Lithium battery electrode structure defect detection method based on Fourier transform

Abstract

The invention discloses a lithium battery electrode structure defect detection method based on Fourier transform, and the method comprises the following steps: S1, collecting multi-modal image data, and carrying out the preprocessing; s2, executing two-dimensional Fourier transform to generate frequency domain feature data; s3, complementing modal data, and optimizing modal consistency; s4, constructing a collaborative feature generation model, and fusing modals to form enhanced features; s5, using deformable convolution optimization features to adjust a receptive field and a fusion weight; s6, inputting feature data for modeling, and generating a defect trend and risk assessment; s7, analyzing a defect and process relationship in combination with causal reasoning, and generating optimization suggestions; and S8, applying the optimization parameters, and outputting a defect and dynamic trend detection result. Through multi-modal feature optimization, dynamic modeling and causal analysis, the precision and efficiency of lithium battery electrode defect detection are remarkably improved, meanwhile, the production process is optimized, the production cost is reduced, and the product quality is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery manufacturing and detection, and particularly to a method for detecting defects in the electrode structure of a lithium battery based on Fourier transform. Background Art

[0002] In recent years, with the wide application of lithium batteries in new energy vehicles, energy storage devices, consumer electronics and other fields, the requirements for the performance and safety of lithium batteries have been increasing day by day. However, the stability of lithium battery performance and service life are largely affected by defects in the electrode structure. During the manufacturing process of lithium battery electrodes, structural defects such as cracks, particle accumulation, and uneven coating are likely to occur, and these defects may lead to a decline in electrochemical performance, local overheating, and even safety accidents. Therefore, quickly and accurately detecting the structural defects of lithium battery electrodes has become one of the important links in improving the quality of lithium batteries.

[0003] In the prior art, the detection of lithium battery electrode defects mainly relies on analysis methods based on single-modal data such as visible light images, infrared thermal images, and X-ray images. These methods extract obvious defect features on the electrode surface through image processing techniques, such as uneven coating, particle accumulation, etc., and have achieved certain results. However, for hidden defects such as microcracks, internal stress distribution of materials, and changes in local microstructures, the expression ability of single-modal data is limited, and it is difficult to conduct a comprehensive and in-depth detection of complex defects. In addition, single-modal data has a high dependence on environmental conditions and hardware acquisition devices. When there are missing or poor-quality modal data during the data acquisition process, the detection performance will decline significantly.

[0004] To address the limitations of single-modal methods, methods for detecting electrode defects based on multi-modal data fusion have been proposed in recent years. These methods utilize multi-modal data such as visible light, infrared thermal images, and X-rays, and comprehensively characterize the electrode structure features by fusing the characteristics of different modal data, thereby improving the coverage rate of defect detection. However, in practical applications, multi-modal data fusion technology still faces many challenges. On the one hand, there are significant differences in the characteristics of different modal data, and there may be conflicts and redundant information between modalities, and direct fusion is likely to lead to a decline in the reliability of the detection results; on the other hand, when there are missing or poor-quality multi-modal data, existing fusion methods usually cannot effectively compensate for the missing modal information, thereby affecting the detection performance. In addition, most existing multi-modal fusion methods are static processing, and it is difficult to capture the dynamic characteristics of defects evolving over time, and they cannot provide the ability to predict the trend of defect expansion and risk assessment.

[0005] In the research of defect propagation trends, traditional static image processing techniques are difficult to capture dynamic evolution characteristics, while time series modeling methods can only describe the temporal variation laws of single-modal data, ignoring the correlation between multi-modal data. This results in significant limitations of existing dynamic defect detection methods when dealing with complex defect propagation scenarios. In addition, the propagation of defects is usually directly related to production process parameters, and existing technologies lack effective causal analysis tools and cannot accurately reveal the correlation between defect characteristics and process parameters. As a result, the detection results are often difficult to directly guide process optimization, limiting the improvement of the yield rate in the electrode manufacturing process.

[0006] Another shortcoming of the existing technology regarding the above problems is the lack of the ability for closed-loop optimization. Traditional electrode defect detection systems usually can only provide static detection results and cannot achieve real-time optimization and feedback adjustment of process parameters. This static detection method cannot adapt to complex and changeable manufacturing environments and is difficult to meet the requirements of high-precision and high-efficiency production. In addition, in actual industrial production, complex electrode surface textures and diverse defect types further increase the difficulty of detection. The existing models are still insufficient in robustness to complex backgrounds and are easily interfered by background noise and irrelevant features, leading to false detection and missed detection problems.

[0007] Therefore, how to provide a lithium battery electrode structure defect detection method based on Fourier transform is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to propose a lithium battery electrode structure defect detection method based on Fourier transform. The present invention comprehensively applies multi-modal feature dynamic reconstruction, efficient deformable convolutional network, spatio-temporal graph neural network, and cross-level multi-scale causal inference algorithm to solve the technical bottlenecks of the existing technology in limited single-modal detection ability, modal conflict and redundancy in the multi-modal data fusion process, insufficient dynamic defect propagation trend prediction ability, and lack of causal analysis of defects and production process parameters. The present invention effectively improves the detection accuracy of obvious and hidden defects by dynamically optimizing multi-modal feature fusion; combines spatio-temporal graph neural network to model the temporal evolution characteristics of defects and realize dynamic risk assessment; reveals the correlation between defects and process parameters through cross-level causal inference, deduces multi-level causal paths, and finally generates scientific process optimization suggestions to realize the closed-loop feedback of detection and production optimization, thereby significantly improving the yield rate, safety, and quality of lithium battery electrode production.

