Three-dimensional reconstruction method of gas diffusion layer in proton exchange membrane fuel cell

Abstract

The present invention discloses a three-dimensional reconstruction method for a gas diffusion layer of a proton exchange membrane fuel cell. The method uses X-ray computed tomography technology to obtain multiple two-dimensional slice images of the gas diffusion layer, uses a K-means clustering algorithm to pre-process the image to separate the voids, uses the carbon paper porosity as a judgment condition, and gradually obtains the correct pre-processed CT image by changing the cluster center system; uses SegFormer semantic segmentation to perform feature extraction, captures the pores, carbon fibers and adhesives in the pre-processed CT image obtained in step 1 and assigns different gray values ​​to generate a phase segmentation slice image; uses three-dimensional reconstruction to stack all phase segmentation slice images to increase the dimension to obtain a three-dimensional gas diffusion layer structure; the present invention can accurately identify the components of a finished gas diffusion layer, obtain component information, adjust the product production material ratio based on this information combined with performance testing, and guide gas diffusion layer and other porous media related product research and development companies to better plan process flows and manufacturing plans.

Description

Technical Field

[0001] The invention belongs to the technical field of fuel cells, and in particular relates to a three-dimensional reconstruction method for a gas diffusion layer of a proton exchange membrane fuel cell. Background Art

[0002] At present, in the field of gas diffusion layer research for proton exchange membrane fuel cells, there is no method to accurately distinguish the components of the gas diffusion layer, reconstruct the model, and obtain the proportion of each component. People can only roughly calculate the proportion of adhesive and fiber in carbon paper through thermogravimetric analysis during the experimental preparation process, but the data obtained by thermogravimetric analysis is only for a certain type of base paper and process conditions, there are certain errors, and it cannot be widely transplanted. In addition, there are also methods that use the apparent scanning electron microscope (SEM) image of GDL to manually calculate the proportion of each component, which is time-consuming and labor-intensive, and because SEM only represents the appearance of GDL, the statistical component proportion cannot represent the entire GDL model, so the statistical results have large errors. These two methods of obtaining the proportion of GDL components cannot obtain the microscopic model of GDL, and cannot use microscopic pore scale analysis to provide scientific optimization suggestions for GDL performance.

[0003] At present, there are two main methods for obtaining the three-dimensional microstructure of GDL, the method combining programming reconstruction with morphological image processing and X-ray computed tomography (XCT). Programming reconstruction and morphological processing can obtain the multi-component microstructure of GDL, but it is necessary to judge the model based on the known component ratio, which has certain value for GDL preparation. However, programming reconstruction and morphological processing have the following disadvantages: the reconstructed fiber skeleton does not take into account the bending and warping of the fiber, it is difficult to strictly control the various factors that affect the reconstruction results, and it cannot accurately reflect the characteristics of the fiber skeleton. There is a certain difference between the adhesive added in the morphological processing and the adhesive of the real shape, which cannot truly reflect the microstructure of GDL.

[0004] The XCT method can obtain the structure of GDL, but XCT has the following disadvantages: XCT cannot distinguish the carbon fiber and adhesive in the carbon paper through threshold segmentation, and it is difficult to accurately evaluate the performance of carbon paper. The XCT threshold segmentation process is difficult to accurately control and is prone to noise, resulting in inaccurate carbon paper models. Due to the material properties of GDL, it is impossible to distinguish the fibers and adhesives in GDL, and it is impossible to accurately perform structural and performance analysis.

[0005] At present, there is no technology that can distinguish the distribution of GDL microscopic components, and it is impossible to obtain the true three-dimensional microstructure. The current status of scientific research on gas diffusion layers at home and abroad is based on simplified structures obtained by programming or XCT. Programming reconstruction does not truly consider factors such as fiber orientation, local porosity, fiber diameter, adhesive shape, distribution characteristics, etc. XCT reconstruction cannot distinguish between fibers and adhesives, and it is difficult to match the simulation and performance test results with experiments, and it is impossible to accurately find the optimization plan for carbon paper preparation. Existing technologies cannot truly reconstruct the gas diffusion layer model of proton exchange membrane fuel cells, and cannot accurately guide experimental preparation and performance testing. This is a technical problem that needs to be solved urgently. Summary of the invention

[0006] The purpose of the present invention is to provide a three-dimensional reconstruction method for the gas diffusion layer of a proton exchange membrane fuel cell. Through artificial intelligence (AI) technology, CT slice images are subjected to image recognition, the position of each component region is marked, and then three-dimensional reconstruction is performed to accurately distinguish the components of the gas diffusion layer in three-dimensional space and obtain the real structure.

