Intelligent detection method and system for sheet metal shell of energy storage equipment

Abstract

The invention relates to the technical field of energy storage equipment detection, and discloses an intelligent detection method and system for a metal plate shell of energy storage equipment. The method comprises the following steps: acquiring a multi-modal detection data set containing three-dimensional laser scanning data, surface defect image data and stress distribution time sequence data from a production line in real time; performing spatial topological feature analysis, multi-scale texture feature extraction and dynamic abnormal mode recognition processing on the three types of data to generate corresponding feature vectors; performing cross-modal association fusion processing on the three types of feature vectors to obtain a fusion detection feature set; based on the set, calling a pre-trained deep decision tree model to carry out defect root cause weight distribution, and generating a defect association degree score set of each detection area; and performing joint defect tracing on the multi-modal data according to the score set, and determining a processing link of a defect source and a material anomaly type. According to the method, multi-dimensional real-time detection and accurate traceability are realized, and the comprehensiveness and accuracy of sheet metal shell detection are improved.

Description

Technical Field

[0001] This invention relates to the field of energy storage equipment testing technology, specifically to an intelligent testing method and system for sheet metal casings of energy storage equipment. Background Technology

[0002] In the field of energy storage equipment manufacturing, sheet metal shells serve as core protective and support components, and their quality directly affects the overall operational stability and safety of the equipment. With the large-scale development of the energy storage industry, sheet metal shell production is gradually moving towards higher precision and mass production, making traditional testing methods insufficient to meet current production demands.

[0003] Currently, the industry primarily employs single-dimensional inspection methods for sheet metal casings. Some companies rely on manual visual inspection of surface defects. This method is heavily influenced by the subjective experience of the inspectors and has a low recognition rate for minor scratches and dents. Furthermore, it struggles to achieve real-time monitoring of the production process, easily leading to defective products flowing into subsequent stages. Other companies use 3D scanning equipment to inspect the structural dimensions of sheet metal casings. While this can improve the accuracy of structural inspection to some extent, it cannot simultaneously acquire information on surface defects and stress distribution. Abnormal stress distribution is often a significant cause of cracking and deformation in sheet metal casings during use, making a single structural inspection insufficient for a comprehensive assessment of product quality. Current testing technologies lack the ability to effectively integrate and analyze multi-source testing data. In the sheet metal casing production process, data such as structural dimensions, surface defects, and stress distribution reflect different aspects of product quality information. These data are correlated; for example, surface dents may be accompanied by localized stress concentration. However, current testing methods typically process data from each dimension independently, failing to establish correlations between them. This leads to biased test results and makes it difficult to accurately trace the root cause of defects. When quality problems are detected, workers must check multiple processing stages, including raw materials, stamping, welding, and painting. This process is not only time-consuming but also makes it difficult to accurately pinpoint the problem area, hindering rapid optimization of the production process and consequently affecting production efficiency and product yield. Traditional testing methods mostly employ offline testing, meaning testing is conducted uniformly after product manufacturing. This approach cannot promptly detect anomalies during the production process, and batch quality issues can lead to significant waste of raw materials and increased production costs. As the quality requirements for sheet metal casings in energy storage devices continue to rise, and production pace accelerates, the shortcomings of traditional testing methods in terms of efficiency, accuracy, and quality traceability are becoming increasingly apparent. Therefore, an intelligent testing method is needed that enables multi-dimensional real-time detection, data fusion analysis, and precise defect tracing. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent detection method and system for sheet metal casings of energy storage devices, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides an intelligent detection method for sheet metal casing of energy storage devices, the method comprising: The multimodal detection data set of the sheet metal shell is collected in real time from the production line. The multimodal detection data set includes three-dimensional laser scanning data, surface defect image data and stress distribution time series data. The three-dimensional laser scanning data is subjected to spatial topological feature parsing processing to generate structural deformation feature vectors; the surface defect image data is subjected to multi-scale texture feature extraction processing to generate defect characterization feature vectors; and the stress distribution time series data is subjected to dynamic anomaly pattern recognition processing to generate stress anomaly feature vectors. The structural deformation feature vector, defect characterization feature vector, and stress anomaly feature vector are subjected to cross-modal correlation and fusion processing to generate a fused detection feature set. Based on the fused detection feature set, a pre-trained deep decision tree model is invoked to perform defect root cause weight allocation processing, generating a defect correlation score set corresponding to each detection area of ​​the sheet metal shell. Based on the defect correlation score set, the three-dimensional laser scanning data, surface defect image data, and stress distribution time series data are jointly processed to trace the source of the defects, thereby determining the processing stage and material anomaly type from which the defects originate.

[0006] Preferably, the step of performing spatial topological feature analysis processing on the three-dimensional laser scanning data to generate structural deformation feature vectors includes: The three-dimensional laser scanning data is divided into grids, and the spatial curvature parameters and thickness distribution parameters of each grid cell are extracted. The spatial curvature parameters of adjacent grid cells are encoded using a pre-trained graph neural network to generate local deformation feature vectors. The thickness distribution parameters of each grid cell are concatenated with the corresponding local deformation feature vectors to generate the structural deformation feature vector.

[0007] Preferably, the step of performing multi-scale texture feature extraction processing on the surface defect image data to generate a defect characterization feature vector includes: The surface defect image data is subjected to Gaussian pyramid layering to obtain image slices at different resolution levels; The image slices at each resolution level are processed by extracting directional gradient histogram features through a pre-trained convolutional neural network to generate a multi-scale texture feature matrix. The multi-scale texture feature matrix is ​​subjected to inter-level feature alignment and dimensionality reduction to generate the defect characterization feature vector.

[0008] Preferably, the step of performing dynamic anomaly pattern recognition processing on the stress distribution time-series data to generate a stress anomaly feature vector includes: The stress distribution time series data is processed by sliding window segmentation to extract the mean stress, range and fluctuation frequency parameters within each time window; By using a pre-trained temporal encoder to dynamically compare the stress parameters of adjacent time windows, abnormal fluctuation ranges can be identified. The stress parameters in the abnormal fluctuation range are compared and encoded with those in the normal fluctuation range to generate the stress anomaly feature vector.

[0009] Preferably, the step of performing cross-modal correlation fusion processing on the structural deformation feature vector, defect characterization feature vector, and stress anomaly feature vector to generate a fused detection feature set includes: The structural deformation feature vector is interpolated in the time dimension to align it with the time step of the stress anomaly feature vector. Calculate the Euclidean distance matrix between each texture feature in the defect characterization feature vector and the corresponding spatial feature in the structural deformation feature vector; Based on the Euclidean distance matrix, a two-way feature weighted fusion process is performed on the defect characterization feature vector and the structural deformation feature vector to generate a spatial-texture fusion feature vector. The spatial-texture fusion feature vector and the stress anomaly feature vector are integrated into a temporal feature through a gated loop unit to generate the fusion detection feature set.

