Lithium electrode group defect recognition method based on deep learning

Abstract

The present application relates to the technical field of deep learning, in particular to a lithium electrode group defect identification method based on deep learning, comprising the following steps: collecting a three-dimensional multi-modal image set of a lithium electrode group to be detected; extracting two-dimensional texture features and three-dimensional elevation features of each three-dimensional multi-modal image in the three-dimensional multi-modal image set; fusing the two-dimensional texture features and the three-dimensional elevation features according to a cross-attention mechanism to obtain cross-modal joint features; and processing the cross-modal joint features through a preset unsupervised normalization flow network combined with a coating thickness prior distribution. In the present application, when the feature distribution of a certain electrode group deviates from a standard model, a quantitative anomaly score is calculated, and a pixel-level defect positioning heat map can be directly generated on the original image space by using reverse mapping technology, so that the specific position and contour of the defect are highlighted, thereby realizing a closed loop from defect detection to defect positioning.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method for identifying defects in lithium electrode assemblies based on deep learning. Background Technology

[0002] Existing technologies for lithium electrode assembly defect identification suffer from inherent limitations in practical operation, including a single perception dimension and insufficient model generalization ability. Many detection schemes rely primarily on two-dimensional image analysis. While effective for defects with clear color and texture contrasts, such as foreign objects and scratches, this approach is almost ineffective for geometric defects exhibiting only significant variations, such as tiny dents, wrinkles, or uneven coating thickness. This is because these defects may not show any optical differences in a two-dimensional image, leading to a risk of missed detections. For example, a slight dent caused by a drying process issue may appear identical to a normal area in a two-dimensional image, but it can significantly impact the battery's internal stress distribution and electrochemical performance, rendering traditional two-dimensional vision methods ineffective. On the other hand, most current deep learning-based solutions employ supervised learning, whose performance heavily depends on a large and comprehensive defect sample database. This is virtually impossible to achieve in actual production, as defect morphologies are highly variable, and new defect types constantly emerge with process fluctuations. Therefore, improvements are needed. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing a deep learning-based method for identifying defects in lithium electrode packs.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a lithium electrode assembly defect identification method based on deep learning, comprising the following steps:

[0005] Acquire a set of three-dimensional multimodal images of the lithium electrode assembly to be tested;

[0006] Extract the two-dimensional texture features and three-dimensional elevation features of each three-dimensional multimodal image in the three-dimensional multimodal image set;

[0007] The two-dimensional texture features and the three-dimensional elevation features are fused according to the cross-attention mechanism to obtain cross-modal joint features;

[0008] The cross-modal joint features are processed by a pre-defined unsupervised normalized flow network that incorporates the prior distribution of coating thickness to obtain feature mapping distribution data.

[0009] Calculate the distribution offset of the feature mapping distribution data to obtain the anomaly score;

[0010] The defect status of the lithium electrode assembly is determined based on the anomaly score and the preset anomaly determination threshold.

[0011] When it is determined that the lithium electrode group is in the defective state, the feature mapping distribution data is reverse-mapped to obtain a pixel-level mapping matrix.

[0012] An anomaly heatmap is generated based on the pixel-level mapping matrix, and the defect locations are marked on the anomaly heatmap.

[0013] Preferably, the steps for acquiring a set of three-dimensional multimodal images of the lithium electrode assembly to be tested are as follows:

[0014] The scanning frame rate parameter is obtained, and the production line laser profilometer is controlled to scan the three-dimensional contour of the lithium electrode assembly to be inspected according to the scanning frame rate parameter to obtain the first scanning data.

[0015] The production line laser profilometer is controlled to scan the three-dimensional contour of the edge of the lithium electrode assembly to be inspected, and second scan data is obtained;

[0016] The first scan data and the second scan data are stitched together to obtain three-dimensional point cloud data;

[0017] Obtain the nearest neighbor number parameter, and use the statistical outlier filtering algorithm to filter out noise points in the 3D point cloud data based on the nearest neighbor number parameter to obtain cleaned 3D point cloud data. Then, perform projection processing on the cleaned 3D point cloud data to obtain a 3D depth map.

[0018] The exposure parameters are obtained, and the binocular vision camera is controlled to capture the surface texture of the lithium electrode assembly to be detected based on the exposure parameters to obtain a two-dimensional color image;

[0019] A reflective mask is generated by extracting the highly reflective metallic areas from the two-dimensional color image. After removing the high-brightness interference information from the two-dimensional color image and the three-dimensional depth image using the reflective mask, the corner features of the two-dimensional color image and the edge features of the three-dimensional depth image are extracted respectively.

[0020] The three-dimensional depth map and the two-dimensional color image are aligned based on the corner features and the edge features to obtain aligned multimodal image data;

[0021] The aligned multimodal image data is continuously collected, and all the aligned multimodal image data are merged to obtain the three-dimensional multimodal image set.

[0022] Preferably, the steps for extracting the two-dimensional texture features and three-dimensional elevation features of each three-dimensional multimodal image in the three-dimensional multimodal image set are as follows:

[0023] Extract the channel-level data of each three-dimensional multimodal image in the three-dimensional multimodal image set, separate the channel-level data, and obtain the first visual channel data and the second depth channel data;

[0024] A backbone visual converter network is constructed, the first visual channel data is divided into multiple image block data, the image block data is input into the backbone visual converter network, the self-attention mechanism of the backbone visual converter network is used to extract the texture distribution information of the image block data in the planar dimension to obtain the first texture feature vector, and the first texture feature vector is flattened to obtain the two-dimensional texture feature.

[0025] A backbone 3D convolutional network with residual connection structure is constructed. The second depth channel data is input into the backbone 3D convolutional network. The spatial convolution kernel of the backbone 3D convolutional network is used to extract the geometric distribution information of the second depth channel data in the spatial height dimension to obtain the first elevation feature tensor. The first elevation feature tensor is downsampled using the downsampling layer of the backbone 3D convolutional network to obtain the 3D elevation feature.