[0007] In order to achieve the above purpose, the technical solutions adopted are as follows:

[0008] A three-dimensional reconstruction method for a gas diffusion layer of a proton exchange membrane fuel cell comprises the following steps:

[0009] (1) Using X-ray computed tomography technology to obtain multiple two-dimensional slices of the gas diffusion layer (GDL), the images were preprocessed using the K-means clustering algorithm to separate the voids, and the carbon paper porosity was used as the judgment condition. The correct preprocessed CT images were gradually obtained by changing the cluster center system;

[0010] (2) Feature extraction is performed through SegFormer semantic segmentation to capture the pores, carbon fibers, and adhesives in the preprocessed CT image obtained in step 1 and assign different grayscale values ​​to generate phase segmentation slices;

[0011] (3) Using three-dimensional reconstruction, all phase segmentation slices are stacked to increase the dimension and obtain a three-dimensional gas diffusion layer structure.

[0012] According to the above scheme, in step 2, in feature extraction, the training set is the preprocessed CT image, the black area is the pore, the straight rod-shaped objects are carbon fibers, and other irregular shapes are adhesives.

[0013] According to the above scheme, step 2 uses pixel accuracy PA as an evaluation indicator in feature extraction; the pixel accuracy is the ratio of the number of pixels with the correct prediction category to the total number of pixels, and the specific formula is as follows:

[0014] PA=(TP+TN) / (TP+TN+FP+FN)

[0015] Among them, TP positive class is judged as positive class, FP negative class is judged as positive class, FN positive class is judged as negative class, and TN negative class is judged as negative class.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] Based on the differences in the structural characteristics of the gas diffusion layer components, the true three-dimensional structure of the gas diffusion layer can be established by accurately identifying and summarizing the growth laws of the fuel cell gas diffusion layer components, which is beneficial to the subsequent structural performance analysis.

[0018] The present invention can accurately identify the components of the finished gas diffusion layer, obtain component information, adjust the material ratio of product production based on this information combined with performance testing, and guide gas diffusion layer and other porous media related product research and development companies to better plan process flow and manufacturing solutions.

[0019] The present invention provides accurate gas diffusion layer microstructure parameters for scientific research, such as local porosity distribution, carbon fiber diameter, carbon fiber orientation probability distribution, adhesive ratio and its distribution law, which can help improve the accuracy of the reconstructed model and make the simulation research closer to the real situation.

[0020] The present invention provides a method for scientific research to accurately reconstruct the real microstructure of materials, which is beneficial for carrying out various studies using the obtained models, and for a deeper understanding of the microscopic mechanism of the heat and mass transfer mechanical behavior of materials, which is conducive to accelerating material development.

[0021] The present invention lays a foundation for characterizing the microstructure of the gas diffusion layer and predicting the performance of the gas diffusion layer, and provides a scientific approach.

[0022] The present invention can be extended from fuel cell gas diffusion layers to other carbon papers and porous media fields, and can provide guidance for solving other material manufacturing and development problems, and has great scientific significance and engineering value. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 : Schematic diagram of the process of the present invention.

[0024] Figure 2 :The two-dimensional slice image of the gas diffusion layer was obtained by X-ray computed tomography.

[0025] Figure 3 : K-means clustering algorithm preprocessing CT images.

[0026] Figure 4 : Phase segmentation slices with feature extraction via SegFormer semantic segmentation.

[0027] Figure 5: Schematic diagram of the gas diffusion layer structure obtained using three-dimensional reconstruction. DETAILED DESCRIPTION

[0028] The following examples further illustrate the technical solution of the present invention, but are not intended to limit the scope of protection of the present invention.

[0029] The specific embodiment provides a three-dimensional reconstruction method of a gas diffusion layer of a proton exchange membrane fuel cell. Figure 1 As shown:

[0030] (1) Using X-ray computed tomography technology to obtain multiple two-dimensional slices of the gas diffusion layer (GDL), the images were preprocessed using the K-means clustering algorithm to separate the voids, and the carbon paper porosity was used as the judgment condition. The correct preprocessed CT images were gradually obtained by changing the cluster center system;

[0031] (2) Feature extraction is performed through SegFormer semantic segmentation to capture the pores, carbon fibers, and adhesives in the preprocessed CT image obtained in step 1 and assign different grayscale values ​​to generate phase segmentation slices;