[0009] A lithium battery electrode structure defect detection method based on Fourier transform according to an embodiment of the present invention includes the following steps:

[0010] S1. Collect multi-modal image data on the surface of the lithium battery electrode, including visible light images, infrared thermal images, and X-ray images, and preprocess the collected multi-modal data. The preprocessing includes denoising, contrast enhancement, and geometric calibration;

[0011] S2. Perform two-dimensional Fourier transform on the preprocessed multi-modal data to generate frequency domain feature data, including low-frequency components and high-frequency components;

[0012] S3. For the case of missing modalities in the frequency domain feature data, use the existing modality data as conditional input, generate the missing modality data through a conditional generative adversarial network, and optimize through cross-modal consistency constraints to output complete multi-modal feature data;

[0013] S4. Based on the complete multi-modal feature data, construct a collaborative feature generation model to dynamically generate new feature modalities, and fuse the generated feature modalities with the original modal features through a cross-modal attention mechanism to form enhanced multi-modal fusion feature data;

[0014] S5. Use an efficient deformable convolutional network to dynamically optimize the enhanced multi-modal fusion feature data, adaptively adjust the receptive field according to the spatial deformation characteristics of different modalities, dynamically adjust the modal fusion weights, and output optimized multi-modal fusion feature data;

[0015] S6. Input the optimized multi-modal fusion feature data into a spatio-temporal graph neural network, model the dynamic defect expansion trend by combining time series features, and generate time evolution characteristic data of defect expansion and dynamic risk assessment results;

[0016] S7. Based on the dynamic risk assessment results and time evolution characteristic data, analyze the causal relationship between defect features and production process parameters by combining a cross-level multi-scale causal inference algorithm, deduce multi-level causal paths, and generate suggestions for optimizing production process parameters;

[0017] S8. Apply the suggestions for optimizing production process parameters to the production process, and output the final detection results including defect distribution, defect type, dynamic expansion trend, and optimization suggestions.

[0018] Optionally, the S3 specifically includes:

[0019] S31. Analyze the preprocessed multi-modal data to identify the situation of missing modalities;

[0020] S32. Input the existing modality data into a conditional generative adversarial network, generate the missing modality data through the generator network, and combine the discriminator network to discriminate the authenticity of the generated data, optimize the adversarial training of the generator network and the discriminator network, and output the generated modality data;

[0021] S33. Optimize the cross-modal consistency constraint for the generated modal data:

[0022]

[0023] Among them, is the objective of the generator, is the objective of the discriminator, V(D, G) is the core objective function of the generative adversarial network, G is the generator, D is the discriminator, E represents the expected value, D r is the real data, P data is the real data distribution, log is the logarithmic function, D c is the generated modal data, P G is the forged data distribution; S34. Integrate the optimized generated modal data with the existing modal data to construct a complete multi-modal feature data set;

[0024] S35. Verify the complete multi-modal feature data and evaluate the effectiveness of the generated modality through multi-modal similarity measurement:

[0025]

[0026] Among them, S represents the similarity measurement, D e is the existing modal data.

[0027] Optionally, the specific steps of S4 include:

[0028] S41. Standardize the complete multi-modal feature data, convert each modal feature data into a unified feature space representation, and generate a standardized feature set;

[0029] S42. Based on the standardized feature set, construct a collaborative feature generation model and extract deep features through a multi-layer convolutional neural network:

[0030]

[0031] Among them, H m is the deep feature extracted for modality m, ReLU represents the activation function, W m and b m are the convolutional kernel weights and biases respectively, is the standardized feature set;

[0032] S43. Introduce an adaptive weight allocation mechanism in the collaborative feature generation model, dynamically adjust the modal feature weights according to the contribution degree of each modal feature to the target task, and generate a weighted modal feature set;

[0033] S44. Use the cross-modal attention mechanism to perform feature fusion on the weighted modal feature set, establish multi-level attention relationships between modalities, and generate preliminary fusion features:

[0034]

[0035] where a ij represents the attention weight between modality i and modality j, exp is the exponential function, sim is the similarity function, m and k represent different modalities, represents the weighted feature representation of modality i, represents the weighted feature representation of modality j, represents the weighted feature representation of modality k;

[0036] S45. Input the preliminary fusion features into the feature enhancement module and perform feature mapping through a multi-layer perceptron;

[0037] S46. Perform feature verification on the enhanced fusion feature data, and evaluate the effectiveness of the fusion features by calculating the modality consistency and task adaptability metrics.