[0026] Preferably, the step of fusing the two-dimensional texture features and the three-dimensional elevation features according to the cross-attention mechanism to obtain cross-modal joint features is as follows:

[0027] Extract the query vector matrix of the two-dimensional texture features, and extract the key vector matrix and value vector matrix of the three-dimensional elevation features;

[0028] Obtain the feature dimension scaling factor parameter, perform a dot product operation on the transpose of the query vector matrix and the key vector matrix to obtain the initial association weight matrix, and divide the initial association weight matrix by the feature dimension scaling factor parameter to obtain the scaled association weight matrix;

[0029] The scaling correlation weight matrix is ​​nonlinearly mapped using an exponential normalization function to obtain a normalized attention matrix;

[0030] Calculate the spatial geometric gradient information of the three-dimensional elevation features, and generate a local elevation mask matrix based on the spatial geometric gradient information;

[0031] After spatial weight modulation of the normalized attention matrix using the local elevation mask matrix, matrix multiplication is performed on the value vector matrix to obtain the three-dimensional activation features.

[0032] The three-dimensional activation features and the two-dimensional texture features are concatenated to obtain multi-dimensional fused tensor data. The multi-dimensional fused tensor data is then input into a fully connected perceptron network for parameter dimensionality reduction to obtain the cross-modal joint features.

[0033] Preferably, the step of processing the cross-modal joint features through a pre-defined unsupervised normalized flow network that incorporates the prior distribution of coating thickness to obtain the feature mapping distribution data specifically includes:

[0034] Obtain a normal sample training dataset containing normal lithium electrode group samples, and use the normal sample training dataset to perform unsupervised training on the initial mapping network model to obtain the preset unsupervised normalized flow network that combines the prior distribution of coating thickness.

[0035] The cross-modal joint features are input into the preset unsupervised normalized flow network that combines the prior distribution of coating thickness, and the scaling parameter matrix and translation offset parameter matrix in the preset unsupervised normalized flow network that combines the prior distribution of coating thickness are obtained. The cross-modal joint features are then subjected to a nonlinear invertible mapping transformation based on the scaling parameter matrix and the translation offset parameter matrix to obtain the forward mapping features.

[0036] Obtain the Jacobian matrix of the preset unsupervised normalized flow network that combines the prior distribution of coating thickness, and calculate the determinant parameter value corresponding to the Jacobian matrix.

[0037] Obtain pre-constructed reference thickness gradient data for lithium electrode assembly, and construct a spatial prior correction factor based on the reference thickness gradient data for lithium electrode assembly;

[0038] Based on the determinant parameter value and the spatial prior correction factor, the log probability density of the forward mapping feature is evaluated and calculated to obtain the target feature vector distribution.

[0039] Align the target feature vector distribution with the standard normal distribution, extract the distribution node data from the target feature vector distribution, and merge the distribution node data to obtain the feature mapping distribution data.

[0040] Preferably, the step of calculating the distribution offset of the feature mapping distribution data to obtain the anomaly score specifically includes:

[0041] Obtain standard high-dimensional normal distribution template data, extract the first mean parameter of the feature mapping distribution data and the second mean parameter of the standard high-dimensional normal distribution template data, calculate the absolute difference between the first mean parameter and the second mean parameter, and obtain the mean distribution difference;

[0042] Extract the first variance parameter of the feature mapping distribution data and the second variance parameter of the standard high-dimensional normal distribution template data, calculate the deviation between the first variance parameter and the second variance parameter, and obtain the variance distribution difference;

[0043] Obtain the joint evaluation function model, input the mean distribution difference and the variance distribution difference into the joint evaluation function model, evaluate the cumulative result of the mean distribution difference and the variance distribution difference, and obtain the distribution offset;

[0044] Obtain the preset maximum tolerance distribution offset parameter and the preset minimum tolerance distribution offset parameter, and extract the relative position data of the distribution offset between the maximum tolerance distribution offset parameter and the minimum tolerance distribution offset parameter;

[0045] The distribution offset is mapped to a continuous numerical range of scores based on the relative position data to obtain the abnormal score.

[0046] Preferably, the step of determining the defect state of the lithium electrode assembly based on the anomaly score and the preset anomaly determination threshold specifically includes:

[0047] Extract the first floating-point value corresponding to the abnormal score, obtain the second floating-point value corresponding to the preset abnormal judgment threshold, and compare the first floating-point value with the second floating-point value.

[0048] When the first floating-point value is greater than the second floating-point value, it is determined that the lithium electrode assembly has a defect, and the lithium electrode assembly is determined to be in the defective state.

[0049] When the first floating-point value is not greater than the second floating-point value, it is determined that the feature mapping distribution data is within the allowable fluctuation range, and the lithium electrode group is determined to be in a normal state.

[0050] Preferably, when the lithium electrode assembly is determined to be in the defective state, the step of performing reverse mapping on the feature mapping distribution data to obtain a pixel-level mapping matrix specifically includes:

[0051] Obtain feature mapping distribution data of lithium electrode groups in defective state, and construct a defective state data set from the obtained feature mapping distribution data;

[0052] Extract the set of parameters of the inverse computation neuron of the unsupervised normalized flow network with the prior distribution of coating thickness, and establish the inverse affine transformation function of the unsupervised normalized flow network with the prior distribution of coating thickness based on the set of parameters of the inverse computation neuron.

[0053] The defect state data set is input into the inverse affine transformation function, and the inverse affine transformation function is controlled to perform a layer-by-layer reverse invertible mapping calculation on the defect state data set to obtain high-dimensional restored features in the two-dimensional spatial domain.

[0054] Extract the feature response intensity value corresponding to each pixel position in the high-dimensional restored feature;

[0055] Obtain the original image width and height parameters corresponding to the three-dimensional multimodal image set, and arrange all the feature response intensity values ​​in an array according to the original image width and height parameters to obtain the pixel-level mapping matrix.

[0056] Preferably, the step of generating an anomaly heatmap based on the pixel-level mapping matrix and marking the defect location in the anomaly heatmap specifically includes:

[0057] Traverse and extract the response value of each mapping node in the pixel-level mapping matrix, establish a color gradient mapping reference dictionary, and look up the corresponding rendering color channel value in the color gradient mapping reference dictionary according to the response value.