[0032] Specifically, in the AI ​​algorithm feature extraction, the training set is a two-dimensional CT slice image. Labelme is used to mark the black area as pores, the straight rod-shaped objects as carbon fibers, and other irregular shapes as adhesives. In the gas diffusion layer, the carbon fibers are staggered and present a certain layered distribution in the thickness direction. Due to the wettability of the phenolic resin liquid during the manufacturing process, the adhesive appears at the overlap of the two carbon fibers to fill the holes and has a good distinction from the carbon fibers. By using the SegFormer semantic segmentation algorithm for training, pixel accuracy (PA) is used as an evaluation indicator in the feature extraction process; the pixel accuracy is the ratio of the number of pixels with the correct predicted category to the total number of pixels. The specific formula is as follows:

[0033] PA=(TP+TN) / (TP+TN+FP+FN)

[0034] Among them, TP positive class is judged as positive class, FP negative class is judged as positive class, FN positive class is judged as negative class, and TN negative class is judged as negative class. The specific meaning is: positive and negative classes represent fibers and adhesives, and the fiber / adhesive pixel results judged by the algorithm are consistent with the annotations. Figure 1 The total number of pixels is TP+TN, and the overall number of pixels is TP+TN+FP+FN. The accuracy of pixel judgment can be obtained by calculating PA through the formula.

[0035] (3) Using three-dimensional reconstruction, all phase segmentation slices are stacked to increase the dimension and obtain a three-dimensional gas diffusion layer structure.

[0036] 1. XCT technology tomography

[0037] XCT technology obtains a series of two-dimensional slice images of a specific area by scanning a rotating object, so as to clearly and accurately observe the internal microstructure of an object without cutting the object. It is a non-invasive and non-invasive imaging method. The basic principle of XCT technology is based on the attenuation and absorption characteristics of radiation in the object being detected. When a beam of radiation with a certain energy passes through the object being detected, the transmission energy received by the detector is different according to the different attenuation coefficients of each volume element in each transmission direction. According to a certain image reconstruction algorithm, a thin layer of tomographic image without image overlap at the interface of the object being detected can be obtained. Repeating the above process can obtain a new tomographic image. When enough two-dimensional tomographic images are measured, the two-dimensional images can be further stacked to synthesize the three-dimensional structure of the sample. In the specific implementation method, the two-dimensional slice image of the gas diffusion layer obtained is a grayscale image such as Figure 2 shown.

[0038] 2. K-means clustering algorithm image segmentation

[0039] For a given sample set, the sample set is divided into K clusters according to the distance between the samples. The points within the cluster are connected as closely as possible, and the distance between clusters is as large as possible. In the present invention, the K-means clustering algorithm is used to divide the CT grayscale image. First, by reading the image, a two-dimensional matrix m×n of CT image grayscale values ​​is obtained, and the two-dimensional matrix is ​​converted into a one-dimensional matrix 1×(m×n); then two center values ​​are taken in the one-dimensional matrix (used to distinguish the pores and solid materials in the CT image) as the initial clustering points of the K-means algorithm, and each point in the one-dimensional array is traversed to calculate the distance from the center point and divide each point. For the two categories after clustering, the center point is recalculated and the previous step is repeated until the two center points no longer change (or the new and old distances float within 2). Finally, the average value of the two center points is calculated to determine the image grayscale threshold, and the points around the threshold are divided (the grayscale of points less than the threshold is marked as 0, and the grayscale of points greater than the threshold is marked as 255). The preprocessed CT image of the white porous phase and the black solid phase is thus distinguished as follows: Figure 3 shown.

[0040] 3. SegFormer feature extraction

[0041] For a given image, it is first divided into multiple small blocks and then input into the hierarchical Transformer encoder to obtain multi-level features of (1 / 4, 1 / 8, 1 / 16, 1 / 32) of the original image resolution. For a small block, the block merging process used in VisionTransformer (ViT) unifies an N×N×3 block into a 1×1×C vector. This can be easily extended to unify a 2×2×C i ; The feature paths are unified into a 1×1×C i+1vector to obtain a hierarchical feature map. Using this method, the hierarchical features can be transformed from Zoom out to Then any other feature maps in the hierarchy are selected to combine non-overlapping images or feature blocks. A reduction ratio is used in the design of the self-attention layer to reduce the length of the sequence and reduce time complexity. Mix-FFN is used to replace the position encoding system. 3×3 convolutions and MLPs are mixed into each FFN, and deep convolutions are used to reduce the number of parameters and improve efficiency. Finally, four steps are performed in the MLP decoder: (1) The multi-level features from the MiT encoder are passed through the MLP layer to unify the channel dimension; (2) The features are upsampled to 1 / 4 and concatenated together; (3) The cascaded features F are fused using the MLP layer; (4) Another MLP layer uses the fused H / 4×W / 4×N_{cls} resolution features to predict the segmentation mask M, where N_{cls} represents the number of categories. The number of fiber and adhesive categories is thus obtained, and the training accuracy is calculated according to the pixel accuracy calculation formula, and the parameters are adjusted and optimized by comparing with the training set images. Black represents carbon fiber, gray represents adhesive, and the phase segmentation slice extracted by SegFormer features in the specific implementation method is shown in the figure. Figure 4 shown.

