Search

CN-122023311-A - Pole piece endpoint detection method based on state space optimization

CN122023311ACN 122023311 ACN122023311 ACN 122023311ACN-122023311-A

Abstract

The invention provides a pole piece endpoint detection method based on state space optimization, and belongs to the technical field of artificial intelligence and computer vision. The invention constructs an end-to-end deep neural network detection model. According to the method, the pole piece structure information is extracted through multi-scale feature coding, a feature modeling mechanism based on a state space is introduced to enhance the distinguishing capability of pole piece endpoints and the background, and the detection result of a dense area is adaptively optimized by combining a reordering strategy of point level prediction and density perception, so that the accurate and stable positioning of the pole piece endpoints is realized. The invention can solve the problems of low manual detection efficiency and poor accuracy caused by the fact that the number of pole piece endpoints in an X-ray image of the power battery for the vehicle is large, the distribution is dense, the contrast is low, noise and artifact are easy to interfere, and can be applied to quality detection in the production process of the power battery, and the detection efficiency and the result consistency are effectively improved.

Inventors

  • ZHANG LIHE
  • Cao Peiqian

Assignees

  • 大连理工大学

Dates

Publication Date
20260512
Application Date
20260126

Claims (9)

  1. 1. The pole piece endpoint detection method based on state space optimization is characterized by adopting an encoder-decoder type end-to-end deep learning network structure, wherein an integral network framework comprises a feature encoder, a prompt pipeline, a decoder, a point predictor and a point segmentation optimizer, an industrial X-ray image is taken as an input image, and automatic positioning of pole piece endpoints is realized by carrying out point segmentation prediction on anode pole piece endpoints and cathode pole piece endpoints in the input image, and the method specifically comprises the following steps: step 1, constructing a feature encoder which consists of a layer of two-dimensional convolution and a residual type two-dimensional convolution neural network; The feature encoder consists of a layer of two-dimensional convolution and a residual two-dimensional convolution neural network, and obtains multi-scale features with different spatial resolutions and semantic layers through feature extraction of an input image ; Step 2, constructing a prompt pipeline based on the residual two-dimensional convolutional neural network in the feature encoder in step 1, and carrying out layer-by-layer feature encoding on the prompt image through the prompt pipeline to obtain a plurality of groups of prompt features with different spatial resolutions and semantic layers , The feature will provide guidance information for pole piece endpoint-related features during the decoding stage; Step 3, constructing a decoder, wherein the decoder is formed by connecting 4 characteristic decoding layers in series according to the sequence from deep to shallow, and is used for recovering the spatial resolution of the characteristics step by step, strengthening the relevant characteristics of the pole piece end points in the decoding process, and inhibiting background noise and imaging interference; Before the decoding process starts, the deepest layer coding feature output by the feature encoder is firstly coded Mapping the two-dimensional convolution of a layer into decoding characteristics to obtain initial decoding characteristics The method is characterized in that the method comprises the steps of performing a step-by-step upsampling and cross-layer feature fusion on the basis of the follow-up decoding, continuously improving the spatial resolution of the decoding features, and providing accurate spatial information support for endpoint detection, wherein a prompt filtering state space module PFSSM is embedded in each feature decoding layer in the decoding process; Step 4, constructing a point predictor, namely after the characteristic of the decoder stage is recovered step by step, constructing the point predictor for generating a rough point segmentation result of the pole piece end points so as to realize the preliminary positioning of the pole piece end point spatial distribution; Step 5, constructing a point segmentation optimizer; the point segmentation optimizer adopts a density perception reordering state space module which uses the fusion characteristics output in the step 4 Coarse point segmentation results generated by a point predictor Based on coarse-point segmentation results for input Reordering pixel positions and generating index mapping and inverse mapping relation between original positions and new positions, and fusing features by semantic guidance of index mapping and global modeling capability of 2D selective scanning layer After being processed by the density sensing reordering state space module, the dual targets of end point region semantic consistency enhancement and background interference suppression are realized, and finally, the finishing point segmentation result is output ; Step 6, constructing a total loss function; The method comprises the steps of adopting a mode of combining weighted cross-over ratio loss and binary cross entropy loss as a point segmentation supervision signal, wherein the weighted cross-over ratio loss is used for restraining the overall consistency of an endpoint region, and the binary cross entropy loss is used for restraining a pixel classification result; Step 7, network training; Based on PyTorch deep learning framework, a multi-card parallel training mode is adopted, industrial X-ray images and corresponding point labels are used as input in the training process, network parameters are jointly optimized according to the total loss function constructed in the step 6, and accurate positioning of pole piece endpoints in the industrial X-ray images is achieved.
