programming interface is used to synchronize the anomaly types, handling suggestions, and other results generated by the diagnostic unit to the host computer to realize the diagnostic process.

[0083] The human-computer interaction interface has a built-in classification test model library (such as electrothermal performance, mechanical reliability performance, safety reliability performance, environmental reliability performance, and cycle durability performance). It binds different test types, such as mechanical reliability test (such as extrusion and impact test), safety reliability test (such as thermal runaway, short circuit, and needle penetration test), environmental reliability test (such as high and low temperature and humidity cycle test), electrothermal performance test (such as charge and discharge efficiency and internal resistance test), and cycle durability test (such as cycle life and expansion force test), to the corresponding dedicated training model.

[0084] After a user selects a specific test type through the human-computer interaction interface, the interface automatically loads the core algorithm of the pre-trained model corresponding to that category, eliminating the need for manual model file import. This ensures accurate matching between the model call and the test scenario, and the input and output interfaces differ for different test categories. For example, the input interface for mechanical reliability testing displays feature input boxes such as displacement, pressure value, and deformation image, while the interface for safety and reliability testing displays parameter configuration items such as temperature threshold and acoustic emission energy. The output interface for mechanical reliability testing displays force-displacement curves and deformation curves, while the interface for safety and reliability testing displays temperature change curves and acoustic emission energy change curves. This allows for real-time visualization of test data, synchronous display of warning information, causes of anomalies (abnormal scores output by different models), diagnostic reports, and handling suggestions. It also supports user parameter configuration and system interaction, enabling dynamic adjustment of thresholds.

[0085] The technology in this application is based on the fusion of five modes: electrical, thermal, mechanical, acoustic, and visual. It not only includes electrical signal characteristics but also integrates cross-physical domain information such as temperature field, structural deformation, acoustic emission, and surface vision. It captures multi-dimensional anomaly features through a two-layer anomaly detection model with a single-modal baseline model and a cross-modal correlation model. By using multi-physical quantity collaborative diagnosis, it avoids the limitations of a single electrical signal, improves the accuracy of anomaly diagnosis, and is applicable to a wider range of testing projects. In contrast, the aforementioned patents primarily focus on early warning of thermal runaway processes.

[0086] The technology in this application is not limited to single algorithm optimization. It focuses on the characterization of the normal distribution of the modality itself, designs cross-modal correlation sub-models, regards the modality as graph nodes and the correlation strength as edge weights, learns the physical correlation law under normal conditions, focuses on the disorder detection of the coupling relationship between modalities, and forms a complete closed-loop process of early warning-localization-source tracing-suggestion.

[0087] Furthermore, the technology of this application, by synchronously collecting multimodal sensing data during the battery assembly testing process, relies on machine learning models to locate the main abnormal parameters through contribution analysis, and combines a pre-set battery fault knowledge graph to output structured diagnostic results, without the need to disassemble the battery, making it convenient for widespread use.

[0088] Example 1: Abnormal diagnosis of battery needle penetration test based on the present invention

[0089] 1. Synchronous acquisition of multimodal information data

[0090] The lithium-ion battery to be tested is fixed in the battery nail penetration testing machine. The battery test anomaly diagnosis system is started. In the human-machine interface, the safety and reliability performance - nail penetration test category is selected. The system automatically loads the scenario-specific model library, based on the S of previous normal nail penetration test samples. total Output distribution, determine the initial threshold T total =0.3, the three-level early warning interval is simultaneously set as: Level 1 (0.3 < S total ≤0.5), Level 2 (0.5 < S) total ≤0.7), Level 3 (S) total >0.7).

[0091] The required sensor array to be arranged on the battery includes:

[0092] Electrical sensors, specifically high-precision voltage and current sensors, are connected to the two terminals of the battery.

[0093] Thermal sensors: Eight K-type thermocouples are placed on key parts of the battery surface, while an infrared thermal imager is aimed at the battery to record the temperature field distribution of the entire battery surface at a rate of 30 frames per second.

[0094] Mechanical sensor: A pressure sensor is integrated into the needle tip to measure the force change during the needle puncture process in real time. Two stress plates are attached to the surface of the tooling to collect the strain of the shell.