[0042] 4. 3D reconstruction of multiphase gas diffusion layer

[0043] The main way to obtain a three-dimensional structure is to first obtain a two-dimensional slice image through dimensionality reduction, perform threshold segmentation on the slice image, and then stack the two-dimensional image to expand it into a three-dimensional structure. The present invention uses the SegFormer semantic segmentation algorithm to train and judge according to the growth law, distribution characteristics, and structural shape of the carbon fiber and adhesive in the CT slice image of the gas diffusion layer, distinguish the segmentation of carbon fiber and adhesive, and then stack the phase-separated slice images to form a three-dimensional structure. The three-dimensional model reconstructed in this way has a higher accuracy and is closer to the real GDL structure. The dark gray represents the carbon fiber and the light gray represents the adhesive. The schematic diagram of the gas diffusion layer structure obtained by three-dimensional reconstruction is as follows Figure 5 shown.

Claims

1. A three-dimensional reconstruction method for a gas diffusion layer of a proton exchange membrane fuel cell, characterized in that The following steps are involved: (1) Using X-ray computed tomography technology to obtain multiple two-dimensional slices of the gas diffusion layer, the images were preprocessed using the K-means clustering algorithm to separate the voids. The porosity of the carbon paper was used as the judgment condition, and the correct preprocessed CT images were gradually obtained by changing the cluster center system. (2) Feature extraction is performed through SegFormer semantic segmentation to capture the pores, carbon fibers, and adhesives in the preprocessed CT image obtained in step 1 and assign different grayscale values ​​to generate phase segmentation slices. The specific process is as follows: For a given image, it is first divided into multiple small blocks and then input into the hierarchical Transformer encoder to obtain multi-level features with resolutions of 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image. For a small block, the block merging process used in Vision Transformer unifies an N×N×3 block into a 1×1×C vector; it is extended to unify a 2×2× The feature paths are unified into a 1×1× vector to obtain a hierarchical feature map; Use this method to change the hierarchy attributes from Zoom out to , and then iterate over any other feature maps in the hierarchy to combine non-overlapping images or feature patches; In the design of the self-attention layer, a reduction ratio is used to reduce the length of the sequence and reduce the time complexity; Mix-FFN is used to replace the position encoding system; 3×3 convolution and MLP are mixed into each FFN, and deep convolution is used to reduce the number of parameters and improve efficiency; Finally, four steps are performed in the MLP decoder: S1. The multi-level features from the MiT encoder are passed through the MLP layer to unify the channel dimension; S2. The features are upsampled to 1 / 4 and concatenated together; S3. The cascaded features F are fused using the MLP layer; S4. Another MLP layer uses the fused H / 4×W / 4×N_{cls} resolution features to predict the segmentation mask M, where N_{cls} represents the number of categories; thus, the number of fiber and adhesive categories is obtained, and the training accuracy is calculated according to the pixel accuracy calculation formula, and the parameters are adjusted and optimized by comparing with the training set images; (3) Using three-dimensional reconstruction, all phase segmentation slices are stacked to increase the dimension and obtain a three-dimensional gas diffusion layer structure.

2. The three-dimensional reconstruction method of the gas diffusion layer of a proton exchange membrane fuel cell according to claim 1, characterized in that Step 2 In feature extraction, the training set is the preprocessed CT image, the black area is the pores, the straight rods are the carbon fibers, and the other irregular shapes are the adhesives.

3. The three-dimensional reconstruction method of the gas diffusion layer of a proton exchange membrane fuel cell according to claim 1, characterized in that Step 2 uses pixel accuracy PA as an evaluation indicator in feature extraction; the pixel accuracy is the ratio of the number of pixels with the correct prediction category to the total number of pixels, and the specific formula is as follows: PA = (TP+TN) / (TP+TN+FP+FN) Among them, TP positive class is judged as positive class, FP negative class is judged as positive class, FN positive class is judged as negative class, and TN negative class is judged as negative class.

Images