[0058] The two-dimensional plane of the pixel-level mapping matrix is ​​rendered and colored according to the rendering color channel values ​​to obtain the abnormal heatmap;

[0059] Obtain the preset red highlight color threshold parameter, and extract the extreme connected regions in the abnormal heatmap where the rendered color channel value is greater than the red highlight color threshold parameter;

[0060] Extract the center coordinates and boundary contour information of the extreme connected region, and calculate the size parameters of the bounding rectangle surrounding the extreme connected region based on the boundary contour information.

[0061] In the abnormal heatmap, a rectangular defect bounding box is drawn in the local area corresponding to the center coordinate point location information using the circumscribed rectangle size parameter, and the defect location is marked on the abnormal heatmap based on the rectangular defect bounding box.

[0062] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0063] This invention, by acquiring composite multimodal images containing two-dimensional texture and three-dimensional elevation, can construct a complete and three-dimensional digital description of lithium electrode assemblies, laying a solid data foundation for subsequent analysis. In processing this data, instead of simple information overlay, a deep correlation fusion operation is performed to intelligently discover and combine the intrinsic connections between two-dimensional texture features and three-dimensional elevation features. This allows the characterization of defects to no longer be limited to a single dimension, enabling the identification of complex defects that only appear under multidimensional information coupling. The core logic of the detection is based on the unsupervised learning paradigm. Only normal, qualified electrode assembly samples need to be learned to construct a high-dimensional feature distribution model of the standard product. This fundamentally avoids the huge challenge of traditional supervised learning methods requiring the prior collection and labeling of massive amounts of diverse defect samples, improving the model's ability to generalize and identify unknown and novel defects. When the feature distribution of an electrode assembly deviates from the standard model, a quantified anomaly score is calculated. Using inverse mapping technology, a pixel-level defect location heatmap can be directly generated in the original image space, highlighting the specific location and contour of the defect, thus achieving a closed loop from defect detection to defect location. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0066] Please see Figure 1 This invention provides a technical solution: a method for identifying defects in lithium electrode assemblies based on deep learning, comprising the following steps:

[0067] Acquire a set of three-dimensional multimodal images of the lithium electrode assembly to be tested;

[0068] Extract the two-dimensional texture features and three-dimensional elevation features of each three-dimensional multimodal image in the three-dimensional multimodal image set;

[0069] By fusing two-dimensional texture features and three-dimensional elevation features using a cross-attention mechanism, cross-modal joint features are obtained;

[0070] The cross-modal joint features are processed by a pre-defined unsupervised normalized flow network that combines the prior distribution of coating thickness to obtain feature mapping distribution data;

[0071] Calculate the distribution offset of the feature map distribution data to obtain the anomaly score;

[0072] The defect status of the lithium electrode assembly is determined based on the anomaly score and the preset anomaly judgment threshold.

[0073] When the lithium electrode assembly is determined to be in a defective state, the feature mapping distribution data is reverse-mapped to obtain a pixel-level mapping matrix.

[0074] An anomaly heatmap is generated based on a pixel-level mapping matrix, and the location of defects is marked on the anomaly heatmap.

[0075] In this embodiment, the step of acquiring a set of three-dimensional multimodal images of the lithium electrode assembly to be detected specifically includes: obtaining a scanning frame rate parameter; controlling the production line laser profilometer to scan the three-dimensional contour of the lithium electrode assembly to be detected according to the scanning frame rate parameter to obtain first scanning data; controlling the production line laser profilometer to scan the edge three-dimensional contour of the lithium electrode assembly to be detected to obtain second scanning data; stitching the first scanning data and the second scanning data together to obtain three-dimensional point cloud data; obtaining a nearest neighbor point quantity parameter; using a statistical outlier filtering algorithm to filter out noise points in the three-dimensional point cloud data according to the nearest neighbor point quantity parameter to obtain cleaned three-dimensional point cloud data; and performing... Projection processing yields a 3D depth map; exposure parameters are acquired, and a binocular vision camera is used to capture the surface texture of the lithium electrode assembly under inspection, resulting in a 2D color image; highly reflective metal areas are extracted from the 2D color image to generate a reflective mask; after removing high-brightness interference information from both the 2D color image and the 3D depth map using the reflective mask, corner features from the 2D color image and edge features from the 3D depth map are extracted respectively; the 3D depth map and the 2D color image are aligned based on the corner and edge features to obtain aligned multimodal image data; the aligned multimodal image data is continuously acquired, and all aligned multimodal image data are merged to obtain a 3D multimodal image set.

[0076] Specifically, in order to acquire a set of three-dimensional multimodal images of the lithium electrode assembly to be inspected, the implementation process first requires setting a scanning frame rate parameter adapted to the production line, for example, setting it to 200 frames per second based on the speed of the production line conveyor belt. Based on this parameter, the production line's linear laser profilometer is controlled to perform high-speed scanning of the main three-dimensional contour of the lithium electrode assembly to be inspected, recording its contour point cloud as the first scan data. Then, the laser power and exposure time are adjusted to control the production line's linear laser profilometer to perform a fine scan of the edge-wound area of ​​the lithium electrode assembly to be inspected, obtaining a second scan containing more details. The first and second scan data are registered and stitched together using the Iterative Nearest Point (ICP) algorithm to form a complete 3D point cloud. After obtaining the nearest neighbor parameter (e.g., set to 50 nearest neighbors), a statistical outlier filtering algorithm is used to process the 3D point cloud data. The outlier detection threshold for this algorithm is set to the global point cloud average distance plus 1.5 times the global distance standard deviation. Any point whose average distance to its 50 nearest neighbors exceeds this threshold is identified as a noise point and filtered out, resulting in cleaned 3D point cloud data. The cleaned 3D point cloud data is processed using orthogonal projection to obtain a pixel-level 3D depth map. Simultaneously, exposure parameters are acquired, for example, set to 100 microseconds. Based on these parameters, a binocular vision camera is used to capture the surface texture of the lithium electrode assembly under inspection, resulting in a high-resolution 2D color image. This 2D color image is then converted to grayscale, and pixels with grayscale values ​​greater than 240 are identified as highly reflective metallic areas, generating corresponding reflective masks. These masks are then used to simultaneously remove high-brightness interference information from both the 2D color image and the 3D depth map using a zeroing operation. Harris corner features from the 2D color image and Canny edge features from the 3D depth map are then extracted. Based on these matched corner and edge features, the optimal affine transformation matrix is ​​calculated using the Random Sample Consensus (RANSAC) algorithm. Pixel-level registration and alignment are then performed on the 3D depth map and the 2D color image to obtain aligned multimodal image data. During production line operation, aligned multimodal image data is continuously collected, and each frame is merged as an element in a set to obtain the final 3D multimodal image set.

