A method for predicting the compression performance of gas diffusion layer assembly

Abstract

The present invention discloses a method for predicting the compression performance of a gas diffusion layer assembly, comprising the following steps: extracting features from a three-dimensional point cloud of a gas diffusion layer to establish a data set; training a vectorized encoder to vectorize compression values ​​to obtain performance parameters under a compression ratio; outputting the data set after three-dimensional feature extraction and the performance parameters under the compression ratio to a 1-dimensional feature model for fusion, and performing deep learning model regression as a new feature vector to train a new network model; inputting the three-dimensional structure point cloud under the compression ratio and the compression ratio information into the trained new network model for prediction, and finally obtaining the three-dimensional structure point cloud under the corresponding compression ratio, and outputting the predicted performance parameter parameters. The present invention inputs a three-dimensional slice 3D tiff image of carbon paper and the compression ratio and the performance value AI algorithm training under the corresponding compression ratio to achieve performance prediction of the three-dimensional structure of the gas diffusion layer under different compression ratios, effectively solving the deficiencies of the prior art.

Description

Technical Field

[0001] The invention relates to the field of new energy technology, and in particular to a method for predicting the compression performance of a gas diffusion layer assembly. Background Art

[0002] With the shortage of fossil energy such as oil and the urgent need for environmental protection, new energy vehicles have become a research hotspot for major automobile manufacturers and research and development institutions around the world. Among them, fuel cell technology is a clean energy conversion technology with great application prospects. Since it directly converts the chemical energy in the fuel into electrical energy without the need for fuel combustion, its energy conversion efficiency is not limited by the Carnot cycle, and the total energy utilization efficiency is theoretically greater than 60%. Fuel cells mainly include alkaline fuel cells, proton exchange membrane fuel cells (Polymer electrolyte membrane fuel cell, phosphoric acid fuel cell, molten carbonate fuel cell and solid oxide fuel cell, among which the most commercially successful is the proton exchange membrane fuel cell. PEMFC vehicles use hydrogen as input fuel, and the reaction output is only pure water. They have the advantages of no noise, no pollution, short refueling time (3-5 minutes) and long driving range (500-700 kilometers). As one of the most important components of proton exchange membrane fuel cells, the gas diffusion layer (Gas Diffusion Layer; GDL) is mainly responsible for transmitting reaction gases, exporting electrons, and discharging product water. The main characterization parameters of the gas diffusion layer include thickness, electronic conductivity, porosity, tortuosity, pore size, permeability and hydrophilicity. The thickness of the diffusion layer is generally between 100 and 400um. The diffusion layer provides a path for current and gas transmission. The thinner the thickness, the shorter the gas transmission and the smaller the mass transfer resistance, but its mechanical strength will also be reduced, resulting in a weakened support for the catalyst layer and membrane, reducing the durability of the battery.

[0003] When the gas diffusion layer is working in the fuel cell stack, oxygen, hydrogen and water are transmitted through the pores of the gas diffusion layer, and electrons and heat are conducted through the solid of the gas diffusion layer. When the assembly force of the fuel cell stack is applied to the gas diffusion layer through the bipolar plate, different microstructural deformations will occur in the areas under the flow channel and the rib plate (the part where the gas diffusion layer contacts the flow field plate). The microstructure of the gas diffusion layer is deformed by 10%-30%, and some carbon fibers are lifted and invade the flow channel. The carbon fibers under the rib plate are compressed more tightly, which will affect the transmission and conduction capacity of gas-water-heat-electricity in the gas diffusion layer, such as diffusion coefficient, permeability, thermal conductivity and electrical conductivity, and further change the output performance of the battery. There are currently few studies on the assembly compression of the gas diffusion layer, and most of the current studies are mainly based on numerical simulation because the experimental price is expensive and the accuracy of the results cannot be guaranteed. The pre-processing of the numerical model used in the fuel cell assembly compression simulation takes a long time, and the calculation process is slow, which makes it difficult to guide the assembly process in the actual production of the stack in real time.