[0095] Piezoelectric acoustic emission sensor: An acoustic emission sensor is coupled to the surface of the battery via a waveguide rod to capture signals generated by internal diaphragm rupture, electrode short circuit, electrolyte boiling, etc.

[0096] All sensors are connected to a synchronous data acquisition card (PCIe high-precision DAQ), which is triggered by TTL pulses. Data is transmitted to an industrial computer in real time, ensuring that all data streams have a unified timestamp.

[0097] 2. Data Preprocessing and High-Dimensional Feature Extraction

[0098] After the test begins, assuming that the voltage drops only slightly by 2% and the maximum temperature rises by less than 5°C 3 seconds after the needle prick, the battery test anomaly diagnosis system will perform the following actions in real time:

[0099] Preprocessing: Low-pass filtering is applied to voltage and current signals to eliminate high-frequency noise; band-pass filtering (e.g., 100kHz-400kHz) is applied to acoustic emission signals to extract effective frequency bands; non-uniformity correction is applied to infrared image sequences.

[0100] Feature extraction: Every 1 second is a time window, and the following features are extracted in real time to form a feature vector, specifically:

[0101] Electrical characteristics: instantaneous voltage value 3.63V, voltage drop rate (dV / dt=-0.023V / s).

[0102] Thermal characteristics: maximum temperature 29℃, average temperature rise rate 0.3℃ / ms, maximum temperature difference on battery surface 5℃.

[0103] Mechanical characteristics: peak needle puncture force 400N, slope of force-displacement curve 60N / mm.

[0104] Acoustic characteristics: The acoustic emission signal has an RMS value of 40dB, an impact number of 80, and a duration of 250ms within a specific frequency band of 200kHz-300kHz.

[0105] Visual characteristics: hot spot area extracted from infrared images is 0.5~1cm², hot spot center movement speed is 0.1~0.3mm / s, and deformation expansion is 0.1~0.3mm.

[0106] 3. Multimodal anomaly detection model based on a two-layer model

[0107] After needle puncture, the voltage drop rate is -0.023V / s, the instantaneous voltage value is 3.63V, the electrical self-encoder calculates the reconstruction characteristics, the reconstruction voltage is 3.66V, the reconstruction rate is -0.012V / s, and the reconstruction error MAE = 1 / 2 [(3.63-3.66)]. 2 +(-0.023-(-0.012)) 2 S1≈0.29, based on the previous normal needle puncture test sample detection model, the maximum sum of reconstruction errors of voltage and rate is 0.00176.

[0108] After acupuncture, the temperature rose from 25℃ to 29℃, an increase of 4℃, with a temperature rise rate of 0.3℃ / ms. The thermal autoencoder calculated the reconstruction error, and the reconstruction normal characteristic temperature was 26℃, the rate was 0.18℃ / ms, the temperature difference was 2℃, and MAE=1 / 3[(29-26)]. 2 +(0.3-0.18) 2 +(5-2) 2 ]≈6.0048, based on the previous normal acupuncture test sample detection model, the maximum sum of reconstruction errors of temperature, rate and temperature difference is 15.8, and normalized to get S2=0.38.

[0109] The needle puncture force is stable at 400N, 60N / mm. The mechanical autoencoder calculates and reconstructs normal characteristics, with a peak force of 400N, a force-displacement slope of 65N / mm, and MAE=1 / 2[(400-400)]. 2 + (60-65) 2 =12.5. Based on the previous normal acupuncture test sample detection model, the maximum value of the reconstruction error of the acupuncture force and force-displacement slope is 138.89, which is normalized to S3=0.09.

[0110] The acoustic emission sensor detected a sudden increase in RMS energy from 20dB to 40dB in the 200-300kHz frequency band, which did not decay for 250m. The reconstructed normal characteristic RMS was 32dB, with 30 impacts and a duration of 100ms. The reconstruction error was calculated as MAE = 1 / 3 [(40-32)]. 2 + (80-30) 2 + (250-100) 2 The maximum sum of reconstruction errors in the detection model is 12853.34, which is approximately 8354.67. After normalization, S4 = 0.65.