[0077] In this embodiment, the steps for extracting the two-dimensional texture features and three-dimensional elevation features of each three-dimensional multimodal image in the three-dimensional multimodal image set are as follows: extracting the channel-level data of each three-dimensional multimodal image in the three-dimensional multimodal image set, separating the channel-level data to obtain the first visual channel data and the second depth channel data; constructing a backbone visual converter network, dividing the first visual channel data into multiple image block data, inputting the image block data into the backbone visual converter network, using the self-attention mechanism of the backbone visual converter network to extract the texture distribution information of the image block data in the planar dimension to obtain the first texture feature vector, flattening the first texture feature vector to obtain the two-dimensional texture features; constructing a backbone three-dimensional convolutional network containing residual connection structures, inputting the second depth channel data into the backbone three-dimensional convolutional network, using the spatial convolution kernel of the backbone three-dimensional convolutional network to extract the geometric distribution information of the second depth channel data in the spatial height dimension to obtain the first elevation feature tensor, using the downsampling layer of the backbone three-dimensional convolutional network to downsample the first elevation feature tensor to obtain the three-dimensional elevation features.

[0078] Specifically, to extract the 2D texture features and 3D elevation features of each 3D multimodal image in the 3D multimodal image set, the implementation first extracts the channel-level data of each 3D multimodal image, separating it into first visual channel data representing RGB color information and second depth channel data representing height information. Then, a backbone visual converter network is constructed to process the first visual channel data. The network model structure is as follows: the input layer receives a 224×224 pixel image; the image segmentation layer segments the input image into 196 16×16 pixel image blocks without overlap; the linear embedding layer linearly maps each flattened image block data into a 768-dimensional vector and adds a learnable positional code to each vector; followed by a 12-layer stacked Transformer encoder, each encoder consisting of a 12-head multi-head self-attention module and a feedforward network, using GELU as the activation function, and applying layer normalization and residual connections to input the positionally encoded image block data sequence. The data is fed into the backbone visual converter network, where its self-attention mechanism is used to calculate the dependencies between image patches, thereby extracting the global texture distribution information of the image patch data in the planar dimension to obtain the first texture feature vector. The first texture feature vector is then subjected to global average pooling and flattening operations to obtain two-dimensional texture features. At the same time, a backbone three-dimensional convolutional network with residual connection structure is constructed to process the second depth channel data. This network contains 5 three-dimensional convolutional blocks, each consisting of a 3×3×3 three-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. The second depth channel data is input into the network as 224×224×1 volume data. The spatial convolutional kernel is used to extract the local geometric distribution information of the second depth channel data in the spatial height dimension layer by layer to obtain the first elevation feature tensor containing multi-scale geometric information. The global average pooling layer at the end of the backbone three-dimensional convolutional network is used as a downsampling layer to downsample the first elevation feature tensor to obtain a fixed-dimensional three-dimensional elevation feature.

[0079] In this embodiment, the steps for fusing two-dimensional texture features and three-dimensional elevation features according to the cross-attention mechanism to obtain cross-modal joint features are as follows: Extract the query vector matrix of the two-dimensional texture features, and extract the key vector matrix and value vector matrix of the three-dimensional elevation features; obtain the feature dimension scaling factor parameter, perform a dot product operation on the transpose of the query vector matrix and the key vector matrix to obtain the initial association weight matrix, divide the initial association weight matrix by the feature dimension scaling factor parameter to obtain the scaled association weight matrix; perform nonlinear mapping processing on the scaled association weight matrix using an exponential normalization function to obtain the normalized attention matrix; calculate the spatial geometric gradient information of the three-dimensional elevation features, and generate a local elevation mask matrix based on the spatial geometric gradient information; after spatial weight modulation of the normalized attention matrix using the local elevation mask matrix, perform matrix multiplication on the value vector matrix to obtain the three-dimensional activation features; concatenate the three-dimensional activation features and the two-dimensional texture features to obtain multidimensional fused tensor data, and input the multidimensional fused tensor data into a fully connected perceptron network for parameter dimensionality reduction to obtain the cross-modal joint features.

[0080] Specifically, to fuse 2D texture features and 3D elevation features based on the cross-attention mechanism, the 2D texture features obtained in the previous step are linearly transformed and used as the query vector matrix. The 3D elevation features are transformed by two other independent linear transformations and used as the key vector matrix and value vector matrix, respectively. A feature dimension scaling factor parameter is obtained; the value of this parameter is the square root of the feature dimension of the key vector matrix. For example, if the feature dimension of the key vector matrix is ​​64, the scaling factor parameter is 8. A matrix dot product operation is performed on the transposes of the query vector matrix and the key vector matrix to obtain an initial association weight matrix reflecting the initial association strength between the two modal features. Each element in the initial association weight matrix is ​​divided by the feature dimension scaling factor parameter to obtain the scaled association weight matrix. Subsequently, the exponential normalization function, i.e., the Softmax function, is used to perform non-linear mapping on each row of the scaled association weight matrix to obtain a normalized attention matrix with values ​​between 0 and 1. Simultaneously, a 3×3 Sobel operator is applied to the 3D elevation features to calculate their spatial geometric gradient information in the horizontal and vertical directions. Regions with gradient magnitudes greater than a preset gradient threshold of 0.8 are marked as 1, and other regions are marked as 0.2, thereby generating a local elevation mask matrix. The gradient threshold of 0.8 is set based on the gradient statistical analysis of a large number of normal samples to characterize regions of geometric abrupt change. The local elevation mask matrix is ​​multiplied element-wise with the normalized attention matrix to achieve spatial weight modulation. The modulated matrix is ​​then multiplied with the value vector matrix to obtain a 3D activation feature that incorporates elevation abrupt change information. This 3D activation feature is then concatenated with the original 2D texture feature in the feature dimension to obtain a multidimensional fusion tensor data with an increased dimension. The multidimensional fusion tensor data is input into a fully connected perceptron network consisting of two linear layers and a ReLU activation function in between for parameter dimensionality reduction and information integration to obtain the final cross-modal joint feature.