[0038] Optionally, the specific steps of S5 are as follows:

[0039] S51. Input the enhanced fusion feature data into the efficient deformable convolutional network, initialize the offset parameters of the convolutional kernel, and assign initial weights to each modal feature;

[0040] S52. Dynamically adjust the receptive field range and position of the convolutional kernel by analyzing the spatial deformation characteristics of the modal features:

[0041]

[0042] where F(p) represents the optimized feature value of pixel point p, K is the convolutional kernel size, Δp k is the offset, represents the enhanced fusion feature data, w k is the weight of modality K;

[0043] S53. Calculate the importance of each modal feature, dynamically adjust the modal fusion weights according to the contribution of the modal features to the overall task, so that important modalities have higher priorities in the fusion process:

[0044]

[0045] where w m is the importance weight of each modality, M is the number of modalities, m represents the modality, β k represents the feature importance score of modality k, β m represents the feature importance score of modality m, and exp is the exponential function;

[0046] S54. Enhance the optimized modal features and the initial fusion features through residual connection to preserve the integrity of the initial features and generate optimized multi-modal fusion feature data;

[0047]

[0048] Among them, represents the feature data after residual optimization, is the initial fusion feature;

[0049] S55. Extract global information from the optimized multi-modal fusion feature data to capture the correlation between features;

[0050] S56. Verify the effectiveness of the optimized multi-modal fusion feature data, and evaluate it using similarity metrics and the loss function of the defect detection task.

[0051] Optionally, the specific steps of S6 are as follows:

[0052] S61. Map the optimized multi-modal fusion feature data into a spatio-temporal graph structure, define the nodes as defect feature points, the edges as the spatial or temporal correlation relationships between feature points, and assign an initial feature vector to each node. At the same time, establish an adjacency matrix to describe the connection relationships between nodes;

[0053] S62. Input the time-varying data of the defect feature points into the time module of the spatio-temporal graph neural network, extract time series features through a temporal convolutional network to capture the evolution law of the defect over time, and generate a dynamic time feature representation of the nodes;

[0054] S63. Use a graph convolutional network to model the spatial relationships in the graph structure, extract the spatial correlation characteristics between defect feature points, generate a spatial feature representation of the nodes, and update the weights of the edges between nodes at the same time;

[0055] S64. Fuse the time feature representation and the spatial feature representation through a spatio-temporal attention mechanism to generate a comprehensive feature representation containing temporal and spatial correlations, and assign a spatio-temporally fused feature vector to each node:

[0056]

[0057] Among them, is the comprehensive feature representation of node v, is the time feature representation of node v, is the spatial feature representation of node v, and α is the attention weight;

[0058] S65. Based on the feature representation after spatio-temporal fusion, perform recursive learning on the graph structure, predict the expansion trend of defects in the future time, generate the time evolution path of defect feature points, and describe the dynamic change pattern of defects;

[0059] S66. According to the time evolution path of defect expansion, evaluate the risk level of the defect, generate a dynamic risk assessment result by comprehensively considering the defect location and expansion speed, and output an assessment report including the defect expansion trend and risk level.

[0060] Optionally, the specific steps of S7 are as follows:

[0061] S71. Based on the dynamic risk assessment result and time evolution characteristic data, construct a causal graph model, define the nodes as defect features and production process parameters, the edges represent the causal relationships between the nodes, and initialize the node feature values and causal association matrix;

[0062] S72. Use the causal graph model to layer the nodes, dividing them into a defect feature layer, a local process parameter layer, and a global process parameter layer;

[0063] S73. Combine the cross-level multi-scale causal reasoning algorithm to calculate the direct causal relationship and the indirect causal relationship respectively. The direct relationship is calculated based on the causal intensity function between the nodes, and the indirect relationship is generated through recursive path derivation, obtaining a multi-scale causal weight matrix:

[0064]

[0065] Among them, represents the direct causal impact of node v i on v j , A ij is the hierarchical adjacency matrix, MI represents the causal association intensity between two nodes, represents the indirect causal impact of node v i on v j , P(i,j) represents the set of paths from node i to node j, |z| represents the length of path z, L is the index of the edge in the path, and A zL represents the weight or connection value of the L-th edge in path z;

[0066] S74. According to the results of multi-scale causal reasoning, generate multi-level causal paths, including the path from defect features to local process parameters, the path from local process parameters to global process parameters, and the comprehensive path, and calculate the causal intensity and path score of each path:

[0067]

[0068] Among them, P ij represents node v ito node v j The total causal strength to ij node v, S(P ij ) represents the path score, and len(P

[0069] S75. Analyze the sensitivity of each process parameter to defect characteristics based on the scoring results of causal paths, identify key process parameters, determine the adjustment priority by combining sensitivity analysis and causal path scoring, and generate a set of key parameters;

[0070] S76. Combine the adjustment priority of key process parameters and the causal path results to generate production process optimization suggestions, including the parameter adjustment range, optimization direction, and their potential impacts, and output an optimization suggestion report.

[0071] The beneficial effects of the present invention are as follows:

[0072] Through the method for detecting lithium battery electrode structure defects based on Fourier transform proposed by the present invention, the deficiencies of the prior art in defect detection, data fusion, dynamic analysis, and process optimization are significantly overcome, achieving the beneficial effects of comprehensively improving the defect detection ability and production process optimization efficiency. The present invention solves the problem of missing or degraded modal data caused by hardware limitations or environmental interference during the multi-modal data acquisition process through the multi-modal feature dynamic reconstruction technology. The introduction of the conditional generative adversarial network (CGAN) not only effectively compensates for the missing modal data but also ensures the consistency between the compensated data and the actual modal characteristics. The application of the cross-modal attention mechanism further optimizes the multi-modal data fusion process, effectively suppressing the conflicts and redundant information between modalities, thereby realizing the comprehensive expression of defect characteristics. This fusion method improves the detection ability for both obvious defects (such as uneven coating and particle accumulation) and hidden defects (such as microcracks and internal stress distribution) of the electrode, significantly improving the detection accuracy and reliability.