[0004] Current methods for exploring compression deformation of gas diffusion layer materials mainly include experimental testing and numerical simulation methods: 1. Experimental testing Currently, the compression test of the gas diffusion layer can be completed through experimental instruments, and performance test experiments are carried out after the compression test, such as vertical conductivity, in-plane conductivity, vertical thermal conductivity, in-plane thermal conductivity, vertical gas diffusivity, in-plane gas diffusivity, etc. However, the experimental instruments for the gas diffusion layer are very expensive, and the test environment is very harsh. It is difficult to find a place in China that can complete all test items. Therefore, experiments are not the mainstream method for current research on gas diffusion layer assembly compression.

[0005] 2. Numerical simulation First, the three-dimensional structure is compressed and calculated. The three-dimensional structure is scanned by CT, and the gas diffusion layer assembly compression simulation is performed using display dynamics simulation software. It takes several weeks for pre-processing and several months for calculation. The robustness is poor and the calculation often does not converge. After completing the three-dimensional structure calculation, various performances are calculated in the compressed three-dimensional structure, which requires high computer configuration. Therefore, although the numerical simulation method is the mainstream method, it is difficult to implement and time-consuming. It is difficult to apply it to the engineering level to guide battery stack manufacturers to upgrade the assembly process.

[0006] However, the current method has the following defects. Experimental testing has the following disadvantages: 1) The instrument is expensive and only a few units can do it; 2) The test takes a long time; 3) The test environment is harsh and the operation requirements are high, making it difficult for non-professionals to test. Numerical simulation has the following disadvantages: 1) It requires high professionalism, and the pre-processing process is complex and time-consuming; 2) The calculation robustness is poor and convergence is difficult; 3) Due to the long calculation time and poor convergence, this method cannot be applied to large-scale state calculations.

[0007] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is to provide a method for predicting the compression performance of the gas diffusion layer assembly, inputting the 3D slice 3D tiff image of the carbon paper and the compression ratio and the performance value under the corresponding compression ratio into the AI ​​algorithm training, so as to realize the performance prediction of the gas diffusion layer 3D structure under different compression ratios. The compression value is vectorized and then fused with the feature vector extracted by deep learning from the 3D slice image, and then used as a new feature vector to be regressed by the deep learning model. Summary of the invention

[0008] To achieve the above object, the present invention provides a method for predicting the compression performance of a gas diffusion layer assembly, comprising the following steps:

[0009] Step 1, extracting features from the three-dimensional point cloud of the gas diffusion layer and establishing a data set;

[0010] Step 2: Train a vectorization encoder to vectorize the compressed value to obtain a feature vector corresponding to the compressed value;

[0011] Step 3: Output the data set after 3D feature extraction and the feature vector corresponding to the compression value to the 1D feature model for fusion, and use it as a new feature vector for deep learning model regression. Use the fused new feature vector as input and the corresponding performance parameter as output to train a new network model.

[0012] Step 4: input the new gas diffusion layer three-dimensional point cloud data and compression value to be predicted into the trained new network model for prediction, and finally obtain performance parameters under multiple different compression values.

[0013] Furthermore, feature extraction is performed through the 3DUnet deep learning model to establish a data set.

[0014] Furthermore, a vectorized encoder is trained through the Deep AE network.

[0015] Furthermore, the 1D feature model is a 1d-resnet deep learning model.

[0016] Furthermore, the 3DUnet deep learning model mainly includes two parts: the first part is the feature extraction part; the second part is the upsampling part;

[0017] Among them, for the feature extraction part, each pooling layer is a scale, including the original image scale, there are a total of 5 scales; for the upsampling part, each upsampling is fused with the same scale as the number of channels corresponding to the feature extraction part.

[0018] Furthermore, the new network model further upgrades the 3DUNet network structure and adds a residual network mechanism within each module.

[0019] Furthermore, the performance parameters include vertical electrical conductivity, in-plane electrical conductivity, vertical thermal conductivity, in-plane thermal conductivity, vertical diffusivity, and in-plane diffusivity.

[0020] The beneficial effects of the present invention are: 1) The present invention can quickly infer the vertical conductivity, in-plane conductivity, vertical thermal conductivity, in-plane thermal conductivity, vertical gas diffusivity, and in-plane gas diffusivity performance of the gas diffusion layer under different compression ratios, establish an AI model, and achieve second-level reasoning, which has great guiding significance for the fuel cell stack assembly process. Manufacturers can quickly determine a good assembly process parameter range based on the reasoning results.