[0081] In this embodiment, the specific steps for processing cross-modal joint features and obtaining feature mapping distribution data through a pre-defined unsupervised normalized flow network combining the prior distribution of coating thickness are as follows: Obtain a normal sample training dataset containing normal lithium electrode group samples; perform unsupervised training on the initial mapping network model using the normal sample training dataset to obtain a pre-defined unsupervised normalized flow network combining the prior distribution of coating thickness; input the cross-modal joint features into the pre-defined unsupervised normalized flow network combining the prior distribution of coating thickness, obtain the scaling parameter matrix and translation offset parameter matrix in the pre-defined unsupervised normalized flow network combining the prior distribution of coating thickness, and then... The forward mapping features are obtained by performing a nonlinear invertible mapping transformation on the cross-modal joint features. The Jacobian matrix of the unsupervised normalized flow network combining the prior distribution of coating thickness is obtained, and the determinant parameter value corresponding to the Jacobian matrix is ​​calculated. The pre-constructed reference thickness gradient data of the lithium electrode group is obtained, and a spatial prior correction factor is constructed based on the reference thickness gradient data of the lithium electrode group. The log probability density of the forward mapping features is evaluated and calculated based on the determinant parameter value and the spatial prior correction factor to obtain the target feature vector distribution. The target feature vector distribution is aligned with the standard normal distribution, the distribution node data in the target feature vector distribution is extracted, and the distribution node data is merged to obtain the feature mapping distribution data.

[0082] Specifically, to process cross-modal joint features using a pre-defined unsupervised normalized flow network incorporating the prior distribution of coating thickness, the process is divided into two phases: training and inference. In the training phase, a normal sample training dataset containing only 5000 normal lithium electrode assembly samples is first acquired. This dataset is used to unsupervised train an initial mapping network model based on the RealNVP architecture. This network consists of eight stacked affine coupling layers. Each layer halves the input feature vector, and one part is processed by a three-layer residual network to calculate scaling and translation parameters for transforming the other part. The training objective is to learn an invertible mapping function that transforms the input normal sample feature distribution into a standard normal distribution. Its loss function is negative log-likelihood. By minimizing this loss, the pre-defined unsupervised normalized flow network incorporating the prior distribution of coating thickness is obtained. In the inference phase, the cross-modal joint features of the samples to be detected are input into the trained network. In the network, the pre-stored scaling parameter matrix and translation offset parameter matrix of each layer are obtained. Based on these parameters, a nonlinear invertible mapping transformation is performed on the cross-modal joint features to obtain their representation in the latent space, i.e., the forward mapping features. The Jacobian matrix of the network with respect to the input variables is obtained, and the logarithm of its determinant is calculated. At the same time, the reference thickness gradient data of the lithium electrode assembly, which is pre-constructed by meshing and gradient calculation of the ideal CAD model, is obtained, and a spatial prior correction factor is constructed based on this. This factor assigns higher weights to the thickness-sensitive regions. According to the variable substitution formula, combined with the determinant parameter value and the spatial prior correction factor, the log probability density of the forward mapping features is evaluated and calculated to obtain the target feature vector distribution. The target feature vector distribution is aligned and compared with the standard normal distribution. All distribution node data in the target feature vector distribution are extracted and merged into a set to obtain the feature mapping distribution data.

[0083] In this embodiment, the steps for calculating the distribution offset of the feature mapping distribution data to obtain the abnormal score are as follows: Obtain standard high-dimensional normal distribution template data; extract the first mean parameter of the feature mapping distribution data and the second mean parameter of the standard high-dimensional normal distribution template data; calculate the absolute difference between the first mean parameter and the second mean parameter to obtain the mean distribution difference; extract the first variance parameter of the feature mapping distribution data and the second variance parameter of the standard high-dimensional normal distribution template data; calculate the deviation between the first variance parameter and the second variance parameter to obtain the variance distribution difference; obtain a joint evaluation function model; input the mean distribution difference and the variance distribution difference into the joint evaluation function model; evaluate the cumulative result of the mean distribution difference and the variance distribution difference to obtain the distribution offset; obtain a preset maximum tolerance distribution offset parameter and a preset minimum tolerance distribution offset parameter; extract the relative position data of the distribution offset between the maximum tolerance distribution offset parameter and the minimum tolerance distribution offset parameter; map the distribution offset to a continuous numerical interval of the score based on the relative position data to obtain the abnormal score.

[0084] Specifically, to calculate the distribution offset of the feature map distribution data, a standard high-dimensional normal distribution template data needs to be obtained first. This template data has a zero mean vector and a covariance matrix that is an identity matrix. Then, the mean vectors of the multi-dimensional features of the obtained feature map distribution data are calculated as the first mean parameter, and the mean vector of the standard high-dimensional normal distribution template data is extracted as the second mean parameter. The mean distribution difference is obtained by summing the absolute values ​​of the differences between the first and second mean parameters in each corresponding dimension. The covariance matrix of the feature map distribution data is calculated as the first variance parameter, and the covariance matrix of the standard high-dimensional normal distribution template data is extracted as the second variance parameter. The deviation between these two covariance matrices is quantified by calculating the Kullback-Leibler divergence, thus obtaining the variance distribution difference. Finally, a pre-defined joint evaluation function model is obtained. This model is an additive... The weighted summation function uses weights set based on historical data analysis. For example, the weight of mean difference is set to 0.6, and the weight of variance difference is set to 0.4, because mean deviation usually reflects systematic anomalies better. The calculated mean and variance distribution differences are substituted into the function model, and their weighted summation is evaluated to obtain the distribution deviation. Then, the preset maximum and minimum tolerance distribution deviation parameters are obtained. These two parameters are obtained by statistically analyzing the distribution deviations of all normal samples during the training phase, taking the maximum value as the maximum tolerance distribution deviation parameter, and the minimum value is usually set to 0. For example, the maximum tolerance value is 12.5 and the minimum is 0. The relative position data of the currently calculated distribution deviation within the interval defined by these two parameters is extracted. Based on this relative position data, the distribution deviation is mapped to a continuous numerical range of scores from 0 to 100 through linear normalization to obtain the abnormal score.

