CN-121982319-A - Fuel cell bipolar plate adhesive tape segmentation method, system, medium and device based on gradient attention and multi-scale cavity residual error structure

CN121982319ACN 121982319 ACN121982319 ACN 121982319ACN-121982319-A

Abstract

The invention relates to the technical field of fuel cell production, in particular to a method, a system, a medium and equipment for segmenting a bipolar plate adhesive tape of a fuel cell based on a gradient Attention and multi-scale cavity residual error structure, which adopt MobileNetV-U-Net fusion architecture, take light MobileNetV as an encoder, gradually extract high-level semantic features of images through a multi-layer inverse residual error block, simultaneously embed a DG Attention module at the bottommost layer in the encoding stage, lead Attention weight calculation through extracting gradient features and strip features, strengthen feature capture on the edge of the adhesive tape and the slender structure, and replace the traditional convolution operation with a Res2DilatedConv module while a decoder keeps multi-layer jump connection of U-Net, realize multi-scale feature extraction through channel grouping, multi-cavity rate convolution and cross-group feature fusion, and adapt structural differences of adhesive tapes under different widths and bending degrees so as to realize efficient and accurate segmentation of the adhesive tape region in the bipolar plate of the fuel cell.

Inventors

FU XIN
TANG LU
YANG SHANGDONG
Yan Shuzong
CHEN RUI
HAN YONG

Assignees

厦门理工学院

Dates

Publication Date: 20260505
Application Date: 20260130

Claims (10)