[0010] Preferably, the step of calling a pre-trained deep decision tree model based on the fused detection feature set to perform defect root cause weight allocation processing and generate a defect correlation score set corresponding to each detection area of ​​the sheet metal shell includes: The fused detection feature set is input into the feature importance evaluation layer of the deep decision tree model to calculate the contribution of each feature dimension to defect classification. The deep decision tree model generates a set of association rules between feature combinations and defect types for each detection region through its multi-path inference layer. The fusion detection features are dynamically weighted according to the set of association rules to generate the defect association score set.

[0011] Preferably, the joint defect tracing processing of the three-dimensional laser scanning data, surface defect image data, and stress distribution time series data based on the defect correlation score set includes: Filter the score items in the defect correlation score set that exceed the dynamic threshold to generate a high-confidence defect feature subset; The high-confidence defect feature subset is subjected to inverse feature mapping processing to locate abnormal deformation regions in the three-dimensional laser scanning data, abnormal texture regions in the surface defect image data, and stress abrupt change intervals in the stress distribution time series data. The process defect transmission path between the abnormal deformation region, the texture abnormal region and the stress mutation interval is established by a pre-trained causal reasoning network.

[0012] Preferably, establishing the process defect transmission path between the abnormal deformation region, the texture abnormal region, and the stress abrupt change interval through a pre-trained causal inference network includes: The grid coordinates of the abnormal deformation region, the pixel coordinates of the texture abnormal region, and the timestamps of the stress mutation intervals are input into the space-time alignment layer of the causal inference network. The spatial proximity and temporal continuity parameters between each anomalous region are calculated using the graph attention mechanism of the causal inference network. Based on the spatial proximity and temporal continuity parameters, a multi-level causal link from material defects to processing deformation is generated.

[0013] Preferably, the process step and material anomaly type for determining the source of the defect include: The characteristic attributes of each node in the multi-level causal link are analyzed and matched with abnormal equipment parameter records in the preset processing technology knowledge base; Based on the abnormal equipment parameter records, identify the stamping die number and bending process parameter deviation value that caused the material stress concentration; A complete process responsibility chain for sheet metal shell defects is generated by combining the stamping die number with the bending process parameter deviation value.

[0014] Preferably, the present invention also includes an intelligent detection system for sheet metal casing of energy storage equipment, the system including a memory, a processor and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the intelligent detection method for sheet metal casing of energy storage equipment as described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By collecting multimodal inspection data from the production line in real time, the limitations of traditional single-dimensional inspection are overcome, enabling comprehensive monitoring of sheet metal shell structure, surface defects, and stress distribution. The real-time data collection promptly reflects the product quality status during production, avoiding the risks of batches of defective products generated under traditional offline inspection methods. This helps to quickly identify anomalies during production, reducing raw material waste and production cost losses. Targeted feature processing was performed on different types of multimodal detection data, providing a high-quality feature foundation for subsequent data fusion and analysis. Among them, spatial topology feature parsing processing can accurately extract structural deformation information from 3D laser scanning data, and the generated structural deformation feature vector can clearly reflect the deviations of sheet metal shells in terms of size and shape; multi-scale texture feature extraction processing can capture the details of defects of different sizes and types in surface defect images, and even minor scratches and spots can be reflected through defect characterization feature vectors; dynamic anomaly pattern recognition processing can mine the stress anomaly patterns that change over time from stress distribution time series data, and the generated stress anomaly feature vector provides a key basis for evaluating the mechanical properties of sheet metal shells. Cross-modal correlation fusion processing organically combines feature vectors from different dimensions, establishing a correlation between structural deformation, surface defects, and stress distribution. The resulting fused detection feature set can comprehensively and holistically reflect the quality status of the sheet metal shell, avoiding the one-sidedness of detection results caused by processing single feature data. This fusion method fully utilizes the complementarity of data from different modalities, making the detection results more comprehensive and accurate, and able to more realistically reflect the product quality level. The pre-trained deep decision tree model performs defect root cause weighting on the fused detection feature set. Based on the correlation between feature vectors and learning experience from historical data, it can accurately score the defect correlation of each detection area on the sheet metal shell. This scoring set provides clear guidance for subsequent defect tracing, eliminating the need for blind investigation by staff and significantly improving the efficiency and accuracy of defect analysis.

[0016] Joint defect tracing based on defect correlation scoring sets enables the correlation analysis of multimodal detection data with processing steps and material properties, accurately determining the processing step and material anomaly type at the source of the defect. This process not only quickly locates the root cause of quality problems but also provides specific directions for production process optimization. For example, if tracing reveals that the defect originates from unreasonable parameter settings in the stamping process, the stamping process parameters can be adjusted accordingly; if it is determined to be a material anomaly, the raw material supplier can be changed in a timely manner or the material inspection standards can be optimized. Through a virtuous cycle of accurate tracing and process optimization, the production quality and pass rate of sheet metal shells can be continuously improved, while shortening the problem-solving cycle, ensuring the stable and efficient operation of the production line, and further meeting the high-quality and high-reliability requirements of energy storage equipment for sheet metal shells. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent detection method for sheet metal casing of energy storage equipment according to the present invention. Figure 2 A flowchart for generating structural deformation feature vectors; Figure 3 The flowchart for generating stress anomaly feature vectors. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides an intelligent inspection method for sheet metal shells of energy storage devices. The method includes: achieving efficient and accurate defect detection and tracing by integrating multimodal data acquisition, feature extraction, fusion analysis, and intelligent decision-making. Specifically, in the implementation process, three-dimensional laser scanning data, surface defect image data, and stress distribution time-series data of the sheet metal shell are collected in real time from the production line to form a multimodal detection data set. Then, spatial topological feature parsing processing is performed on the three-dimensional laser scanning data to generate structural deformation feature vectors; multi-scale texture feature extraction processing is performed on the surface defect image data to generate defect characterization feature vectors; and dynamic anomaly pattern recognition processing is performed on the stress distribution time-series data to generate stress anomaly feature vectors. These feature vectors are then subjected to cross-modal correlation fusion processing to form a fused detection feature set. Based on this set, a pre-trained deep decision tree model is called to perform defect root cause weight allocation processing, outputting a defect correlation score set for each detection area. Finally, joint defect tracing processing is performed on the original data based on the score set to determine the processing stage and material anomaly type from which the defect originates.