[0085] In this embodiment, the step of determining the defect status of the lithium electrode assembly based on the anomaly score and the preset anomaly judgment threshold is as follows: extract the first floating-point value result corresponding to the anomaly score, obtain the second floating-point value result corresponding to the preset anomaly judgment threshold, and compare the first floating-point value result with the second floating-point value result; when the first floating-point value result is greater than the second floating-point value result, it is determined that the lithium electrode assembly has a defect and the lithium electrode assembly is in a defective state; when the first floating-point value result is not greater than the second floating-point value result, it is determined that the feature mapping distribution data is within the allowable fluctuation range and the lithium electrode assembly is in a normal state.

[0086] Specifically, to determine the defect status of the lithium electrode assembly based on the anomaly score and a preset anomaly threshold, the first floating-point value corresponding to the anomaly score is extracted from the calculation results of the previous step, for example, 78.5. Simultaneously, the second floating-point value corresponding to the preset anomaly threshold is obtained. The process of setting this threshold is as follows: Prepare a validation dataset containing 1000 normal samples and 1000 known defect samples. Calculate the anomaly score for all samples in this dataset, plot the receiver operating characteristic (ROC) curve, and calculate the F1 score under different thresholds. Select the anomaly score that maximizes the F1 score as the final anomaly threshold. For example, calculations show that when the threshold is 65.0, the F1 score... Since the highest value is found, the second floating-point value is set to 65.0. The extracted first floating-point value of 78.5 is compared with the second floating-point value of 65.0. In this example, since the first floating-point value is greater than the second floating-point value, the lithium electrode assembly being tested is determined to have a defect, and its state is determined to be defective. Conversely, if the first floating-point value calculated for another sample is 32.0, since it is not greater than the second floating-point value of 65.0, the feature mapping distribution data corresponding to that sample is determined to be within the allowable normal fluctuation range, and its state is determined to be normal. In this way, the final state determination is made for each lithium electrode assembly to be tested.

[0087] In this embodiment, when the lithium electrode assembly is determined to be in a defective state, the step of performing reverse mapping on the feature mapping distribution data to obtain a pixel-level mapping matrix is ​​as follows: The feature mapping distribution data of the lithium electrode assembly in a defective state is obtained, and the obtained feature mapping distribution data is used to construct a defective state data set; a preset set of parameters for the inverse computation neuron of an unsupervised normalized flow network combining a prior distribution of coating thickness is extracted; a preset inverse affine transformation function of an unsupervised normalized flow network combining a prior distribution of coating thickness is established based on the set of parameters for the inverse computation neuron; the defective state data set is input into the inverse affine transformation function, and the inverse affine transformation function is controlled to perform layer-by-layer reverse inverse mapping calculation on the defective state data set to obtain high-dimensional restored features in the two-dimensional spatial domain; the feature response intensity value corresponding to each pixel position in the high-dimensional restored features is extracted; the original image width and height parameters corresponding to the three-dimensional multimodal image set are obtained, and all feature response intensity values ​​are arrayed according to the original image width and height parameters to obtain a pixel-level mapping matrix.

[0088] Specifically, when the lithium electrode assembly is determined to be in a defective state, it is necessary to perform a reverse mapping of the feature mapping distribution data to obtain a pixel-level mapping matrix. The specific implementation details are as follows: First, the feature mapping distribution data corresponding to the lithium electrode assembly in the defective state is obtained. This dataset is used as the input for the inverse calculation, i.e., the defective state data set. Simultaneously, from a pre-trained, pre-defined unsupervised normalized flow network that incorporates the prior distribution of coating thickness, the set of neuron parameters used for performing the inverse calculation is extracted from all affine coupling layers. These parameters are consistent with those used in the forward propagation, but the calculation process is reversed. Based on this inverse calculation neuron parameter set, an inverse affine transformation function of the network is established. This function is the exact mathematical inverse of the forward mapping function. The defective state data set is input into this inverse affine transformation function, controlling the function to perform a layer-by-layer reverse calculation of the defective state data set. Reversible mapping computation involves recovering the original high-dimensional feature space layer by layer from the standard normal distribution space until it returns to the original high-dimensional feature space. This results in a high-dimensional restored feature in a two-dimensional spatial domain with the same dimension as the cross-modal joint feature input during forward propagation. To quantify the degree of anomaly, an average feature of a normal sample is also calculated as a reference benchmark. This is obtained by averaging the cross-modal joint features of all normal samples in the training set. Then, the Euclidean distance between the high-dimensional restored feature and the reference benchmark feature is calculated, and this distance value is used as the feature response intensity value corresponding to each pixel position. The image size parameters of the original three-dimensional multimodal image set are obtained, namely the original image width and height parameters, for example, 224×224. According to these two size parameters, the feature response intensity values ​​corresponding to all pixels are rearranged into a two-dimensional matrix to obtain the pixel-level mapping matrix.

[0089] In this embodiment, the steps of generating an anomaly heatmap based on the pixel-level mapping matrix and marking the defect location on the anomaly heatmap are as follows: traversing and extracting the response value of each mapping node in the pixel-level mapping matrix, establishing a color gradient mapping reference dictionary, and querying the corresponding rendering color channel value in the color gradient mapping reference dictionary based on the response value; rendering and coloring the two-dimensional plane of the pixel-level mapping matrix according to the rendering color channel value to obtain the anomaly heatmap; obtaining a preset red highlight color threshold parameter, and extracting extreme connected regions in the anomaly heatmap where the rendering color channel value is greater than the red highlight color threshold parameter; extracting the center coordinate point location information and boundary contour information of the extreme connected regions, and calculating the outer rectangle size parameter surrounding the extreme connected regions based on the boundary contour information; drawing a rectangular defect bounding box in the local area corresponding to the center coordinate point location information in the anomaly heatmap using the outer rectangle size parameter, and marking the defect location on the anomaly heatmap based on the rectangular defect bounding box.