[0073] The use of the efficient deformable convolutional network (DCN) is a major innovative highlight of the present invention. By dynamically adjusting the size and shape of the receptive field, the network can adaptively capture the spatial feature changes of different modalities and regions, especially showing excellent robustness in the detection of complex textures and diverse defect types. In addition, the introduction of residual optimization not only retains the original feature information but also enhances the detection ability for small defects that are difficult to capture, enabling the detection system to maintain stable performance in complex production environments.

[0074] The present invention realizes the dynamic modeling of defect propagation through a spatio-temporal graph neural network (ST-GNN), filling the technical gap that existing static detection methods cannot capture the dynamic defect propagation trend. By jointly modeling time series features and spatial correlation characteristics, it accurately predicts the dynamic propagation path and potential risks of defects, providing intuitive risk assessment results. This ability not only helps to identify potential hazards in advance but also provides a scientific basis for formulating more efficient quality control strategies. At the same time, the innovative application of the cross-level multi-scale causal inference algorithm clarifies the influence path of process parameters on defect formation and propagation by hierarchically analyzing the causal relationship between defect features and production process parameters. This function realizes the logical closed-loop from defect detection to process optimization, effectively transforming the detection results into the guiding basis for production improvement.

[0075] In terms of process optimization, the present invention further generates scientific process parameter optimization suggestions through causal path analysis and dynamic risk assessment. By analyzing the sensitivity and adjustment priority of process parameters and combining the multi-level association of causal paths, it provides detailed suggestions on the parameter adjustment range and optimization direction, ultimately achieving the precise optimization of the production process. Different from traditional detection methods that can only provide static detection results, the present invention forms a dynamic feedback mechanism between detection and optimization, significantly improving production efficiency and product quality. Brief Description of the Drawings

[0076] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0077] Figure 1 is a flowchart of a method for detecting defects in a lithium battery electrode structure based on Fourier transform proposed by the present invention;

[0078] Figure 2 is a schematic structural diagram of a multi-modal feature dynamic reconstruction module of a method for detecting defects in a lithium battery electrode structure based on Fourier transform proposed by the present invention. Detailed Embodiments

[0079] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0080] Reference Figure 1-2 , a method for detecting defects in a lithium battery electrode structure based on Fourier transform, includes the following steps:

[0081] S1. Collect multi-modal image data on the surface of the lithium battery electrode, including visible light images, infrared thermal images, and X-ray images, and preprocess the collected multi-modal data. The preprocessing includes denoising, contrast enhancement, and geometric calibration;

[0082] S2. Perform two-dimensional Fourier transform on the preprocessed multi-modal data to generate frequency domain feature data, including low-frequency components and high-frequency components;

[0083] S3. For the modality missing situation in the frequency domain feature data, use the existing modality data as conditional input, generate the missing modality data through a conditional generative adversarial network, and optimize it through cross-modal consistency constraints to output complete multi-modal feature data;

[0084] S4. Based on the complete multi-modal feature data, construct a collaborative feature generation model to dynamically generate new feature modalities, and fuse the generated feature modalities with the original modal features through a cross-modal attention mechanism to form enhanced multi-modal fusion feature data;

[0085] S5. Use an efficient deformable convolutional network to dynamically optimize the enhanced multi-modal fusion feature data, adaptively adjust the receptive field according to the spatial deformation characteristics of different modalities, dynamically adjust the modal fusion weights, and output optimized multi-modal fusion feature data;

[0086] S6. Input the optimized multi-modal fusion feature data into a spatio-temporal graph neural network, model the dynamic defect expansion trend by combining time series features, and generate time evolution characteristic data of defect expansion and dynamic risk assessment results;

[0087] S7. Based on the dynamic risk assessment results and time evolution characteristic data, analyze the causal relationship between defect features and production process parameters by combining a cross-level multi-scale causal reasoning algorithm, deduce multi-level causal paths, and generate suggestions for optimizing production process parameters;

[0088] S8. Apply the suggestions for optimizing production process parameters to the production process and output the final detection results including defect distribution, defect type, dynamic expansion trend, and optimization suggestions.

[0089] In this embodiment, the specific content of S3 includes:

[0090] S31. Analyze the preprocessed multi-modal data to identify the modality missing situation;

[0091] S32. Input the existing modality data into the conditional generative adversarial network, generate the missing modality data through the generator network, and combine the discriminator network to perform authenticity discrimination on the generated data, optimize the adversarial training of the generator network and the discriminator network, and output the generated modality data;

[0092] S33. Optimize the cross-modal consistency constraint for the generated modal data:

[0093]

[0094] Among them, is the target of the generator, is the target of the discriminator, V(D, G) is the core objective function of the generative adversarial network, G is the generator, D is the discriminator, E represents the expected value, D r is the real data, P data is the real data distribution, log is the logarithmic function, D c is the generated modal data, P G is the forged data distribution; S34. Integrate the optimized generated modal data with the existing modal data to construct a complete multi-modal feature data set;

[0095] S35. Verify the complete multi-modal feature data and evaluate the effectiveness of the generated modality through multi-modal similarity measurement:

[0096]

[0097] Among them, S represents the similarity measurement, D e is the existing modal data.