[0020] Example 1: See Figure 2The process involves meshing the 3D laser scanning data. This step uses a high-precision laser scanner to acquire point cloud data of the sheet metal shell surface. A spatial meshing algorithm converts the point cloud data into a regular array of mesh cells, each representing a local region and containing its 3D coordinate information. Subsequently, the spatial curvature parameters and thickness distribution parameters of each mesh cell are extracted. The spatial curvature parameters are calculated using a surface fitting algorithm to determine the surface curvature of the mesh cell, while the thickness distribution parameters are measured using an ultrasonic thickness gauge or a dual-laser scanning system to determine the physical thickness of each mesh cell. This forms a basic dataset containing geometric shape and physical properties. A pre-trained graph neural network encodes the topological relationships of the spatial curvature parameters of adjacent mesh cells. The pre-training process uses scanning data from historical qualified products as the training set to learn the spatial association patterns between mesh cells. In practical applications, the graph structure uses mesh cells as nodes and adjacent relationships as edges. Graph convolutional layers aggregate the curvature features of neighboring nodes, and an attention mechanism calculates the importance weights between nodes. Finally, a local deformation feature sub-vector for each mesh cell is output. These sub-vectors effectively capture local depressions, bulges, or twists in the sheet metal shell. The thickness distribution parameters of each grid cell are concatenated with the corresponding local deformation feature vectors. After normalization, the thickness parameters are concatenated with the local deformation feature vectors along the feature dimension. The concatenated feature vector contains both spatial deformation information and material thickness information. This concatenation process uses a feature fusion algorithm to ensure the scale consistency of the two types of features. The final generated structural deformation feature vector serves as the data output that comprehensively characterizes the integrity of the sheet metal shell structure, providing a foundation for subsequent cross-modal analysis.

[0021] In the implementation of multi-scale texture feature extraction, the surface defect image data is first processed by Gaussian pyramid layering. A multi-resolution image pyramid is constructed using an image processing library. The original high-resolution image is downsampled multiple times to generate image sequences of different scales, with each level corresponding to the observation granularity. The layering method from coarse to fine granularity allows the system to simultaneously capture macroscopic texture features and microscopic defect details. The image slices at each level retain the color and grayscale information of the original image. The image slices at each resolution level are processed by pre-trained convolutional neural network for histogram of oriented gradients (HARQs). The pre-trained convolutional neural network adopts a ResNet-50 architecture finely tuned on an industrial defect dataset, and its convolutional layers automatically extract the gradient features of the image. The calculation process of the histogram of oriented gradients first calculates the pixel gradient direction and magnitude for each image slice, and then statistically analyzes the gradient direction histogram of the local region, finally forming a multi-scale texture feature matrix. This matrix can effectively characterize the texture characteristics of surface defects such as scratches, dents, and corrosion. The multi-scale texture feature matrix is ​​subjected to inter-level feature alignment and dimensionality reduction. Feature alignment uses bilinear interpolation to adjust feature matrices of different resolutions to a uniform size, ensuring that features at each level correspond in space. Dimensionality reduction uses principal component analysis to preserve the main variation direction of texture features, reducing data redundancy while maintaining feature discriminative power. The resulting defect characterization feature vector serves as a quantitative description of surface quality and is used for subsequent cross-modal fusion analysis.

[0022] The entire implementation process utilizes a parallel computing architecture for efficient processing. The meshing of 3D laser scanning data and the pyramid layering of image data are performed simultaneously. The inference processes of the graph neural network and convolutional neural network are executed in a GPU-accelerated environment. Feature stitching and dimensionality reduction operations are performed in batches to ensure the system meets the timeliness requirements of real-time production line inspection. All generated feature vectors are marked with spatial coordinates and timestamps, providing a basis for subsequent defect tracing. Parameter settings during implementation are determined based on extensive process experiments. The unit size of the meshing partition is set according to the typical defect size of sheet metal shells, typically 5mm × 5mm. The number of Gaussian pyramid levels is calculated based on image resolution and the minimum defect size, generally 4-6 levels. The pre-training cycle of the graph neural network typically requires thousands of training samples. During training, adaptive learning rate adjustment and early stopping strategies are used to prevent overfitting. Data quality control is integrated throughout the entire implementation process. 3D laser scanning data undergoes point cloud denoising and coordinate calibration preprocessing, while surface defect image data requires illumination normalization and color correction. All generated feature vectors undergo consistency verification to check the reasonableness of the feature value distribution range; outliers are automatically removed and recalculated to ensure the reliability of subsequent analysis. In terms of system integration, these processing modules are encapsulated as independent microservices. They receive multimodal data input via message queues and publish the feature vectors to the data bus after processing. Standardized data interfaces are used between services, and the format and dimensions of the feature vectors are clearly defined, facilitating calls and further processing by other modules. The entire implementation process runs on a distributed computing framework, supporting horizontal scaling to meet the inspection needs of different production lines.

[0023] Example 2: See Figure 3 In the implementation of dynamic anomaly pattern recognition processing, the stress distribution time-series data is first segmented using a sliding window method. This step transforms continuous stress sensor readings into discrete time segments by dividing the data into windows with fixed time intervals. Each window contains a specific number of data points and is timestamped. The window size setting needs to comprehensively consider the frequency characteristics of stress changes and the real-time requirements of the detection system. Overlapping window technology is usually used to avoid edge effects and ensure that important signal features are not truncated. The window sliding step size is dynamically adjusted according to the sampling frequency and process cycle. The mean, range, and fluctuation frequency parameters of stress within each time window are extracted. The mean is calculated using an arithmetic mean algorithm to reflect the average stress level within that time period. The range is calculated by the difference between the maximum and minimum values ​​to characterize the stress fluctuation range. The fluctuation frequency is converted from the time domain signal to the frequency domain representation using a fast Fourier transform and the dominant frequency component is extracted. These parameters together constitute the basic feature set describing the dynamic characteristics of stress. Each parameter is standardized to eliminate the influence of dimensions, ultimately forming a time window feature vector containing multi-dimensional indicators.