[0090] Specifically, to generate an anomaly heatmap based on the pixel-level mapping matrix and mark defect locations within it, it is first necessary to traverse and extract the response values ​​of each mapping node in the pixel-level mapping matrix, and establish a color gradient mapping reference dictionary. This dictionary non-linearly maps response values ​​from minimum to maximum to a set of colors. For example, using a color scheme that smoothly transitions from blue to green, then to yellow, and finally to red, the lowest 10% of values ​​are mapped to blue, and the highest 10% to red. Based on the response value of each mapping node, the corresponding rendering color channel value is looked up in the color gradient mapping reference dictionary. Then, the entire two-dimensional plane of the pixel-level mapping matrix is ​​rendered and colored according to the retrieved rendering color channel value to obtain a visualized anomaly heatmap. A preset red highlight color threshold parameter is obtained, the standard for which is to set the threshold value of all response values ​​in the pixel-level mapping matrix. After sorting the values ​​from largest to smallest, the values ​​in the top 5% are used as thresholds. All pixels in the anomaly heatmap whose original response values ​​for the rendered color channel are greater than the red highlight threshold are extracted. A connected component labeling algorithm is applied to these pixels to find all interconnected extreme connected regions. For each extracted extreme connected region, the average coordinates of all its pixels are calculated to obtain the center coordinates of the region. The maximum and minimum coordinates of the region in the horizontal and vertical directions are found to form its boundary contour information. Based on the boundary contour information, the minimum bounding rectangle size parameter that can completely enclose the extreme connected region is calculated. On the anomaly heatmap, using the center coordinates of each extreme connected region as a reference, a prominent rectangular defect bounding box is drawn using its corresponding bounding rectangle size parameter. Based on all drawn rectangular defect bounding boxes, the defect locations are accurately labeled on the anomaly heatmap.

[0091] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for identifying defects in lithium electrode arrays based on deep learning, characterized in that, Includes the following steps: Acquire a set of three-dimensional multimodal images of the lithium electrode assembly to be tested; Extract the two-dimensional texture features and three-dimensional elevation features of each three-dimensional multimodal image in the three-dimensional multimodal image set; The two-dimensional texture features and the three-dimensional elevation features are fused according to the cross-attention mechanism to obtain cross-modal joint features; The cross-modal joint features are processed by a pre-defined unsupervised normalized flow network that incorporates the prior distribution of coating thickness to obtain feature mapping distribution data. Calculate the distribution offset of the feature mapping distribution data to obtain the anomaly score; The defect status of the lithium electrode assembly is determined based on the anomaly score and the preset anomaly determination threshold. When it is determined that the lithium electrode group is in the defective state, the feature mapping distribution data is reverse-mapped to obtain a pixel-level mapping matrix. An anomaly heatmap is generated based on the pixel-level mapping matrix, and the defect locations are marked on the anomaly heatmap.

2. The lithium electrode assembly defect identification method based on deep learning according to claim 1, characterized in that, The specific steps for acquiring a set of three-dimensional multimodal images of the lithium electrode assembly to be tested are as follows: Obtain the scanning frame rate parameter, and control the production line line laser profilometer to scan the three-dimensional contour of the lithium electrode assembly to be detected according to the scanning frame rate parameter to obtain the first scanning data; The production line laser profilometer is controlled to scan the three-dimensional contour of the edge of the lithium electrode assembly to be inspected, and second scan data is obtained. The first scan data and the second scan data are stitched together to obtain three-dimensional point cloud data; Obtain the nearest neighbor number parameter, and use the statistical outlier filtering algorithm to filter out noise points in the 3D point cloud data based on the nearest neighbor number parameter to obtain cleaned 3D point cloud data. Then, perform projection processing on the cleaned 3D point cloud data to obtain a 3D depth map. The exposure parameters are obtained, and the binocular vision camera is controlled to capture the surface texture of the lithium electrode assembly to be detected based on the exposure parameters to obtain a two-dimensional color image; A reflective mask is generated by extracting the highly reflective metallic areas from the two-dimensional color image. After removing the high-brightness interference information from the two-dimensional color image and the three-dimensional depth image using the reflective mask, the corner features of the two-dimensional color image and the edge features of the three-dimensional depth image are extracted respectively. The three-dimensional depth map and the two-dimensional color image are aligned based on the corner features and the edge features to obtain aligned multimodal image data; The aligned multimodal image data is continuously collected, and all the aligned multimodal image data are merged to obtain the three-dimensional multimodal image set.

3. The lithium electrode assembly defect identification method based on deep learning according to claim 1, characterized in that, The specific steps for extracting the two-dimensional texture features and three-dimensional elevation features of each three-dimensional multimodal image in the three-dimensional multimodal image set are as follows: Extract the channel-level data of each three-dimensional multimodal image in the three-dimensional multimodal image set, separate the channel-level data, and obtain the first visual channel data and the second depth channel data; A backbone visual converter network is constructed, the first visual channel data is divided into multiple image block data, the image block data is input into the backbone visual converter network, the self-attention mechanism of the backbone visual converter network is used to extract the texture distribution information of the image block data in the planar dimension to obtain the first texture feature vector, and the first texture feature vector is flattened to obtain the two-dimensional texture feature. A backbone 3D convolutional network with residual connection structure is constructed. The second depth channel data is input into the backbone 3D convolutional network. The spatial convolution kernel of the backbone 3D convolutional network is used to extract the geometric distribution information of the second depth channel data in the spatial height dimension to obtain the first elevation feature tensor. The first elevation feature tensor is downsampled using the downsampling layer of the backbone 3D convolutional network to obtain the 3D elevation feature.

4. The lithium electrode assembly defect identification method based on deep learning according to claim 1, characterized in that, The specific steps for fusing the two-dimensional texture features and the three-dimensional elevation features according to the cross-attention mechanism to obtain the cross-modal joint features are as follows: Extract the query vector matrix of the two-dimensional texture features, and extract the key vector matrix and value vector matrix of the three-dimensional elevation features; Obtain the feature dimension scaling factor parameter, perform a dot product operation on the transpose of the query vector matrix and the key vector matrix to obtain the initial association weight matrix, and divide the initial association weight matrix by the feature dimension scaling factor parameter to obtain the scaled association weight matrix; The scaling correlation weight matrix is ​​nonlinearly mapped using an exponential normalization function to obtain a normalized attention matrix; Calculate the spatial geometric gradient information of the three-dimensional elevation features, and generate a local elevation mask matrix based on the spatial geometric gradient information; After spatial weight modulation of the normalized attention matrix using the local elevation mask matrix, matrix multiplication is performed on the value vector matrix to obtain the three-dimensional activation features. The three-dimensional activation features and the two-dimensional texture features are concatenated to obtain multi-dimensional fused tensor data. The multi-dimensional fused tensor data is then input into a fully connected perceptron network for parameter dimensionality reduction to obtain the cross-modal joint features.