  2. 2. The state space optimization-based pole piece endpoint detection method according to claim 1, wherein the step 1 is specifically: step 1.1, shallow feature mapping is carried out on an input image through one-layer two-dimensional convolution operation, shallow space information of the input image is automatically learned, the shallow space information comprises basic gray level distribution, edge contour and texture change, and 0 th layer coding feature consistent with the space size of the input image is obtained ; Step 1.2, inputting an input image into a residual type two-dimensional convolutional neural network for deep feature extraction, wherein the residual type two-dimensional convolutional neural network constructs 5 feature extraction layers along the depth direction, realizes progressive abstraction of feature representation from local detail to global structural feature while guaranteeing stable gradient propagation, and finally obtains a plurality of groups of multi-scale coding features with different spatial scales and semantic levels , , Representing the feature extraction hierarchy of a residual two-dimensional convolutional neural network, When representing layer 1 coding features , When representing layer 2 coding features , When representing layer 3 coding features , When representing layer 4 coding features , When representing layer 5 coding features 。
  3. 3. The pole piece endpoint detection method based on state space optimization according to claim 2, wherein in the step 1.2, the feature extraction layer is formed by cascading a plurality of residual modules, and feature receptive fields are expanded step by step in a mode of combining multi-layer convolution operation and identity mapping, so that the expression capability of complex structures and context semantic information is enhanced.
  4. 4. A method for detecting a pole piece endpoint based on state space optimization according to claim 3, wherein in the step 2: based on the structure of the residual two-dimensional convolutional neural network, the first 4 feature extraction layers are cut off and reserved, so that a prompt feature extraction pipeline, namely a prompt pipeline, is constructed, network parameters of the prompt pipeline and a feature encoder are completely independent, and the method comprises the following specific steps: The network parameters of the hint pipeline, i.e. the hint image, are selected from the industrial X-ray image dataset, and the hint image is consistent with the input image received by the feature encoder in step 1 in terms of imaging type and structural form, for providing a reliable structural prior for pole piece endpoint detection tasks.
  5. 5. The state space optimization-based pole piece endpoint detection method according to claim 4, wherein the step 3 is specifically: Step 3.1, in each feature decoding layer, firstly, carrying out up-sampling operation on decoding features output by a previous feature decoding layer to enable spatial resolution of the decoding features to be consistent with coding features corresponding to a current feature decoding layer, then carrying out element-by-element addition fusion on the up-sampled decoding features and the coding features so as to simultaneously reserve high-layer semantic information and low-layer spatial structure information, and taking the fused features as input features of a prompt filtering state space PFSSM, and realizing prompt guide filtering and global context modeling through the prompt filtering state space to effectively distinguish pole piece endpoint features from background and imaging interference, wherein the filtering state space is formed by sequentially connecting a linear layer, a prompt filter, a SiLU activation layer, a 2D selective scanning layer, an LN normalization layer and an output linear layer in series; step 3.2, the input features obtained in the step 3.1 are firstly subjected to feature dimension adjustment through a linear layer, and the input features are sent to a prompt filter; Step 3.3, in the prompt filter, firstly selecting the prompt features matched with the spatial resolution of the current feature decoding layer, and executing global average pooling operation on the prompt features to generate global prompt vectors; finally, carrying out channel-by-channel fusion on the channel-level weight and a two-dimensional convolution kernel with trainable parameters to generate a dynamic convolution weight, and executing prompt guide filtering operation on the current decoding characteristics by using the weight; step 3.4, inputting the features output after the step 3.3 is performed with the prompt filter into a SiLU activation layer and a 2D selective scanning layer in sequence, wherein the 2D selective scanning layer completes the state update modeling of the features by executing selective scanning operation along a plurality of directions in a two-dimensional space range; In summary, the feature update relationship in the feature decoding layer is expressed as: (1); Wherein, the Representing upsampling; Representing decoding characteristics therein , Representing highest resolution decoding features Spatial resolution and encoder shallow features In accordance with the method, the device and the system, Representing the next highest resolution decoding feature Spatially resolved into One half of the number (a) of the number (b), Representing mid-resolution decoding features Spatially resolved into One-fourth of the number of (a), Representing low resolution decoding features The spatial resolution is One eighth of (2); Representing the previous layer decoding characteristics.
  6. 6. The state space optimization-based pole piece endpoint detection method according to claim 5, wherein the step 4 is specifically: First, selecting the highest resolution decoding feature output by the decoder in step 3 Up-sampling is performed on the encoded signal to enable the spatial resolution of the encoded signal to be matched with the layer 0 encoded feature output by the feature encoder Maintain consistency and then Upsampled features Adding and fusing, introducing shallow space information and semantic information contained in high-level decoding features simultaneously in a cross-layer fusing operation mode, and obtaining fused features through one-layer two-dimensional convolution ; For fusion features Further applying 2 layers of two-dimensional convolution operation to finish feature channel mapping and space detail refinement, and finally outputting a coarse point segmentation result of the pole piece end points The expression is: (2); wherein, H and W respectively represent the space height and width of the output characteristic diagram, 2 represents the channel number and respectively corresponds to the space distribution of the end points of the anode pole piece and the end points of the cathode pole piece; representing coarse point segmentation results Is of a size H W, a real number domain feature map with the channel number of 2, wherein each channel pixel value is used for representing the confidence coefficient of the corresponding spatial position as the end point of the target pole piece.