[0024] A pre-trained temporal encoder dynamically compares the stress parameters of adjacent time windows. Based on a long short-term memory network architecture, the encoder's pre-training process uses stress data from normal production conditions to learn inherent patterns of parameter changes. In practical applications, the encoder sequentially processes the feature vectors of consecutive time windows, capturing the temporal dependencies between parameters through its internal gating mechanism. It calculates the degree of difference between the current window and historical windows, and when the difference exceeds a preset dynamic threshold, the system marks that time interval as an abnormal fluctuation interval. The process of identifying abnormal fluctuation intervals employs a multi-indicator joint judgment strategy, considering not only the abrupt changes of individual parameters but also analyzing the correlation changes of multiple parameters. The system maintains a sliding benchmark window containing the average parameter values ​​of several recent normal time windows. The current window parameters are compared with the benchmark values ​​in multiple dimensions. When multiple parameters simultaneously deviate significantly, an anomaly marker is triggered. This design effectively avoids false alarms and improves detection sensitivity. The stress parameters in abnormal fluctuation ranges are compared and encoded with those in normal fluctuation ranges. The comparison process first establishes a probability distribution model for normal parameters, using a Gaussian mixture model to fit the distribution characteristics of historical normal data. For each abnormal range, the deviation of its parameter values ​​from the normal distribution is calculated, and the overall abnormality of multi-dimensional parameters is measured using Mahalanobis distance. These deviations are then normalized and encoded to generate a fixed-dimensional stress anomaly feature vector. The encoding method for the stress anomaly feature vector combines one-hot encoding and numerical encoding. Discrete anomaly types (such as sudden peaks and persistent shifts) are represented using one-hot encoding, while continuous anomaly values ​​are scaled and directly embedded into the feature vector. This vector also contains temporal location information of the abnormal range, including start timestamps and duration, providing a temporal alignment basis for subsequent cross-modal fusion.

[0025] The entire processing flow adopts a streaming computing model. Stress data flows into the system in real time, and windowing and feature extraction are performed immediately. The timing encoder processes each newly arriving window incrementally, avoiding delays caused by batch processing. The system employs a dual-buffer mechanism: one buffer for real-time processing of new data, and the other for background model updates and parameter optimization, ensuring the continuity and stability of the processing. Quality control during implementation includes data verification and model monitoring. Raw data for each time window undergoes signal quality testing to remove invalid data caused by sensor malfunctions or interference. The performance of the timing encoder is continuously optimized through an online learning mechanism. The system periodically updates the baseline model with newly collected normal data, enabling anomaly detection capabilities to adapt to the slow changes in production processes. Regarding parameter configuration, the sliding window length is typically set to encompass a time span of several typical stress cycles. For most sheet metal processing, a window length of 10-30 seconds balances detection sensitivity and computational efficiency. The dynamic threshold is set using an adaptive method, dynamically adjusted based on statistical values ​​of recent data fluctuations to avoid oversensitivity or insensitivity caused by fixed thresholds. During system integration, this processing module is deployed as an independent time-series analysis service. It receives stress sensor data streams through a high-speed data interface, and the processing results are published to the message middleware in real time. The service adopts a microservice architecture design, supporting parallel processing of data from multiple sensors. Each data processing instance maintains an independent state machine, ensuring the isolation and scalability of the processing logic. The anomaly handling mechanism includes a multi-level response strategy. For detected abnormal fluctuations, the system handles them according to severity: minor anomalies are only logged and the model is updated; significant anomalies trigger real-time alarms and notify operators; and severe anomalies directly link with the production line control system and may trigger shutdown protection. All abnormal events have detailed feature descriptions and time locations, providing a complete time-series evidence chain for subsequent defect tracing. The data processing pipeline adopts a fault-tolerant design. When an anomaly is processed within a certain time window, the system automatically skips that window and marks it as invalid, while simultaneously initiating a data repair mechanism to attempt to resume processing. A periodic checkpoint mechanism ensures that the system can continue working from the most recent valid state after a restart, avoiding data loss or duplicate processing. The entire implementation process emphasizes a balance between real-time performance, accuracy, and reliability.

[0026] The implementation of dynamic anomaly pattern recognition processing is illustrated using a specific inspection case on a sheet metal shell production line: A batch of sheet metal shells, numbered B-2089, enters the inspection station after the stamping process. Distributed stress sensors installed on the production line begin to collect stress distribution data on the workpiece surface in real time at a sampling frequency of 1000Hz. Data acquisition continues for approximately 180 seconds throughout the entire processing cycle, forming a complete stress distribution time-series dataset. When implementing sliding window segmentation processing, the system sets the window length to 5 seconds and the sliding step size to 1 second, dividing the 180 seconds of time-series data into 176 overlapping time windows. Each window contains 5000 original stress data points. The system assigns a unique timestamp identifier to each window, with window numbers from W001 to W176 corresponding to different time periods in the processing. This overlapping window design ensures that no brief stress anomaly events are missed. When extracting stress characteristic parameters within each time window, the system calculates three key indicators in parallel: the mean stress reflects the average stress condition within that time period, the range reflects the fluctuation range of the stress value, and the fluctuation frequency extracts the dominant frequency component through Fast Fourier Transform. Taking window W123 as an example, this window corresponds to the period from timestamp T+122s to T+127s, and the calculated mean stress is 85.6MPa, the range is 42.3MPa, and the dominant frequency component is concentrated in the range of 120-150Hz. These parameters constitute the feature vector of this window. The time-series encoder adopts a pre-trained long short-term memory network structure, and its training data contains stress characteristic sequences of 2000 normal workpieces in the past three months. The encoder processes the feature vector of each window sequentially, and remembers the feature patterns of historical windows through its internal state. When processing window W123, the network detects that the mean stress of the current window suddenly rises to 1.8 times the normal range, the range expands to 2.3 times the average level, and the frequency distribution also shows abnormal peaks. These changes trigger the encoder's anomaly detection mechanism. During the dynamic difference comparison process, the encoder maintains a sliding reference interval of 30 windows, which contains feature data from windows W093 to W122, representing recent normal stress patterns. The encoder calculates the Mahalanobis distance between the current window W123 and the reference interval across multiple feature dimensions. The results show that the difference in the mean stress dimension reaches 4.3 standard deviations, and the difference in the range dimension reaches 3.8 standard deviations, significantly exceeding the preset threshold of 2.0 standard deviations. When identifying abnormal fluctuation intervals, the system employs a multi-condition triggering mechanism: when the feature differences of three consecutive windows exceed the threshold, or when a single window is abnormal in multiple dimensions simultaneously, it is marked as an abnormal interval. Window W123 is abnormal in both the mean and range dimensions simultaneously, so the system immediately marks this window as abnormal and continuously monitors subsequent windows W124 and W125. These two windows also show abnormal characteristics, thus confirming that T+122s to T+127s is a continuous abnormal fluctuation interval.In the comparative encoding and processing stage, the system first establishes a probabilistic model of normal stress parameters. This model is trained based on feature data from the most recent 500 normal workpieces and uses a Gaussian mixture model to characterize the joint distribution of multi-dimensional features. For the abnormal interval W123-W125, the deviation of each feature value from the normal distribution is calculated. The deviation of the mean stress is 2.8σ, the range deviation is 2.5σ, and the frequency distribution deviation is 1.9σ. When generating the stress anomaly feature vector, the system aggregates and encodes the abnormal features from multiple windows. The vector contains information such as the anomaly intensity score, duration, main anomaly dimensions, and spatial distribution pattern. In this case, the generated feature vector shows: the anomaly intensity level is 3 (out of 5), the duration is 5 seconds, the main anomaly dimensions are the mean stress and the range, and the anomaly pattern is a sudden peak accompanied by continuous fluctuations. The entire processing is performed in real time on a distributed computing framework. After the stress data is acquired, it immediately enters the pipeline processing, and the processing delay of each window is controlled within 50 milliseconds. The system is equipped with an anomaly early warning mechanism. When an abnormal fluctuation is detected, an alarm is immediately sent to the production line monitoring system to remind operators to pay attention to the processing quality of the workpiece. During the data quality monitoring process, a sensor noise interference was detected at time T+125s. The system automatically corrected the anomaly by comparing data from adjacent sensors and using filtering algorithms to avoid false alarms. The processing results were correlated with other detection data of the workpiece to provide a complete record of stress anomalies for subsequent defect tracing analysis. Ultimately, the workpiece was marked as a suspicious product awaiting further inspection and entered the manual re-inspection process.