5. The lithium electrode assembly defect identification method based on deep learning according to claim 1, characterized in that, The specific steps for processing the cross-modal joint features using a pre-defined unsupervised normalized flow network that incorporates the prior distribution of coating thickness to obtain the feature mapping distribution data are as follows: Obtain a normal sample training dataset containing normal lithium electrode group samples, and use the normal sample training dataset to perform unsupervised training on the initial mapping network model to obtain the preset unsupervised normalized flow network that combines the prior distribution of coating thickness. The cross-modal joint features are input into the preset unsupervised normalized flow network that combines the prior distribution of coating thickness, and the scaling parameter matrix and translation offset parameter matrix in the preset unsupervised normalized flow network that combines the prior distribution of coating thickness are obtained. The cross-modal joint features are then subjected to a nonlinear invertible mapping transformation based on the scaling parameter matrix and the translation offset parameter matrix to obtain the forward mapping features. Obtain the Jacobian matrix of the preset unsupervised normalized flow network that combines the prior distribution of coating thickness, and calculate the determinant parameter value corresponding to the Jacobian matrix. Obtain pre-constructed reference thickness gradient data for lithium electrode assembly, and construct a spatial prior correction factor based on the reference thickness gradient data for lithium electrode assembly; Based on the determinant parameter value and the spatial prior correction factor, the log probability density of the forward mapping feature is evaluated and calculated to obtain the target feature vector distribution. Align the target feature vector distribution with the standard normal distribution, extract the distribution node data from the target feature vector distribution, and merge the distribution node data to obtain the feature mapping distribution data.

6. The lithium electrode assembly defect identification method based on deep learning according to claim 1, characterized in that, The specific steps for calculating the distribution offset of the feature mapping distribution data to obtain the anomaly score are as follows: Obtain standard high-dimensional normal distribution template data, extract the first mean parameter of the feature mapping distribution data and the second mean parameter of the standard high-dimensional normal distribution template data, calculate the absolute difference between the first mean parameter and the second mean parameter, and obtain the mean distribution difference; Extract the first variance parameter of the feature mapping distribution data and the second variance parameter of the standard high-dimensional normal distribution template data, calculate the deviation between the first variance parameter and the second variance parameter, and obtain the variance distribution difference; Obtain the joint evaluation function model, input the mean distribution difference and the variance distribution difference into the joint evaluation function model, evaluate the cumulative result of the mean distribution difference and the variance distribution difference, and obtain the distribution offset; Obtain the preset maximum tolerance distribution offset parameter and the preset minimum tolerance distribution offset parameter, and extract the relative position data of the distribution offset between the maximum tolerance distribution offset parameter and the minimum tolerance distribution offset parameter; The distribution offset is mapped to a continuous numerical range of scores based on the relative position data to obtain the abnormal score.

7. The lithium electrode assembly defect identification method based on deep learning according to claim 1, characterized in that, The specific steps for determining the defect status of the lithium electrode assembly based on the anomaly score and the preset anomaly determination threshold are as follows: Extract the first floating-point value corresponding to the abnormal score, obtain the second floating-point value corresponding to the preset abnormal judgment threshold, and compare the first floating-point value with the second floating-point value. When the first floating-point value is greater than the second floating-point value, it is determined that the lithium electrode assembly has a defect, and the lithium electrode assembly is determined to be in the defective state. When the first floating-point value is not greater than the second floating-point value, it is determined that the feature mapping distribution data is within the allowable fluctuation range, and the lithium electrode group is determined to be in a normal state.

8. The lithium electrode assembly defect identification method based on deep learning according to claim 1, characterized in that, When it is determined that the lithium electrode assembly is in the defective state, the specific steps for performing reverse mapping on the feature mapping distribution data to obtain a pixel-level mapping matrix are as follows: Obtain feature mapping distribution data of lithium electrode groups in defective state, and construct a defective state data set from the obtained feature mapping distribution data; Extract the set of parameters of the inverse computation neuron of the unsupervised normalized flow network with the prior distribution of coating thickness, and establish the inverse affine transformation function of the unsupervised normalized flow network with the prior distribution of coating thickness based on the set of parameters of the inverse computation neuron. The defect state data set is input into the inverse affine transformation function, and the inverse affine transformation function is controlled to perform a layer-by-layer reverse invertible mapping calculation on the defect state data set to obtain high-dimensional restored features in the two-dimensional spatial domain. Extract the feature response intensity value corresponding to each pixel position in the high-dimensional restored feature; Obtain the original image width and height parameters corresponding to the three-dimensional multimodal image set, and arrange all the feature response intensity values ​​in an array according to the original image width and height parameters to obtain the pixel-level mapping matrix.

9. The lithium electrode assembly defect identification method based on deep learning according to claim 1, characterized in that, The specific steps for generating an anomaly heatmap based on the pixel-level mapping matrix and marking the defect locations on the anomaly heatmap are as follows: Traverse and extract the response value of each mapping node in the pixel-level mapping matrix, establish a color gradient mapping reference dictionary, and look up the corresponding rendering color channel value in the color gradient mapping reference dictionary according to the response value. The two-dimensional plane of the pixel-level mapping matrix is ​​rendered and colored according to the rendering color channel values ​​to obtain the abnormal heatmap; Obtain the preset red highlight color threshold parameter, and extract the extreme connected regions in the abnormal heatmap where the rendered color channel value is greater than the red highlight color threshold parameter; Extract the center coordinates and boundary contour information of the extreme connected region, and calculate the size parameters of the bounding rectangle surrounding the extreme connected region based on the boundary contour information. In the abnormal heatmap, a rectangular defect bounding box is drawn in the local area corresponding to the center coordinate point location information using the circumscribed rectangle size parameter, and the defect location is marked on the abnormal heatmap based on the rectangular defect bounding box.

Images