[0111] The hot spot area is 0.5~1cm², the hot spot center movement speed is 0.1~0.3mm / s, and the deformation expansion is 0.1~0.3mm. The visual autoencoder calculates the reconstructed features. The reconstructed normal feature area is 0.2cm², the speed is 0mm / s, and the expansion is 0.05mm. The actual reconstruction error MAE is approximately 0.1217 using the above average variance calculation method. The maximum sum of reconstruction errors in the detection model is 0.2898, and after normalization, S5=0.42.

[0112] The GAT model uses the aforementioned five modal features as nodes and inter-modal correlations as edges. Through training, it learns the correlation strength βi of 10 edges under normal conditions. When a single modal feature is input into the model, the normal correlation strength β1 between voltage drop rate and temperature rise rate is 0.85, the actual correlation strength is 0.2, and the deviation is 0.65; the normal correlation strength β2 between needle penetration force and acoustic emission is 0.7, the actual correlation strength is 0.1, and the deviation is 0.6; the normal correlation strength β3 between hotspot area and expansion is 0.60, the actual correlation strength is 0.15, and the deviation is 0.45; the deviations of the remaining seven edges are all <0.2, having a relatively small impact. The weights of the above three edges account for 75%, and GAT aggregates the deviations through an attention mechanism, outputting Scross = 0.78.

[0113] 4. Abnormal Diagnosis and Source Tracing

[0114] The multimodal anomaly detection model calculates a comprehensive score S based on the above. total The optimal weights of the normal acupuncture test model after training and optimization are γ1=0.1, γ2=0.15, γ3=0.05, γ4=0.25, γ5=0.15, δ=0.3, with a total of 1. The threshold trigger T total =0.3; S total =0.1×0.29+0.15×0.38+0.05×0.09+0.25×0.65+0.15×0.42+0.3×0.78=0.316+0.234=0.55, S total =0.55> T total =0.3, triggering a level 2 warning, the system indicates a risk level of: medium.

[0115] Once the warning is triggered, the system initiates the diagnostic reasoning process:

[0116] First, a contribution analysis was conducted. Among the 12 features, the duration K10 of the acoustic emission signal accounted for 33.2%, making it the dominant factor. Combined with its measured value of 250ms, which is much higher than the reconstructed value of 100ms, this indicates that there was continuous mechanical damage inside during the acupuncture process. The electrothermal correlation deviation was 0.65 (normal β=0.85, actual 0.2), with abnormal instantaneous voltage values ​​being the main cause (contributing 16.6%), reflecting that the voltage drop and temperature rise were not synchronized, possibly indicating a local short circuit and delayed heat generation. The thermal-visual correlation deviation was 0.45 (normal β=0.60, actual 0.15), with abnormal hot spot area being the main cause (contributing 15.7%), indicating that the temperature rise and hot spot expansion were not matched, possibly indicating uneven internal heat accumulation.

[0117] High-contribution features (acoustic signal duration > 200ms, voltage-thermal correlation deviation > 0.5, hot spot area > 0.5cm²) were compared with the pre-set fingerprint: the feature combination of the diaphragm micro-rupture fault was matched (matching degree 92% > 85%). The typical mode of this fault is continuous acoustic signal enhancement + voltage and temperature correlation mismatch + hot spot expansion.

[0118] System outputs structured diagnostics:

[0119] Abnormal location: near the battery puncture area (area of ​​concentrated acoustic emission and hot spot).

[0120] Causes of the anomaly: abnormal acoustic emission, abnormal thermo-electrical-visual correlation;

[0121] Anomaly type: Diaphragm micro-rupture accompanied by localized micro-short circuit; Risk level: Medium;

[0122] Recommendation: Pause battery status checks.

[0123] During the testing process of Example 1, the data obtained are shown in Tables 1, 2, and 3 below.