[0027] Example 3: In the implementation of cross-modal correlation fusion processing, time-dimensional interpolation is performed on the structural deformation feature vector. Since the structural deformation data and stress data may have asynchronous acquisition times, a cubic spline interpolation algorithm is used to resample the time series of the structural deformation feature vector. The interpolation process uses the timestamp of the stress data as the reference point to calculate the corresponding value of the structural deformation feature at each new time point, ensuring that the two types of feature vectors have the same time step and temporal correspondence. This processing method lays the foundation for time-series alignment for subsequent multimodal fusion. When calculating the degree of difference between each texture feature in the defect characterization feature vector and the corresponding spatial feature in the structural deformation feature vector, an improved distance metric method is used. This method not only considers the Euclidean distance of the feature values ​​but also introduces spatial constraints. For each texture feature point, its corresponding structural feature point is found in three-dimensional space, and then the degree of difference between the two in the feature space is calculated using the following formula: in: This represents the weighted distance between the i-th texture feature and the j-th structural feature. It is the weight coefficient of the k-th feature dimension. It is the value of the i-th texture feature in the k-th dimension. It is the value of the j-th structural feature in the k-th dimension. The weight coefficients are dynamically adjusted according to the importance of the features, and the distance calculation results form a feature difference matrix.

[0028] Based on the feature difference matrix, a bidirectional weighted feature fusion is performed on the defect characterization feature vector and the structural deformation feature vector. The weighting process adopts an adaptive weight allocation mechanism, assigning higher fusion weights to feature pairs with small differences and reducing the weight contribution of feature pairs with large differences. The fusion operation is performed along the feature dimensions, calculating a weighted average for each feature dimension to generate a spatial-texture fusion feature vector that simultaneously contains texture and structural information. This vector retains the main discriminative information of the original features. The spatial-texture fusion feature vector and the stress anomaly feature vector are integrated temporally through a gated recurrent unit (GRU). The GRU network adopts a three-layer hidden layer structure, and the input at each time step includes the spatial-texture feature and stress anomaly feature at the current moment. The network controls the flow of information through update gates and reset gates, learns the temporal dependencies between multimodal features, and finally outputs a fused detection feature set containing rich spatiotemporal feature information. In the defect root cause weighting process, the fused detection feature set is input into the feature importance evaluation layer of the deep decision tree model. This layer adopts a feature importance calculation method based on Gini impurity. By statistically analyzing the amount of impurity reduction brought by each feature during the decision tree splitting process, the contribution of each feature dimension to defect classification is quantified. The calculation process traverses all splitting nodes of the decision tree, accumulates the number of times each feature is used for splitting and the total amount of impurity reduction brought by it, and finally obtains the importance score of each feature.

[0029] The system generates association rules between feature combinations and defect types for each detection region through a multi-path inference layer of a deep decision tree model. The multi-path inference layer consists of multiple decision trees, with each tree's path from the root node to a leaf node forming a decision rule. The system extracts complete paths from all decision trees, combining the splitting conditions on the path with the defect category corresponding to the final leaf node to form association rules. These rules reveal the correspondence between specific feature combinations and defect types. Dynamic weight allocation is performed on the fused detection features based on the association rule set. The weight allocation process first calculates the confidence and support of each association rule, then adjusts the weights of the corresponding features according to the rule quality. Features included by multiple high-confidence rules are assigned higher weights, while features used less frequently by rules are assigned lower weights. Finally, a defect association score set is generated, reflecting the probability of defects in each detection region. The entire implementation process is executed using a distributed computing framework, with feature fusion and weight allocation tasks distributed across multiple computing nodes for parallel processing. The system incorporates a feature quality monitoring mechanism to detect changes in feature value distribution in real time. When a significant drift in feature distribution is detected, a model recalibration process is automatically triggered to ensure the accuracy and reliability of the evaluation results. In terms of parameter tuning, the hidden layer dimension of the gated recurrent unit network is adaptively determined based on the length of the feature vector, typically set to 1.5 to 2 times the original feature dimension. The number and depth of decision trees are determined through cross-validation, achieving a balance between model complexity and generalization ability. The confidence threshold in weight allocation is dynamically adjusted according to actual application needs, controlling the strictness of association rule selection. The data processing pipeline adopts a pipeline architecture, with each processing step's output undergoing quality checks and scoring; only qualified intermediate results proceed to the next processing stage. The system maintains detailed processing logs, recording the processing path and transformation history of each feature vector, providing a complete data traceability link for subsequent defect tracing. The anomaly handling mechanism includes feature value anomaly detection and data consistency verification. When a feature value is found to be outside a reasonable range, the system automatically replaces it with adjacent value interpolation or model-based predicted values. For severe data anomalies, a data re-acquisition process is triggered to ensure the quality and reliability of the input data. The entire implementation process emphasizes quality control of data processing and robust handling of anomalies.