  7. 7. The state space optimization-based pole piece endpoint detection method according to claim 6, wherein the step 5 is specifically: The module structure of the point segmentation optimizer consists of a linear layer, a depth separable convolution layer, a SiLU activation layer, an index mapping unit, a 2D selective scanning layer, an inverse mapping unit, an LN normalization layer and an output linear layer; Step 5.1, merging the input fusion features The characteristic dimension adjustment is completed by sending the characteristic dimension adjustment into a linear layer, and then the initial optimized characteristic is obtained by sequentially extracting the space characteristic of the depth separable convolution layer and carrying out nonlinear mapping of a SiLU activation layer; Step 5.2, based on the coarse point segmentation result For each pixel position Assigning semantic category labels The value expression is as follows: (3); Wherein, the The width-direction coordinates are indicated, ; The length-direction coordinates are indicated, 0 Represents an anode region, 1 represents a background region, and 2 represents a cathode region; According to the rule of 'same semantic category index priority aggregation', reordering pixel positions, and establishing a forward mapping relation between an original position and a reordered new position Simultaneously constructing inverse mapping relation from new position to original pixel position ; Step 5.3, flattening the initially optimized features output in step 5.1 into a one-dimensional feature sequence according to the space dimension Forward mapping relation based on semantic category Performing a reordering operation on the feature sequence X to obtain reordered features The operation expression is as follows: ; Wherein, the Representing an index mapping operation; Step 5.4, reordering features Inputting the data to a 2D selective scanning layer, executing feature scanning and state updating in a two-dimensional space range, establishing a long-distance dependency relationship with stronger semantic consistency, and outputting the modeled features ; Step 5.5, based on the reverse mapping relationship generated in step 5.2 For modeled features Performing inverse mapping operation to restore the features to the original two-dimensional spatial layout to obtain refined features The operation expression is as follows: (5); In the formula, Representing an inverse mapping operation; step 5.6, for the recovered refined features Sequentially performing LN normalization and dimension transformation of an output linear layer, and outputting a refined point segmentation result after refinement through a layer of convolution layer The expression is: (6); wherein, H and W respectively represent the space height and width of the output characteristic diagram, 2 represents the channel number and respectively corresponds to the space distribution of the end points of the anode pole piece and the end points of the cathode pole piece; representing the coarse point segmentation result Is of a size H W, a real number domain feature map with the channel number of 2, wherein each channel pixel value is used for representing the confidence coefficient of the corresponding spatial position as the end point of the target pole piece.
  8. 8. The state space optimization-based pole piece endpoint detection method according to claim 7, wherein the step 6 is specifically: Coarse-point segmentation results for point predictor output Finishing point segmentation result output by the point segmentation optimizer in step 5 Respectively calculating corresponding point segmentation losses, and defining the losses based on the rough point segmentation result as The loss based on the refinement point segmentation result is Both adopt the point segmentation loss calculation mode with the same structure, wherein the point segmentation loss is calculated by the point segmentation loss calculation method The method is obtained by linearly combining the weighted cross ratio loss and the binary cross entropy loss, and the calculation mode is as follows: (7); Wherein, the Representing a weighted overlap ratio loss for reinforcing a partition overlap constraint of the endpoint region; Representing binary cross entropy loss for precisely optimizing classification boundaries of endpoint and non-endpoint regions; the weighted cross-ratio loss The calculation mode of (a) is as follows: (8); Wherein, the Representing the total number of pixels in the image; represented at pixel locations Prediction results of the position; represented at pixel locations Labeling the positions; Is a pixel weight coefficient; The binary cross entropy loss The calculation mode of (a) is as follows: (9); Wherein, the Representing the total number of pixels in the image; represented at pixel locations Prediction results of the position; represented at pixel locations Labeling the positions; Loss to be obtained based on rough point segmentation result And loss based on refinement point segmentation results Weighted summation to construct a total loss function for training an encoder-decoder type end-to-end deep learning network The double-stage training constraint of early positioning guide and later precision reinforcement is realized, and the calculation mode is as follows: (10); Wherein, the Representing a total loss function; Representing a refinement point segmentation loss weight coefficient; Representing a coarse point segmentation loss weight coefficient; Representing the point segmentation loss calculated based on the refined point segmentation result output by the point segmentation optimizer; the point segmentation loss calculated based on the rough point segmentation result output by the point predictor is represented.
  9. 9. The state space optimization-based pole piece endpoint detection method according to claim 8, wherein the step 7 is specifically: the network training is realized based on PyTorch deep learning frames, and training is performed on 4 blocks NVIDIA TESLA V100 graphic processing units; training updates network parameters using Adam optimizer, wherein first moment estimates parameters Set to 0.5, second moment estimation parameter Set to 0.999, initial learning rate set to Introducing weight attenuation in training process, the weight attenuation coefficient is set as And in the aspect of training strategies, a staged learning rate attenuation mechanism is adopted to adjust the learning rate.