1. A fuel cell bipolar plate adhesive tape segmentation method based on gradient attention and multi-scale cavity residual error structure adopts a MobileNetV-U-Net fusion architecture segmentation model for segmentation, and is characterized by comprising the following steps: Collecting an original image of a fuel cell bipolar plate with an adhesive tape area subjected to segmentation marking, establishing an original data set, preprocessing the original data set, and dividing the original data set into a training set, a verification set and a test set; Inputting the training set into a segmentation model comprising an encoder, a decoder and two core function modules for training, and monitoring the performance of the segmentation model by using the verification set and dynamically adjusting the learning rate to obtain a trained segmentation model; The encoder adopts MobileNetV inverse residual blocks to replace the encoder part in the traditional U-Net as a feature extraction main body, gradually extracts high-level semantic features of images through multi-layer inverse residual blocks, embeds a DG Attention module at the bottommost layer in the encoding stage, guides Attention weight calculation through extracting gradient features and strip features, strengthens feature capture of adhesive tape edges and a strip structure, and replaces traditional convolution operation with a Res2DilatedConv module while retaining multi-layer jump connection of the U-Net, wherein the Res2DilatedConv module realizes multi-scale feature extraction through channel grouping, multi-void-rate convolution and cross-group feature fusion and adapts structural differences of adhesive tapes under different widths and bending degrees; the method comprises the steps of collecting an original image of an adhesive tape area in a fuel cell bipolar plate to be segmented, inputting a segmentation model for segmentation, and outputting an adhesive tape area segmentation result graph consistent with the original image in size, so that efficient and accurate segmentation of the adhesive tape area in the fuel cell bipolar plate is realized.
2. The method for segmenting the bipolar plate adhesive tape of the fuel cell based on the gradient attention and the multi-scale cavity residual error structure of claim 1, wherein the preprocessing of the original data set is data enhancement operation at least comprising random rotation, mirror image inversion, brightness adjustment and Gaussian noise disturbance.
3. The method for segmenting the bipolar plate adhesive tape of the fuel cell based on the gradient Attention and the multi-scale cavity residual error structure of claim 1, wherein the DG Attention module guides Attention weight calculation by extracting gradient features and strip features comprises the following steps: extracting local convolution characteristics, namely extracting low-level characteristics of an image through two-layer depth separable convolution and mapping the low-level characteristics into a high-dimensional characteristic image; Edge feature extraction, namely extracting edge information in an image by Sobel edge detection, convoluting a horizontal edge kernel of a Sobel operator and a vertical edge check image, and calculating the gradient of each pixel point in the horizontal direction And gradient in the vertical direction Calculating the gradient amplitude value, and generating an edge map; strip feature extraction based on horizontal gradient And a vertical gradient Calculating an elongated feature, the elongated feature is expressed as: Wherein, the Is a small constant for preventing the division by zero; Transformer attention fusion-introducing edge graphs and elongated features in the standard self-attention score to adjust the attention weights, the final attention weights are expressed as: Wherein, the Is a hyper-parameter that adjusts the influence of the boot graph, And The horizontal and vertical directions of the edge map, respectively; Carrying out normalization processing on the final attention weight to obtain a final attention weight matrix, wherein each element represents the attention degree of one spatial position of the feature map to another spatial position; And carrying out weighted summation on the value (V) through a final Attention weight matrix to obtain a high-level semantic feature map output by the DG Attention module.
4. The method for segmenting the bipolar plate adhesive tape of the fuel cell based on the gradient attention and the multi-scale cavity residual error structure as set forth in claim 3, wherein the horizontal kernel of the Sobel operator And a vertical core Expressed as: 。
5. The fuel cell bipolar plate adhesive tape segmentation method based on the gradient attention and multi-scale cavity residual error structure of claim 3, wherein the edge map is expressed as: 。
6. The method for segmenting the bipolar plate adhesive tape of the fuel cell based on the gradient attention and the multi-scale cavity residual error structure of claim 1, wherein the Res2DilatedConv module realizes multi-scale feature extraction through channel grouping, multi-cavity convolution and cross-group feature fusion, and the method comprises the following steps: The channel grouping comprises a first layer decoding stage, a subsequent decoding stage and a channel grouping stage, wherein the first layer decoding stage takes a high-layer semantic feature image output by the DG Attention module as an input feature image, and the output of the decoding stage above all serves as an input feature image; And (3) multi-cavity rate convolution, namely directly outputting the first subgroup feature images, and respectively applying cavity convolution with linearly increasing cavity rate to the rest subgroup feature images, wherein the cavity rate is expressed as follows: Wherein, the The void fraction is the basic void fraction; Group-crossing feature accumulation, namely, accumulating the first Convolved output of group subgroup feature map The output of the group subgroup feature graphs are added channel by channel to realize cross-scale information fusion; The method comprises the steps of feature fusion and residual error connection, wherein the output of all the subgroup feature images are spliced in the channel dimension, the channel number is adjusted through 1X 1 convolution, and the fusion feature images are formed by fusion, namely if the channel number of the input feature images is consistent with the channel number of the fusion feature images, the fusion feature images are directly subjected to residual error addition with the input feature images, and if the channel number of the input feature images is inconsistent with the channel number of the fusion feature images, the residual error addition is performed after the input channel number is adjusted through 1X 1 convolution, so that the enhanced feature images are obtained; Feature recovery and fusion, namely taking the enhanced feature map as input of a decoding stage, gradually recovering the spatial resolution in a step-by-step up-sampling mode, and introducing features of corresponding scales of an encoding stage in each stage of decoding process to fuse to obtain a final feature map; And carrying out pixel-level classification processing on the final feature image, and outputting a glue strip segmentation result image with the same size as the original image.
7. The method for segmenting the bipolar plate adhesive tape of the fuel cell based on the gradient attention and the multi-scale cavity residual error structure of claim 1, further comprising the step of evaluating the performance of the trained segmented model through a test set to measure the segmentation performance of the segmented model and obtain a final segmented model.
8. A fuel cell bipolar plate adhesive tape segmentation system based on a gradient attention and multi-scale cavity residual error structure is characterized by being used for realizing the fuel cell bipolar plate adhesive tape segmentation method based on the gradient attention and multi-scale cavity residual error structure as set forth in any one of claims 1-7.
9. A computer readable storage medium, wherein the computer readable storage medium stores computer instructions, and the computer is executed by a processor to implement the fuel cell bipolar plate adhesive tape segmentation method based on the gradient attention and multi-scale cavity residual structure according to any one of claims 1-7.
10. A computer device comprising at least one processor and a memory communicatively coupled to the processor, wherein the memory stores instructions executable by the at least one processor to cause the processor to perform the method of fuel cell bipolar plate adhesive strip segmentation based on gradient attention and multi-scale void residual structure of any one of claims 1-7.