[0030] The implementation process of cross-modal correlation fusion and defect root cause weighting is illustrated using a specific sheet metal shell inspection case: A batch of sheet metal shells, numbered B-2089, simultaneously acquired 3D laser point cloud data, surface image data, and stress time-series data on the inspection line. System timestamp records show that the 3D scan data acquisition interval was 2 seconds, the image data acquisition frequency was 1 frame / second, while the stress data was continuously acquired at 1000Hz. This multi-rate sampling requires time alignment processing first. When implementing time-dimensional interpolation, the system resamples the structural deformation feature vector based on the time axis of the stress data. Since a feature vector is generated every 2 seconds for the 3D scan data, while a stress feature vector is generated every 0.1 seconds, a cubic spline interpolation algorithm is used to insert intermediate values ​​between the scan data points, ensuring that the time resolution of the structural deformation feature vector is consistent with that of the stress feature. The interpolated vector contains a timestamp that completely corresponds to the stress data. When calculating the correlation between defect characterization feature vectors and structural deformation feature vectors, the system establishes a spatial correspondence: the image pixel coordinates are transformed to a three-dimensional coordinate system through calibration to find the spatial position corresponding to each texture feature point. Taking the right corner area of ​​the workpiece as an example, the scratch feature point (X134, Y87) detected in the image corresponds to the three-dimensional mesh unit G-210. The multi-dimensional difference between the texture features and structural features at this location is calculated, including indicators such as gradient direction distribution, curvature change pattern, and thickness uniformity. When performing feature weighted fusion based on the difference matrix, the system assigns adaptive weights to each feature pair: for the scratch area, the weight of the texture gradient feature is set to 0.7, and the weight of the structural curvature feature is 0.3; for the stress concentration area, the weight of the structural thickness feature is increased to 0.6, and the weight of the texture feature is reduced to 0.4; the spatial-texture fusion vector generated after weighted fusion retains both surface defect and structural deformation information, and the vector dimension is the sum of the original feature dimensions. A gated recurrent unit network receives the fused feature sequence. The network structure contains 128 hidden units, with a time step of 10 consecutive time points. The network learns the temporal dependencies between multimodal features, such as discovering that stress anomalies often precede structural deformation and that surface defects are strongly correlated with stress abrupt changes. It outputs a fused detection feature set containing spatiotemporal features, which captures the dynamic evolution of the defect generation process. A deep decision tree model receives the fused features and evaluates their importance. The model contains 100 decision trees, each with a maximum depth of 15 layers. Feature importance calculations show that in corner region detection, the Gini importance score for stress abrupt changes is 0.32, for structural curvature features it is 0.28, for texture gradient features it is 0.25, and for thickness uniformity features it is 0.15. These scores reflect the contribution of each feature to defect detection.When generating association rules in the multi-path inference layer, multiple important paths are extracted from the decision tree: Rule R87 indicates that "when the peak stress is >80MPa and the curvature change is >0.15mm, the defect probability reaches 92%"; Rule R92 shows that "when the texture gradient anomaly is accompanied by a thickness deviation >0.2mm, the defect probability is 87%"; Rule R105 indicates that "when structural deformation occurs within 2 seconds after a stress mutation, the defect probability is 95%". These rules form a knowledge base for defect diagnosis. During the dynamic weight allocation process, the system adjusts the feature weights based on the rule confidence and support: for features that are included by multiple high-confidence rules (such as stress mutation and curvature change), the weight coefficient is increased to 0.9; for features that are used less frequently by rules (such as color uniformity), the weight is reduced to 0.3; in the final generated defect association score set, the corner area receives 0.94 points (out of 1.0), while the scores of other areas are all below 0.3. The entire process is completed on a distributed computing cluster. The feature fusion stage takes approximately 50 milliseconds, and the weight allocation stage takes 30 milliseconds, meeting the real-time detection requirements of the production line. The system regularly updates the decision tree model, and new detection data continuously enriches the association rule base, improving the accuracy of defect diagnosis. Ultimately, the workpiece was accurately identified as having a complex defect at the right corner, and proceeded to the subsequent source tracing analysis process.

[0031] Example 4: In the specific implementation of the joint defect tracing process, firstly, score items exceeding a dynamic threshold are screened from the defect correlation score set. The dynamic threshold is determined based on the statistical distribution of historical detection data, using a percentile-based calculation method. After sorting the scores in descending order, the top 5% of scores are taken as the threshold boundary. The system traverses the score data of all detection areas and extracts score items above the threshold to form a high-confidence defect feature subset. This subset contains metadata such as score values, area identifiers, and timestamps. The high-confidence defect feature subset undergoes reverse feature mapping processing. This step restores the abstract feature vectors to the original data space by querying the coordinate mapping relationship from the feature extraction stage. For structural deformation features, the specific area coordinates in the 3D laser scanning data are located based on the grid cell number. For texture features, the corresponding position in the surface defect image is found through the hierarchical index of the image pyramid and pixel coordinate mapping. For stress features, the specific time interval of the stress distribution time series data is mapped based on the time window number.

[0032] When locating abnormal deformation areas in 3D laser scanning data, the system highlights the mesh cells corresponding to the abnormal values ​​in the feature vector and marks the specific coordinate range of the deformation area in the 3D model. Taking the right corner of a sheet metal shell as an example, the system detects that the curvature parameters of the mesh cells (numbered G-207 to G-215) in this area significantly exceed the normal range, and the thickness distribution shows an uneven pattern. The system automatically marks these abnormal meshes with red areas in the 3D point cloud model. When locating texture abnormal areas in surface defect image data, the system finds the specific level and pixel range in the corresponding image pyramid based on the spatial index of the feature vector. Continuing with the right corner as an example, the system detects an abnormal gradient direction in the pixel area (X:120-135, Y:80-95) in the second-level image slice. This area corresponds to a surface scratch defect on the actual workpiece. The system marks the abnormal area on the original image and records its morphological characteristics. When locating the stress abrupt change interval in the stress distribution time series data, the specific time segment in the stress sensor data is found based on the timestamp information of the feature vector. The detection found that the workpiece had a stress peak during the processing timestamp T+123s to T+127s, with the maximum stress value reaching 2.3 times the normal range. The system recorded the stress curve characteristics and abrupt change pattern of this time interval.

[0033] A pre-trained causal inference network establishes the process defect propagation path between abnormal deformation regions, texture anomaly regions, and stress mutation intervals. The causal inference network employs a graph neural network architecture, where nodes represent detected abnormal regions and edges represent possible causal relationships. Network input includes the spatial coordinates, timestamps, and feature values ​​of the abnormal regions. Multi-layer graph convolution operations are used to learn the association patterns between nodes. In the causal inference network processing, the grid coordinates of the abnormal deformation regions, the pixel coordinates of the texture anomaly regions, and the timestamps of the stress mutation intervals are input to a space-time alignment layer. This layer uses a unified coordinate system transformation method to map all abnormal elements to the same spatiotemporal reference system. Grid coordinates are calibrated to the world coordinate system, pixel coordinates are transformed to the same coordinate system according to camera parameters, and timestamps are unified to the absolute time axis of the production line. The graph attention mechanism of the causal inference network calculates the spatial proximity and temporal continuity parameters between abnormal regions. Spatial proximity is calculated based on Euclidean distance, and temporal continuity is analyzed based on time difference to determine the chronological order of abnormal events. The attention mechanism assigns different importance weights to each abnormal region, highlighting the dominant role of key nodes in the causal chain. Based on spatial proximity and temporal continuity parameters, a multi-level causal link from material defects to processing deformation is generated. The system uses a path search algorithm to find possible causal relationship chains in the anomaly region map. Taking a specific detected anomaly as an example, the system constructs a causal hypothesis from stress mutation to surface scratches to structural deformation, and verifies the rationality of the causal chain through probability calculation.