[0098] In this embodiment, the specific content of S4 includes:

[0099] S41. Standardize the complete multi-modal feature data, convert each modal feature data into a unified feature space representation, and generate a standardized feature set;

[0100] S42. Based on the standardized feature set, construct a collaborative feature generation model and extract deep features through a multi-layer convolutional neural network:

[0101]

[0102] Among them, H m is the deep feature extracted for modality m, ReLU represents the activation function, W m and b m are the convolution kernel weights and biases respectively, is the standardized feature set;

[0103] S43. Introduce an adaptive weight allocation mechanism in the collaborative feature generation model, dynamically adjust the modal feature weights according to the contribution of each modal feature to the target task, and generate a weighted modal feature set;

[0104] S44. Use the cross-modal attention mechanism to perform feature fusion on the weighted modal feature set, establish multi-level attention relationships between modalities, and generate preliminary fusion features:

[0105]

[0106] Among them, a ij represents the attention weight between modality i and modality j, exp is the exponential function, sim is the similarity function, m and k represent different modalities, represents the weighted feature representation of modality i, represents the weighted feature representation of modality j, represents the weighted feature representation of modality k;

[0107] S45. Input the preliminary fusion features into the feature enhancement module and perform feature mapping through a multi-layer perceptron;

[0108] S46. Verify the features of the enhanced fusion feature data, and evaluate the effectiveness of the fusion features by calculating the modality consistency and task adaptability metrics.

[0109] In this embodiment, the S5 specifically includes:

[0110] S51. Input the enhanced fusion feature data into the efficient deformable convolutional network, initialize the offset parameters of the convolutional kernel, and assign initial weights to each modal feature;

[0111] S52. Dynamically adjust the receptive field range and position of the convolutional kernel by analyzing the spatial deformation characteristics of the modal features:

[0112]

[0113] Among them, F(p) represents the optimized feature value of pixel point p, K is the convolutional kernel size, Δp k is the offset, D f enh represents the enhanced fusion feature data, w k the weight of modality K;

[0114] S53. Calculate the importance of each modal feature, and dynamically adjust the modal fusion weights according to the contribution of the modal features to the overall task, so that important modalities get higher priorities in the fusion process:

[0115]

[0116] Among them, w m is the importance weight of each modality, M is the number of modalities, m represents the modality, β k represents the feature importance score of modality k, β mdenotes the feature importance score of modality m, and exp is the exponential function;

[0117] S54. Enhance the optimized modality features and the initial fusion features through residual connection, preserve the integrity of the initial features, and generate optimized multi-modal fusion feature data;

[0118]

[0119] Among them, denotes the feature data after residual optimization, is the initial fusion feature;

[0120] S55. Extract global information from the optimized multi-modal fusion feature data to capture the correlation between features;

[0121] S56. Validate the effectiveness of the optimized multi-modal fusion feature data, and evaluate it using a similarity metric and the loss function of the defect detection task.

[0122] In this embodiment, the specific steps of S6 are as follows:

[0123] S61. Map the optimized multi-modal fusion feature data into a spatio-temporal graph structure, define the nodes as defect feature points, the edges as the spatial or temporal correlation relationships between feature points, and assign an initial feature vector to each node. At the same time, establish an adjacency matrix to describe the connection relationships between nodes;

[0124] S62. Input the time variation data of the defect feature points into the time module of the spatio-temporal graph neural network, extract time series features through a temporal convolutional network, capture the evolution law of the defect over time, and generate a dynamic time feature representation of the nodes;

[0125] S63. Use a graph convolutional network to model the spatial relationships in the graph structure, extract the spatial correlation characteristics between defect feature points, generate a spatial feature representation of the nodes, and update the weights of the edges between nodes;

[0126] S64. Fuse the time feature representation and the spatial feature representation through a spatio-temporal attention mechanism to generate a comprehensive feature representation containing temporal and spatial correlations, and assign a spatio-temporally fused feature vector to each node:

[0127]

[0128] Among them, is the comprehensive feature representation of node v, is the time feature representation of node v, is the spatial feature representation of node v, and α is the attention weight;

[0129] S65. Based on the feature representation after spatio-temporal fusion, recursively learn the graph structure, predict the expansion trend of defects in the future time, generate the time evolution path of defect feature points, and describe the dynamic change pattern of defects.

[0130] S66. According to the time evolution path of defect expansion, evaluate the risk level of defects, generate a dynamic risk assessment result by comprehensively considering the defect location and expansion speed, and output an assessment report including the defect expansion trend and risk level.

[0131] In this embodiment, the specific content of S7 is as follows:

[0132] S71. Based on the dynamic risk assessment result and time evolution characteristic data, construct a causal graph model, define the nodes as defect features and production process parameters, the edges represent the causal relationships between the nodes, and initialize the node feature values and causal association matrix.

[0133] S72. Use the causal graph model to layer the nodes, dividing them into a defect feature layer, a local process parameter layer, and a global process parameter layer.