[0124] Table 1

[0125]

[0126] Table 2

[0127]

[0128] Table 3

[0129]

[0130] The diagnostic method of this invention achieves early warning by capturing weak precursors in a single modality and subtle imbalances in cross-modal correlations. This invention employs multimodal information cross-validation to effectively distinguish normal fluctuations from true anomalies, reducing false alarm rates and improving diagnostic accuracy. This invention uses contribution analysis and battery fault knowledge graph matching to indicate anomalies and cause analysis, thus interpreting diagnostic results. This invention reduces reliance on human experience, improves the efficiency and consistency of test anomaly diagnosis, and achieves high automation and intelligence.

[0131] The battery testing anomaly diagnosis method based on multimodal fusion of the present invention, compared with the existing technology that only covers thermal runaway monitoring and single electrical signal mode, includes a more comprehensive multimodal data fusion system, which simultaneously collects five modal data of electrical, thermal, mechanical, acoustic and visual modes. The features covered include, but are not limited to, cross-physical domain information such as voltage and current, temperature, stress and strain, acoustic emission signals, and deformation images. It can be adapted to various battery testing items such as electrothermal performance and mechanical reliability, and has a wider range of applications.

[0132] Furthermore, this invention utilizes a two-layer intelligent diagnostic framework based on a combination of multimodal baseline analysis and cross-modal correlation analysis. By integrating heterogeneous data from multiple sources, including electrical, thermal, mechanical, acoustic, and visual data, it achieves early, accurate, and interpretable diagnosis of anomalies during battery testing. It locates abnormal parameters through contribution analysis, matches fault patterns using a pre-set fault knowledge graph, and ultimately outputs structured diagnostic results including the cause and type of the anomaly, risk level, and handling suggestions. This forms a complete functional closed loop from early warning triggering to decision support. Existing diagnostic technologies are mostly limited to fault type identification (such as only determining whether there is an internal short circuit) and lack structured output of anomaly types, evolution risks, and handling suggestions.

[0133] The technology of this invention differs from machine learning models that rely on a single algorithm for optimization and lack hierarchical design. It incorporates a two-layer analysis architecture that designs multiple single-modal baseline sub-models and cross-modal correlation sub-models. The first-layer single-modal baseline sub-model, based on the requirements of battery testing and with clearly defined judgment conditions or parameter thresholds, adaptively selects a type of supervised learning model (SVM, such as electrical or thermal) or an unsupervised learning model (such as force, sound, or vision) to independently characterize the data distribution of each test item in the normal state for each modality. This is used to detect the abnormal deviation score S of a single modality. i The second layer is a cross-modal association sub-model, which uses a GAT graph attention network, treating modes as nodes and association strength as dynamic edge weights, and outputs a cross-modal association anomaly score S. cross It is specifically designed to capture imbalances in the physical coupling between modes, resulting in superior diagnostic accuracy.

[0134] The early warning strategy of this invention relies on a dynamic threshold adaptive adjustment mechanism, based on the true labels of the test set and validation set and the comprehensive anomaly score S output by the comprehensive model.total Distribution, dynamic optimization of threshold T total This application addresses the issue of poor adaptability of fixed thresholds in complex testing scenarios by iteratively adjusting thresholds based on feedback model misjudgment results. It also overcomes the shortcomings of existing technologies that often use fixed thresholds or set thresholds based on a single modal distribution, lacking dynamic adaptation capabilities. The early warning strategy of this application involves dynamically allocating early warning weights that integrate cross-modal correlation strength, allowing the early warning sensitivity to dynamically adjust with the degree of anomaly in intermodal coupling. This addresses the deficiency in existing technologies where multimodal fusion weights are mostly fixed values, failing to consider the impact of modal correlation strength on early warning priority. Furthermore, the early warning strategy of this application provides a refined response logic for three-level graded early warnings and associates differentiated handling suggestions, overcoming the shortcomings of existing technologies that often only provide a single early warning level or merely indicate anomalies, lacking a tiered response strategy based on risk level.