[0034] Table 1: Correlation Analysis of Abnormal Regions Exception types Spatial coordinates Timestamp Eigenvalues Proximity score Temporal continuity Stress mutation Equipment station S4 T+123s Peak 2.3x - - Surface scratches (X134,Y87) T+124s Gradient anomalies 0.87 0.92 Structural deformation Grid G-210 T+125s Curvature Exceeds Limit 0.93 0.89 Referring to Table 1, the system refines the causal chain through multiple rounds of iterative reasoning. The first round establishes preliminary causal hypothesis, and subsequent rounds verify and revise this hypothesis by introducing process knowledge and physical constraints. The final output multi-level causal chain contains a complete anomaly propagation path, indicating the possible sources and transmission processes of defects. Real-time visual monitoring is employed during implementation. All detected anomaly areas and established causal chains are dynamically displayed on the monitoring interface, allowing operators to intuitively view defect distribution and propagation paths. The system supports interactive analysis, allowing users to manually adjust and verify the automatically generated causal chains, combining professional experience to refine the defect tracing results. Regarding data storage, all tracing results, including anomaly area location data, causal chain diagrams, and reasoning process records, are saved to the quality database, providing data support for subsequent quality analysis and process improvement. The system periodically performs statistical analysis on the tracing results to identify common defect patterns and recurring process problems, promoting continuous optimization of the production process. The entire implementation process emphasizes practicality and operability. The parameter settings of the causal reasoning network have been verified through a large number of experiments to ensure that reliable traceability results can be generated in different production scenarios. The system is equipped with multiple verification mechanisms, including causal logic consistency checks and physical law conformity verification, to ensure that the output defect traceability conclusions are consistent with data patterns and meet engineering realities.

[0035] Example 5: In the process of determining the processing stage and material anomaly type of the defect source, analyzing the characteristic attributes of each node in the multi-level causal link is a key step. The system extracts the characteristic data of each node from the directed graph output by the causal inference network, including spatial coordinates, timestamps, physical parameters, and anomaly type labels. For the defect case at the right corner of the sheet metal shell, the system analyzes three main nodes: Node A (stress mutation point) is located in the S4 area of ​​the equipment platform, timestamp T+123s, and the peak stress value exceeds the standard range; Node B (surface scratch) is located at workpiece coordinates (X134, Y87), timestamp T+124s, and the gradient feature anomaly degree reaches 0.87; Node C (structural deformation) is located in mesh cell G-210, timestamp T+125s, and the curvature deviation exceeds the threshold of 0.15mm. When matching abnormal equipment parameter records in the preset processing technology knowledge base, the system will perform similarity matching between the parsed node features and historical events in the knowledge base. The processing technology knowledge base is stored in a graph database structure and contains data such as equipment operating parameters, mold usage records, and process adjustment logs. The system can quickly retrieve historical records that match the current abnormal features through multi-dimensional indexes. The retrieval conditions include equipment number, time range, parameter deviation type, etc.

[0036] Taking a corner defect as an example, the system retrieved three relevant records for equipment station S4 in the knowledge base near the detection time point: record R1023 showed that the pressure parameter of stamping die M7-42 fluctuated abnormally at T+120s; record R1045 indicated that the bending process parameter exceeded the set range at T+122s; and record R1067 recorded that material batch B-2089 had uneven hardness in this area. The system calculated the matching degree between each record and the current abnormal feature and selected the record with the highest matching degree as the candidate cause. Based on the abnormal equipment parameter records, the system identified the stamping die number and bending process parameter deviation value that caused the material stress concentration. The identification process adopted the decision tree classification method, which stratified the abnormal equipment parameter records according to equipment type, process section, and parameter category. For the stamping die number, the system analyzed the press work log and die usage record to determine that the die number running at the abnormal time point was M7-42. This die had accumulated 15,200 processing cycles and was close to its design life. The deviation values ​​of bending process parameters are calculated by comparing the actual processing parameters with the standard process requirements. The system reads real-time data of the bending process from the PLC controller, including parameters such as pressing speed, holding time, and pressure value. The test found that the pressing speed of the workpiece was set to 12mm / s, while the standard process requirement is 10±0.5mm / s. The holding time of 3.5s exceeded the standard range by 3.0±0.2s, and the pressure value of 85kN was also 6.25% higher than the standard value of 80kN. By combining the stamping die number with the bending process parameter deviation value, a complete process responsibility chain for sheet metal shell defects is generated. The responsibility chain generation adopts the fault tree analysis method, with the final defect as the top event and the parameter abnormality of each processing link as the basic event, and constructs a logical relationship diagram. The responsibility chain automatically generated by the system shows that: excessive wear of die M7-42 leads to uneven distribution of stamping pressure, which in turn causes local stress concentration in the material. In addition, improper bending process parameter settings, excessive pressing speed and excessive holding time work together to finally produce multiple defects such as stress mutation, surface scratches and structural deformation at the right corner of the sheet metal shell.

[0037] The chain of responsibility verification was conducted by reproducing the processing procedure. The system retrieved inspection data from other workpieces processed under the same process parameters and found that 23% of products processed using the M7-42 mold exhibited similar defects, while the defect rate dropped to 2% after replacing the mold. Experiments adjusting bending process parameters showed that when the pressing speed was controlled within 10.5 mm / s and the holding time did not exceed 3.2 s, stress concentration at corners was significantly improved. The generated chain of responsibility report contains a complete chain of evidence and data support. The report details the inspection data for each abnormal node, the matching equipment parameter records, the calculated deviation value, and the deduced fault propagation path. The report uses a structured format, including defect description, inspection data, cause analysis, responsible links, and improvement suggestions, providing clear direction for production quality improvement. A continuous learning mechanism was established during system implementation. Each identified defect case and its chain of responsibility analysis results were fed back into the process knowledge base, enriching the case accumulation. The system regularly evaluates the accuracy of the chain of responsibility, continuously optimizing the causal reasoning model and matching algorithm through actual production verification and feedback on improvement effects, thereby improving the accuracy and practicality of defect tracing. The entire implementation process is tightly integrated with the production line management system. Once the defect responsibility chain is determined, the corresponding handling process is automatically triggered, including notifying the equipment department to replace the M7-42 mold, adjusting the bending process parameter settings, and conducting key inspections on work using the same batch of materials. The system records the handling process and final effect of each defect case, forming a complete quality closed-loop management system, and promoting continuous improvement and optimization of the production process.