[0134] S73. Combine the cross-level multi-scale causal reasoning algorithm to calculate the direct causal relationship and the indirect causal relationship respectively. The direct relationship is calculated based on the causal intensity function between the nodes, and the indirect relationship is generated through recursive path derivation, obtaining a multi-scale causal weight matrix:

[0135]

[0136] Among them, represents the direct causal influence of node v i on v j , A ij is the hierarchical adjacency matrix, MI represents the causal association intensity between two nodes, represents the indirect causal influence of node v i on v j , P(i,j) represents the set of paths from node i to node j, |z| represents the length of path z, L is the index of the edge in the path, and A zL represents the weight or connection value of the L-th edge in path z;

[0137] S74. According to the results of multi-scale causal reasoning, generate multi-level causal paths, including the path from defect features to local process parameters, the path from local process parameters to global process parameters, and the comprehensive path, and calculate the causal intensity and path score of each path:

[0138]

[0139]

[0140] Among them, P ij represents the total causal strength from node v i to node v j . S(P ij ) represents the path score, and len(P ij ) is the path length;

[0141] S75. Analyze the sensitivity of each process parameter to defect characteristics based on the scoring results of causal paths, identify key process parameters, determine the adjustment priority by combining sensitivity analysis and causal path scoring, and generate a set of key parameters;

[0142] S76. Combine the adjustment priority of key process parameters and the causal path results to generate production process optimization suggestions, including parameter adjustment ranges, optimization directions, and their potential impacts, and output an optimization suggestion report.

[0143] Example 1:

[0144] To verify the feasibility of the present invention in implementation, the present invention is applied to a new energy battery production enterprise. This enterprise focuses on the production of high-performance lithium batteries, mainly used in the field of new energy vehicles. Since the quality of lithium battery electrodes directly affects battery performance and safety, it is necessary to strictly detect electrode defects during production. However, during the actual production process, this enterprise found that when using traditional detection methods, it was impossible to comprehensively capture obvious defects (such as uneven coating and particle accumulation) and hidden defects (such as microcracks and material internal stress) on the electrode surface, and the production yield was relatively low, resulting in relatively large economic losses.

[0145] During a large-scale production, this enterprise found that the performance of multiple batches of batteries fluctuated. Through sampling inspection, it was found that electrode cracks and particle accumulation were the main problems. Traditional detection methods are based on single-modal data and cannot effectively detect these complex defects. For this reason, the enterprise introduced the lithium battery electrode structure defect detection method based on Fourier transform proposed by the present invention, combined with multi-modal feature dynamic reconstruction, efficient deformable convolutional network, and spatio-temporal graph neural network, to conduct full-process detection and optimization of lithium battery electrodes on the production line.

[0146] When applying the present invention, multi-modal data is first collected, including visible light images, infrared thermal images, and X-ray images. However, due to hardware device problems and environmental interference, some modal data is missing. The present invention compensates for the missing modal data using a conditional generative adversarial network, and at the same time fuses the generated modality and the original modality data through a cross-modal attention mechanism to form enhanced fused feature data. Subsequently, an efficient deformable convolutional network is used to dynamically optimize these fused features to capture particle accumulation and microcrack defects existing in multiple batches.

[0147] Through dynamic expansion trend prediction, the system further models the time-evolution characteristics of defects and successfully predicts the speed and range of crack propagation. For example, in a certain batch, the initial inspection found that the crack area was 3.2 mm 2 , and the system predicted that it might expand to 8.4 mm within 10 days 2 . According to the risk assessment results, the predicted expansion area might cause local overheating of the battery and affect its performance. Based on the prediction results, the enterprise strengthened the control of coating uniformity and compaction process in subsequent production, significantly reducing the incidence of similar defects.

[0148] Based on defect detection, the system combines a cross-level multi-scale causal inference algorithm to analyze the causal relationship between defects and process parameters, identifying coating thickness and compaction force as key process parameters. Specifically, a coating thickness deviation exceeding ±3 μm is the main inducement for crack defects, while a compaction force of less than 30 N further exacerbates the incidence of particle accumulation. The system recommends optimizing the control range of coating thickness to 20 ± 2 μm and increasing the compaction force to 35 N, and implementing this optimization strategy in subsequent production.

[0149] The optimized production data shows that the product yield has increased from 91.6% before optimization to 96.8%, and the incidence of electrode defects has decreased by nearly 60%. At the same time, the detection efficiency has been significantly improved, and the detection time for each electrode has been shortened from the original 12 seconds to 8 seconds. Within one month after the full application of the present invention, the scrap rate of this production line has decreased from the original 4.7% to 1.4%, directly saving approximately 300,000 yuan in production costs.

[0150] Table 1 Comparison of Defect Detection Effects

[0151] Detection index Before optimization (traditional method) After optimization (this invention) Detection accuracy (%) 86.4 97.2 Defect recognition rate (%) 72.5 94.8 False detection rate (%) 7.8 1.6 Detection time per sample (seconds) 12 8

[0152] Table 2 Comparison of Production Yield and Scrap Rate

[0153] Index Before optimization After optimization Product yield rate (%) 91.6 96.8 Defect incidence rate (%) 8.4 3.4 Scrap rate (%) 4.7 1.4 Monthly cost savings (ten thousand yuan) - 30

[0154] As can be seen from Table 1, in terms of defect detection performance, the detection accuracy of the present invention has increased from 86.4% of the traditional method to 97.2%, the defect recognition rate has increased from 72.5% to 94.8%, and the false detection rate has decreased from 7.8% to 1.6%. This shows that the present invention exhibits extremely high accuracy in the detection of both obvious defects (such as particle accumulation) and hidden defects (such as microcracks), while significantly reducing the possibility of misjudgment. In addition, the detection time per sample has been shortened from 12 seconds of the traditional method to 8 seconds, and the detection efficiency has increased by more than 33%, greatly optimizing the detection process and meeting the requirements of high-efficiency production.