Description

Fuel cell bipolar plate adhesive tape segmentation method, system, medium and device based on gradient attention and multi-scale cavity residual error structure Technical Field The invention relates to the technical field of fuel cell production, in particular to a fuel cell bipolar plate adhesive tape segmentation method, a system, a medium and equipment based on gradient attention and a multi-scale cavity residual error structure. Background Bipolar plates (also known as separators) are one of the important components of fuel cells and their functions include providing gas flow channels, preventing hydrogen and oxygen in the cell's gas chambers from crossing, and establishing a current path between the anode and cathode in series, etc. In the production process of the fuel cell, the accurate segmentation of the adhesive tape area in the bipolar plate is important for the subsequent air tightness detection, and the packaging quality and the service life of the fuel cell are directly affected. The existing traditional image processing algorithm such as threshold segmentation and edge detection is often insufficient in precision and poor in robustness when facing the problems of complex texture, non-uniform illumination, blurred edges of adhesive tapes and the like of a bipolar plate, and is difficult to meet the high-precision requirement of industrial production. Therefore, how to realize the efficient and accurate segmentation of the adhesive tape region in the bipolar plate of the fuel cell is a technical problem to be solved in the field. It should be noted that the information disclosed in this background section is only for the purpose of increasing the understanding of the general background of the invention and should not be taken as an acknowledgement or any form of suggestion that this information forms the prior art already known to a person of ordinary skill in the art. Disclosure of Invention In order to solve the technical problems of the conventional image processing algorithm, on the one hand, the invention provides a fuel cell bipolar plate adhesive tape segmentation method based on a gradient attention and multi-scale cavity residual error structure, which adopts a MobileNetV-U-Net fusion architecture segmentation model to segment, and comprises the following steps: Collecting an original image of a fuel cell bipolar plate with an adhesive tape area subjected to segmentation marking, establishing an original data set, preprocessing the original data set, and dividing the original data set into a training set, a verification set and a test set; Inputting the training set into a segmentation model comprising an encoder, a decoder and two core function modules for training, and monitoring the performance of the segmentation model by using the verification set and dynamically adjusting the learning rate to obtain a trained segmentation model; The encoder adopts MobileNetV inverse residual blocks to replace the encoder part in the traditional U-Net as a feature extraction main body, gradually extracts high-level semantic features of images through multi-layer inverse residual blocks, embeds a DG Attention module at the bottommost layer in the encoding stage, guides Attention weight calculation through extracting gradient features and strip features, strengthens feature capture of adhesive tape edges and a strip structure, and replaces traditional convolution operation with a Res2DilatedConv module while retaining multi-layer jump connection of the U-Net, wherein the Res2DilatedConv module realizes multi-scale feature extraction through channel grouping, multi-void-rate convolution and cross-group feature fusion and adapts structural differences of adhesive tapes under different widths and bending degrees; the method comprises the steps of collecting an original image of an adhesive tape area in a fuel cell bipolar plate to be segmented, inputting a segmentation model for segmentation, and outputting an adhesive tape area segmentation result graph consistent with the original image in size, so that efficient and accurate segmentation of the adhesive tape area in the fuel cell bipolar plate is realized. On the other hand, the invention also provides a fuel cell bipolar plate adhesive tape segmentation system based on the gradient attention and the multi-scale cavity residual error structure, which is used for realizing the fuel cell bipolar plate adhesive tape segmentation method based on the gradient attention and the multi-scale cavity residual error structure. In a third aspect, the present invention further provides a computer readable storage medium, where computer instructions are stored, where the computer, when executed by a processor, implements the above-mentioned fuel cell bipolar plate adhesive tape segmentation method based on gradient attention and multi-scale cavity residual error structure. In a fourth aspect, the present invention further provides a computer device, including at