[0135] The anomaly tracing and intelligent diagnosis strategy of this application has a dual-dimensional contribution quantification mechanism. It accurately calculates the comprehensive anomaly score S for each feature pair by combining single-modal direct contribution with cross-modal indirect contribution. total Total contribution ξ i， Contribution calculation strictly follows S total The weighted logic ensures that high contribution characteristics are consistent with S. total The anomaly causal relationship is explainable, overcoming the shortcomings of existing technologies that often only analyze the importance of features within a single modality (such as single-modal reconstruction error) or only analyze the contribution of cross-modal correlation strength. This application's anomaly tracing and intelligent diagnosis strategy addresses the differences in single-modal sub-models by employing differentiated feature anomaly quantification methods, then unifying them into a contribution value through normalization and weight allocation. This ensures that the feature importance output by different models can be compared horizontally, overcoming the limitation of existing technologies in adapting to the heterogeneity of different sub-models in multimodal scenarios.

[0136] This invention proposes a two-layer analysis architecture. The first layer is a single-modal baseline sub-model, which models five modes, including electrical and thermal modes, and adaptively selects SVM or autoencoder based on the scenario, offering high flexibility. The second layer is a cross-modal correlation sub-model, which utilizes GAT to specifically capture intermodal coupling relationships. The GAT model is explicitly designed to handle 10 sets of cross-modal correlations, and dynamic edge weights β are introduced. i The architecture is more refined and has a greater advantage in accuracy. It locates the main abnormal parameters through contribution analysis and outputs structured diagnostic results by combining the pre-built battery fault knowledge graph, making the function more complete.

[0137] The system of this invention features a microsecond-level cross-modal synchronization and adaptive preprocessing mechanism. It achieves precise timestamp alignment of electrical, thermal, acoustic, and visual signals through a PCIe synchronization DAQ card and hardware triggering. Differential normalization (e.g., truncation of extreme values ​​Min-Max, anti-outlier Z-score) and time-varying noise reduction strategies are designed for different modal data characteristics (e.g., electrical signals with abrupt changes, force signals with impacts), solving the problems of multimodal data time misalignment and dynamic noise interference, and achieving more refined synchronization and preprocessing adaptation. It supports dynamic adjustment of thresholds and parameters through a human-machine interface, outputting a structured report containing anomaly location, cause, and handling suggestions, forming a complete closed loop of synchronous acquisition, real-time detection, intelligent diagnosis, and dynamic interaction. The interface has a built-in classification model library, with different test types bound to corresponding dedicated training models. After the user selects a specific test type, the interface automatically loads the core algorithm of the pre-trained model corresponding to that category, eliminating the need for manual model file import and ensuring accurate matching between model calls and test scenarios. Furthermore, the input and output interfaces differ for different test categories, overcoming the shortcomings of existing technologies that often focus on a single aspect (e.g., only early warning or diagnosis) and lack full-process collaborative capabilities.

[0138] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects. The scope of the present invention is defined by the appended claims rather than the foregoing description, and therefore all changes falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention.

[0139] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A battery test abnormality diagnosis method based on multi-modal data fusion, characterized in that, Includes the following steps: Multimodal information data reflecting battery status is collected synchronously, and the multimodal information data includes at least electrical parameters, thermal parameters, mechanical parameters, acoustic parameters, and visual parameters; After preprocessing the collected multimodal information data, high-dimensional features are extracted. The high-dimensional features include at least time-domain, frequency-domain, time-frequency-domain, and spatiotemporal features that characterize changes in battery state. The high-dimensional features are input into a pre-trained multimodal anomaly detection model with a two-layer model structure for detection, and a comprehensive anomaly score is output. The comprehensive anomaly score is compared with a set threshold, and an anomaly is determined based on the comparison result. If an anomaly is confirmed, an alarm is triggered according to the alarm control strategy.