Description

Pole piece endpoint detection method based on state space optimization Technical Field The invention belongs to the technical field of artificial intelligence and computer vision, relates to a pole piece endpoint detection method based on state space optimization, in particular to an industrial X-ray image analysis method based on deep learning, and provides a pole piece endpoint detection method based on state space optimization, which is suitable for industrial application scenes such as quality detection and safety evaluation of a vehicle power battery in a manufacturing process. Background With the rapid development of the electric automobile industry, the power type battery for vehicles is widely focused on the manufacturing quality as a core energy unit for determining the performance and safety of the whole vehicle. The interior of the battery cell is formed by alternately stacking a plurality of anode plates and cathode plates, and the arrangement sequence, the end point positions and the overall consistency of the electrode plates directly influence the electrochemical performance, the cycle life and the use safety of the battery. In the process of assembling the battery cell, the internal pole piece is influenced by factors such as processing precision, stacking consistency, assembly process and the like, and the internal pole piece can have problems such as dislocation, protrusion or abnormal arrangement, so that serious potential safety hazards such as internal short circuit, abnormal heating and even thermal runaway are caused. Therefore, after the battery cell is packaged, the high-reliability and high-precision nondestructive detection is carried out on the internal pole piece structure, and the method has important significance for guaranteeing the manufacturing quality of power battery products. Currently, in industrial production, a nondestructive detection mode based on X-ray imaging is generally adopted to evaluate the quality of the battery monomer. According to the method, the space positions of the endpoints of the positive electrode plate and the negative electrode plate in the X-ray image are analyzed, and the number of the electrode plates and whether the abnormality such as protrusion or deletion exists or not are further calculated, so that the automatic detection of the internal structural integrity of the battery cell is realized. However, the related X-ray images generally have remarkable structural complexity, namely, on one hand, the pole piece end points are tiny in size and low in gray contrast in the images and are easily interfered by non-target structures such as diaphragms, trays and the like to cause weaker end point feature expression, and on the other hand, the pole pieces are large in number and dense in arrangement, the adjacent pole pieces are smaller in distance, the local structures are similar in height, and feature confusion and positioning deviation are easily caused. In actual production, part of enterprises still rely on a manual visual interpretation mode to detect, so that the efficiency is low, the detection is easily influenced by experience differences and visual fatigue of operators, and quality consistency and traceability requirements in high-beat and mass manufacturing scenes are difficult to meet. Although research has been attempted to introduce computer vision technology to realize automatic detection, traditional image processing means (such as edge detection, threshold segmentation, morphological operation and the like) are sensitive to imaging quality and have insufficient robustness, and general target detection or semantic segmentation models are not fully fused with pole piece dense stacking structural features and arrangement sequence semantic priori, so that stable and accurate pole piece endpoint positioning is difficult to realize under the conditions of complex background, micro-spacing and multi-scale imaging. In summary, the prior art has obvious defects in the aspects of weak feature perceptibility, dense structure resolution capability, structural semantic modeling and the like, and is difficult to meet the actual demands of the vehicle power battery on the pole piece endpoint detection in the aspects of precision and high reliability. Therefore, it is needed to propose an industrial X-ray imaging pole piece endpoint analysis method which can fuse pole piece structure priori, enhance weak endpoint feature expression and realize high-precision and stable detection under complex interference and variable imaging conditions. Disclosure of Invention The technical problem to be solved by the invention is that under the conditions of weak pole piece endpoint characteristics, dense pole piece stacking and complex imaging interference, the conventional industrial X-ray image detection method is difficult to realize uniform modeling of fine local endpoint positioning and global structure consistency, and the problems of endpoint confusion