[0155] As can be seen from Table 2, the application effect of the present invention is more remarkable. Before optimization, the product yield was only 91.6%, with a defect incidence rate of 8.4% and a scrap rate of 4.7%. After optimization, the product yield increased to 96.8%, the defect incidence rate decreased to 3.4%, and the scrap rate dropped to 1.4%. This result indicates that through the dynamic defect prediction and cross-level multi-scale causal reasoning algorithm of the present invention, not only the scrap losses caused by defects in production are significantly reduced, but also the overall quality of the production line is effectively improved. In addition, by optimizing process parameters (such as coating thickness and compaction force), the monthly production cost is saved by approximately 300,000 yuan, reflecting the practical value of the present invention in reducing production costs and improving economic benefits.

[0156] Generally speaking, through comprehensive defect detection, multi-modal feature optimization, and process parameter adjustment, the present invention not only improves production efficiency and product quality, but also effectively reduces the defect risk in the production process, providing an efficient and reliable intelligent solution for the lithium battery manufacturing industry. The analysis results of the data table fully verify the technical innovation and practical application value of the present invention.

[0157] The above is only the preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A method for detecting structural defects of lithium battery electrodes based on Fourier transform, characterized in that: The steps include: S1. Collect multimodal image data of the surface of lithium battery electrodes, including visible light images, infrared thermal images and X-ray images, and preprocess the collected multimodal data, including denoising, contrast enhancement and geometric calibration; S2, performing a two-dimensional Fourier transform on the preprocessed multimodal data to generate frequency domain feature data, including low-frequency components and high-frequency components; S3. For the missing modalities in the frequency domain feature data, the existing modal data is used as conditional input, and the missing modal data is generated through the conditional generative adversarial network. After cross-modal consistency constraint optimization, the complete multimodal feature data is output; S4. Based on the complete multimodal feature data, a collaborative feature generation model is constructed to dynamically generate new feature modalities, and the generated feature modalities are fused with the original modal features through a cross-modal attention mechanism to form enhanced multimodal fusion feature data; S5. Use an efficient deformable convolutional network to dynamically optimize the enhanced multimodal fusion feature data, adaptively adjust the receptive field according to the spatial deformation characteristics of different modalities, dynamically adjust the modal fusion weights, and output the optimized multimodal fusion feature data; S6. Input the optimized multimodal fusion feature data into the spatiotemporal graph neural network, perform dynamic defect expansion trend modeling by combining time series features, and generate time evolution characteristic data of defect expansion and dynamic risk assessment results; S7. Based on the dynamic risk assessment results and time evolution characteristic data, combined with the cross-level multi-scale causal reasoning algorithm, the causal relationship between defect characteristics and production process parameters is analyzed, multi-level causal paths are derived, and production process optimization parameter recommendations are generated; S8. Apply the production process optimization parameter recommendations to the production process and output the final inspection results including defect distribution, defect type, dynamic expansion trend and optimization recommendations.

2. A lithium battery electrode structure defect detection method based on Fourier transform according to claim 1, characterized in that: The S3 specifically includes: S31, analyzing the preprocessed multimodal data to identify modality missing conditions; S32, input the existing modal data into the conditional generative adversarial network, generate missing modal data through the generator network, and use the discriminant network to perform authenticity discrimination on the generated data, optimize the adversarial training of the generator network and the discriminant network, and output the generated modal data; S33. Optimize the cross-modal consistency constraints on the generated modal data: in, is the target of the generator, is the target of the discriminator, V(D,G) is the core objective function of the generative adversarial network, G is the generator, D is the discriminator, E represents the expected value, and D r is the real data, P data is the real data distribution, log is the logarithmic function, D c is the generated modal data, P G To forge data distribution; S34, fusing the optimized generated modal data with the existing modal data to construct a complete multimodal feature data set; S35. Verify the complete multimodal feature data and evaluate the effectiveness of the generated modality through multimodal similarity measurement: Among them, S represents the similarity measure, D e For existing modal data.

3. The method for detecting structural defects of lithium battery electrodes based on Fourier transform according to claim 1, characterized in that: The S4 specifically includes: S41, performing standardization processing on the complete multimodal feature data, converting each modal feature data into a unified feature space representation, and generating a standardized feature set; S42. Based on the standardized feature set, a collaborative feature generation model is constructed to extract deep features through a multi-layer convolutional neural network: Among them, H m is the deep features extracted for modality m, ReLU represents the activation function, W m and b m are the convolution kernel weight and bias respectively, is a standardized feature set; S43. Introducing an adaptive weight allocation mechanism into the collaborative feature generation model, dynamically adjusting the modal feature weights according to the contribution of each modal feature to the target task, and generating a weighted modal feature set; S44. Using the cross-modal attention mechanism, perform feature fusion on the weighted modal feature set, establish a multi-level attention relationship between modalities, and generate preliminary fusion features: Among them, a ij represents the attention weights of modality i and modality j, exp is the exponential function, sim is the similarity function, m and k represent different modalities, represents the weighted feature representation of modality i, represents the weighted feature representation of modality j, represents the weighted feature representation of modality k; S45, inputting the preliminary fusion features into the feature enhancement module, and performing feature mapping through a multi-layer perceptron; S46. Perform feature verification on the enhanced fusion feature data and evaluate the effectiveness of the fusion feature by calculating the modal consistency and task adaptability indicators.