[0038] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent detection of sheet metal casing of energy storage equipment, characterized in that, The method includes: The multimodal detection data set of the sheet metal shell is collected in real time from the production line. The multimodal detection data set includes three-dimensional laser scanning data, surface defect image data and stress distribution time series data. The three-dimensional laser scanning data is subjected to spatial topological feature parsing processing to generate structural deformation feature vectors; the surface defect image data is subjected to multi-scale texture feature extraction processing to generate defect characterization feature vectors; and the stress distribution time series data is subjected to dynamic anomaly pattern recognition processing to generate stress anomaly feature vectors. The structural deformation feature vector, defect characterization feature vector, and stress anomaly feature vector are subjected to cross-modal correlation and fusion processing to generate a fused detection feature set. Based on the fused detection feature set, a pre-trained deep decision tree model is invoked to perform defect root cause weight allocation processing, generating a defect correlation score set corresponding to each detection area of ​​the sheet metal shell. Based on the defect correlation score set, the three-dimensional laser scanning data, surface defect image data, and stress distribution time series data are jointly processed to trace the source of the defects, thereby determining the processing stage and material anomaly type from which the defects originate.

2. The intelligent detection method for sheet metal casing of energy storage equipment according to claim 1, characterized in that, The step of performing spatial topological feature analysis on the three-dimensional laser scanning data to generate structural deformation feature vectors includes: The three-dimensional laser scanning data is divided into grids, and the spatial curvature parameters and thickness distribution parameters of each grid cell are extracted. The spatial curvature parameters of adjacent grid cells are encoded using a pre-trained graph neural network to generate local deformation feature vectors. The thickness distribution parameters of each grid cell are concatenated with the corresponding local deformation feature vectors to generate the structural deformation feature vector.

3. The intelligent detection method for sheet metal casing of energy storage equipment according to claim 1, characterized in that, The step of performing multi-scale texture feature extraction processing on the surface defect image data to generate a defect characterization feature vector includes: The surface defect image data is subjected to Gaussian pyramid layering to obtain image slices at different resolution levels; The image slices at each resolution level are processed by extracting directional gradient histogram features through a pre-trained convolutional neural network to generate a multi-scale texture feature matrix. The multi-scale texture feature matrix is ​​subjected to inter-level feature alignment and dimensionality reduction to generate the defect characterization feature vector.

4. The intelligent detection method for sheet metal casing of energy storage equipment according to claim 1, characterized in that, The step of performing dynamic anomaly pattern recognition processing on the stress distribution time-series data to generate a stress anomaly feature vector includes: The stress distribution time series data is processed by sliding window segmentation to extract the mean stress, range and fluctuation frequency parameters within each time window; By using a pre-trained temporal encoder to dynamically compare the stress parameters of adjacent time windows, abnormal fluctuation ranges can be identified. The stress parameters in the abnormal fluctuation range are compared and encoded with those in the normal fluctuation range to generate the stress anomaly feature vector.

5. The intelligent detection method for sheet metal casing of energy storage equipment according to claim 1, characterized in that, The step of performing cross-modal correlation and fusion processing on the structural deformation feature vector, defect characterization feature vector, and stress anomaly feature vector to generate a fused detection feature set includes: The structural deformation feature vector is interpolated in the time dimension to align it with the time step of the stress anomaly feature vector. Calculate the Euclidean distance matrix between each texture feature in the defect characterization feature vector and the corresponding spatial feature in the structural deformation feature vector; Based on the Euclidean distance matrix, a two-way feature weighted fusion process is performed on the defect characterization feature vector and the structural deformation feature vector to generate a spatial-texture fusion feature vector. The spatial-texture fusion feature vector and the stress anomaly feature vector are integrated into a temporal feature through a gated loop unit to generate the fusion detection feature set.

6. The intelligent detection method for sheet metal casing of energy storage equipment according to claim 1, characterized in that, The process involves calling a pre-trained deep decision tree model based on the fused detection feature set to perform defect root cause weight allocation, generating a defect correlation score set corresponding to each detection area of ​​the sheet metal shell, including: The fused detection feature set is input into the feature importance evaluation layer of the deep decision tree model to calculate the contribution of each feature dimension to defect classification. The deep decision tree model generates a set of association rules between feature combinations and defect types for each detection region through its multi-path inference layer. The fusion detection features are dynamically weighted according to the set of association rules to generate the defect association score set.

7. The intelligent detection method for sheet metal casing of energy storage equipment according to claim 6, characterized in that, The joint defect tracing process based on the defect correlation score set for three-dimensional laser scanning data, surface defect image data, and stress distribution time-series data includes: Filter the score items in the defect correlation score set that exceed the dynamic threshold to generate a high-confidence defect feature subset; The high-confidence defect feature subset is subjected to inverse feature mapping processing to locate abnormal deformation regions in the three-dimensional laser scanning data, abnormal texture regions in the surface defect image data, and stress abrupt change intervals in the stress distribution time series data. The process defect transmission path between the abnormal deformation region, the texture abnormal region and the stress mutation interval is established by a pre-trained causal reasoning network.

8. The intelligent detection method for sheet metal casing of energy storage equipment according to claim 7, characterized in that, The establishment of the process defect transmission path between the abnormal deformation region, the texture abnormal region, and the stress mutation interval through a pre-trained causal reasoning network includes: The grid coordinates of the abnormal deformation region, the pixel coordinates of the texture abnormal region, and the timestamps of the stress mutation intervals are input into the space-time alignment layer of the causal inference network. The spatial proximity and temporal continuity parameters between each anomalous region are calculated using the graph attention mechanism of the causal inference network. Based on the spatial proximity and temporal continuity parameters, a multi-level causal link from material defects to processing deformation is generated.

9. The intelligent detection method for sheet metal casing of energy storage equipment according to claim 8, characterized in that, The processing steps and material anomaly types used to determine the source of defects include: The characteristic attributes of each node in the multi-level causal link are analyzed and matched with abnormal equipment parameter records in the preset processing technology knowledge base; Based on the abnormal equipment parameter records, identify the stamping die number and bending process parameter deviation value that caused the material stress concentration; A complete process responsibility chain for sheet metal shell defects is generated by combining the stamping die number with the bending process parameter deviation value.

10. An intelligent inspection system for sheet metal casing of an energy storage device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent detection method for sheet metal casing of energy storage equipment as described in any one of claims 1 to 9.

Images