2. The battery test abnormality diagnosis method based on multi-modal data fusion according to claim 1, characterized in that, The battery state characterized by the high-dimensional features includes at least the following electrical features: voltage change rate, plateau voltage length, static voltage drift, current change rate, shelving leakage current, and DC internal resistance; thermal features: temperature change rate, maximum temperature difference, and surface temperature uniformity; mechanical features: the ratio of internal pressure to design withstand pressure, surface extrusion / puncture force, and pressure change rate of the battery casing; acoustic features: pulse energy, characteristic frequency peak, and number of acoustic emission events of the acoustic emission signal inside the battery; and visual features: leakage area, surface protrusion height, surface crack length, deformation rate of abnormal points, hot spot diffusion location, and infrared hot spot fixing base in the image. 3.The battery test abnormal diagnosis method based on multi-modal data fusion according to claim 1, characterized in that, The multimodal anomaly detection model with the two-layer model structure includes: Multiple single-modal baseline sub-models, each of which processes one type of modality information in multimodal information, are used to independently characterize the data distribution pattern of a single modality information under normal conditions based on supervised learning model / unsupervised learning model, and output the abnormal deviation score of the single modality information itself; The cross-modal association sub-model treats each modal feature as a graph node and the inter-modal association as an undirected edge. The weight of each undirected edge represents the association strength. The graph attention network processes the feature vectors of multiple input graph nodes. For any two nodes, the mapped features are concatenated, and the original attention score is calculated through a linear layer and the LeakyReLU activation function. The original attention score of each node and all its neighboring nodes is normalized using the Softmax function to obtain the edge weight of the undirected edge. The features of each node and the features of its neighboring nodes are weighted and summed according to the edge weight to calculate the aggregated node features. The aggregated node features of the cross-modal point pairs are concatenated into a feature vector, which is then passed through a fully connected layer, multiple linear transformations, and activation to obtain a single real-valued input. The Sigmoid function maps to [0,1] and outputs the cross-modal association anomaly score.

4. The battery test abnormality diagnosis method based on multi-modal data fusion according to claim 3, characterized in that, The comprehensive anomaly score is obtained by weighting and summing the anomaly deviation score of the individual modal information output by each single-modal baseline sub-model and the cross-modal correlation anomaly score output by the cross-modal correlation sub-model. The comprehensive anomaly score of normal samples is close to 0, while the comprehensive anomaly score of abnormal samples exceeds a set threshold.

5. The battery test anomaly diagnosis method based on multimodal data fusion according to claim 1, characterized in that, The alarm is triggered according to the alarm control strategy. Based on the magnitude of the comprehensive anomaly score, there are at least three levels of warning: low risk warning, medium risk warning, and high risk warning. Each risk warning corresponds to a different test and handling strategy.

6. The battery test anomaly diagnosis method based on multimodal data fusion according to claim 1, characterized in that, After triggering the alarm according to the alarm control strategy, the following is also included: The diagnostic reasoning process is initiated to analyze and judge the contribution of each feature in the comprehensive anomaly score and locate the feature that causes the anomaly. The feature whose contribution to the anomaly is greater than the threshold and its changing trend are matched with the feature fingerprint of the pre-set battery fault knowledge graph, and a structured diagnostic prompt is output based on the matching result.

7. The battery test anomaly diagnosis method based on multimodal data fusion according to claim 6, characterized in that, The pre-set battery fault knowledge graph storage contains a variety of typical faults, including at least the multi-parameter abnormal fingerprint patterns corresponding to internal short circuits, lithium plating, separator rupture, and gas production in the battery.

8. The battery test anomaly diagnosis method based on multimodal data fusion according to claim 6, characterized in that, The structured diagnostic prompts include at least the type of abnormality, the cause of the abnormality, the risk level assessment, and treatment recommendations.

9. The battery test anomaly diagnosis method based on multimodal data fusion according to claim 6, characterized in that, The matching of the feature fingerprint includes at least core feature matching, numerical range matching, and trend consistency matching.

10. A battery testing anomaly diagnosis system based on multimodal data fusion, characterized in that, The method for diagnosing battery test anomalies according to any one of claims 1-9 includes: The data acquisition module is used to simultaneously acquire multimodal information data reflecting battery status using multiple sensors; The data processing and storage module is used for preprocessing and storing the collected multimodal information data; The anomaly diagnosis and analysis module is used to extract high-dimensional features from the preprocessed multimodal information data, input the high-dimensional features into a pre-trained two-layer model structure multimodal anomaly detection model for detection, and output a comprehensive anomaly score; compare the comprehensive anomaly score with a set threshold, determine whether it is abnormal based on the comparison result, and trigger an alarm according to the alarm control strategy when an anomaly is confirmed.

Images