4. The method for detecting structural defects of lithium battery electrodes based on Fourier transform according to claim 1, characterized in that: The S5 specifically includes: S51, inputting the enhanced fusion feature data into the efficient deformable convolutional network, initializing the offset parameters of the convolution kernel, and assigning an initial weight to each modal feature; S52. Dynamically adjust the receptive field range and position of the convolution kernel by analyzing the spatial deformation characteristics of the modal features: Among them, F(p) represents the optimized feature value of pixel p, K is the convolution kernel size, Δp k is the offset, represents the enhanced fusion feature data, w k The weight of mode K; S53. Calculate the importance of each modal feature and dynamically adjust the modal fusion weight according to the contribution of the modal feature to the overall task, so that important modalities get higher priority in the fusion process: Among them, w m The importance weight of each mode, M is the number of modes, m represents the mode, β k represents the feature importance score of modality k, β m represents the feature importance score of mode m, exp is the exponential function; S54, enhancing the optimized modal features and the initial fusion features through residual connection, retaining the integrity of the initial features, and generating optimized multimodal fusion feature data; in, represents the feature data after residual optimization, is the initial fusion feature; S55, extracting global information from the optimized multimodal fusion feature data to capture the correlation between features; S56. Verify the effectiveness of the optimized multimodal fusion feature data and evaluate it using similarity indicators and the loss function of the defect detection task.

5. The method for detecting structural defects of lithium battery electrodes based on Fourier transform according to claim 1, characterized in that: The S6 specifically includes: S61, mapping the optimized multimodal fusion feature data into a spatiotemporal graph structure, defining nodes as defect feature points, edges as spatial or temporal associations between feature points, assigning an initial feature vector to each node, and establishing an adjacency matrix to describe the connection relationship between nodes; S62, inputting the time variation data of the defect feature points into the time module of the spatiotemporal graph neural network, extracting the time series features through the time convolution network, capturing the evolution law of the defect over time, and generating a dynamic time feature representation of the node; S63, using a graph convolutional network to model spatial relationships in a graph structure, extracting spatial correlation characteristics between defect feature points, generating spatial feature representations of nodes, and updating weights of edges between nodes; S64. The temporal feature representation and the spatial feature representation are fused through the spatiotemporal attention mechanism to generate a comprehensive feature representation containing temporal and spatial associations, and a spatiotemporal fused feature vector is assigned to each node: in, is the comprehensive feature representation of node v, is the temporal feature representation of node v, is the spatial feature representation of node v, α is the attention weight; S65. Based on the feature representation after spatiotemporal fusion, recursive learning is performed on the graph structure to predict the expansion trend of defects in the future, generate the time evolution path of defect feature points, and describe the dynamic change pattern of defects; S66. Evaluate the risk level of the defect based on the time evolution path of the defect expansion, generate a dynamic risk assessment result by comprehensively considering the defect location and expansion speed, and output an assessment report including the defect expansion trend and risk level.

6. The method for detecting structural defects of lithium battery electrodes based on Fourier transform according to claim 1, characterized in that: The S7 specifically includes: S71. Based on the dynamic risk assessment results and time evolution characteristic data, a causal graph model is constructed, nodes are defined as defect characteristics and production process parameters, edges represent causal relationships between nodes, and node feature values ​​and causal association matrix are initialized; S72, using a cause-and-effect graph model to layer the nodes into a defect feature layer, a local process parameter layer, and a global process parameter layer; S73. Combined with the cross-level multi-scale causal inference algorithm, direct causal relationships and indirect causal relationships are calculated respectively. The direct relationship is calculated based on the causal strength function between nodes, and the indirect relationship is generated by recursive path deduction to obtain the multi-scale causal weight matrix: in, Represents node v i v j The direct causal effect of A ij is a hierarchical adjacency matrix, MI represents the causal relationship strength between two nodes, Represents node v i v j Indirect causal influence, P(i,j) represents the set of paths from node i to node j, |z| represents the length of path z, L is the index of the edge in the path, A zL represents the weight or connection value of the Lth edge in path z; S74. Based on the results of multi-scale causal reasoning, generate multi-level causal paths, including the path from defect features to local process parameters, the path from local process parameters to global process parameters, and the comprehensive path, and calculate the causal strength and path score of each path: Among them, P ij Represents node v i To node v j The total causal strength, S(P ij ) represents the path score, len(P ij ) is the path length; S75. Based on the scoring results of the causal path, analyze the sensitivity of each process parameter to the defect characteristics, identify the key process parameters, determine the adjustment priority by combining the sensitivity analysis and the causal path scoring, and generate a set of key parameters; S76. Combine the adjustment priorities of key process parameters and the causal path results to generate production process optimization suggestions, including parameter adjustment range, optimization direction and its potential impact, and output an optimization suggestion report.

Images