[0021] 2) The present invention achieves good training results with a small amount of data sets, and has strong applicability in gas diffusion layer and other composite material assembly processes, numerical calculation research, and performance testing. It can enable fuel cell gas diffusion layer and other carbon paper manufacturers to quickly obtain the performance values ​​of carbon paper after compression, eliminating tedious testing processes and saving R&D expenses.

[0022] 3) The present invention uses a neural network to quickly infer the overall performance of the gas diffusion layer after assembly, which is beneficial for gas diffusion layer assembly manufacturers to quickly select assembly processes, reduce the number of tests, and improve production efficiency.

[0023] 4) This invention proposes a 3D to 1D AI reasoning route, which provides favorable guidance for subsequent AI in 3D simulation of engineering applications. This method can be extended from the fuel cell gas diffusion layer industry to other carbon paper and porous media industries, and can solve the research, manufacturing and testing problems of porous media in the fields of petrochemicals, air purification, textiles, papermaking, aerospace, atomic energy, etc., and has great scientific significance and industrial value.

[0024] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention.

[0025] Figure 1 is a flow chart of a method for predicting compression deformation of a gas diffusion layer according to the present invention;

[0026] Figure 2 is a network structure diagram of the 3DUNet network model of the present invention;

[0027] Figure 3 It is the upgraded 3D ResUnet network structure diagram of the present invention; BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 4 It is a custom autoencoder graph for the present invention;

[0029] Figure 5 It is a diagram of the training process of assembling the compression performance prediction model of the present invention;

[0030] Figure 6 It is a coordinate curve diagram for model training of the present invention.

[0031] like Figure 1 , 5 As shown, the present invention provides a method for predicting the compression performance of a gas diffusion layer assembly.

[0032] Step 1, extracting features from the three-dimensional point cloud of the gas diffusion layer and establishing a data set;

[0033] Step 2: Train a vectorization encoder to vectorize the compressed value to obtain a feature vector corresponding to the compressed value;

[0034] Step 3: Output the data set after 3D feature extraction and the feature vector corresponding to the compression value to the 1D feature model for fusion, and use it as a new feature vector for deep learning model regression. Use the fused new feature vector as input and the corresponding performance parameter as output to train a new network model.

[0035] Step 4: input the new gas diffusion layer three-dimensional point cloud data and compression value to be predicted into the trained new network model for prediction, and finally obtain performance parameters under multiple different compression values.

[0036] The present invention realizes the performance prediction of the three-dimensional structure of the gas diffusion layer under different compression ratios by inputting the three-dimensional slice 3D tiff image of the carbon paper and the compression ratio and the performance values ​​(vertical conductivity, in-plane conductivity, vertical thermal conductivity, in-plane thermal conductivity, vertical gas diffusivity, in-plane gas diffusivity) under the corresponding compression ratio into the AI ​​algorithm training. The compression value is vectorized and then fused with the feature vector extracted by deep learning from the three-dimensional slice image, and then used as a new feature vector to regress the deep learning model. The overall idea of ​​fusion is as follows: Figure 1 shown.

[0037] The key technologies of the gas diffusion layer assembly compression performance prediction method of the present invention can be divided into the following three modules and prediction data post-processing: image feature extraction, which can be performed through deep learning models such as 3DUnet;

[0038] Vectorize the compressed values ​​and train a vectorized encoder through the Deep AE network (the corresponding parameters can be frozen after pre-training);

[0039] 1D feature extraction: can be performed through the 1d-resnet deep learning model. 3DUNet is used as the basic model for 3D image feature extraction in 3D slices. At the same time, the vector features of the compressed value are extracted through the trained Deep AE network. After the two feature functions are fused, they are further processed through the 1d-resnet deep learning model. Among them: 1. The basic model of model compression of the 3DUNet model uses 3DUNet. The network is upgraded and improved based on the overall structure of 2DUNet. The focus is on converting the 2d convolution into 3d convolution, so that it can be suitable for extracting the features of 3D slices.

[0040] Figure 2 The 3DUNet network structure diagram. Each blue box corresponds to a multi-channel feature map, where the number of channels is marked on the top of the box and the size of xy is located in the lower left corner of the box; the white box represents the copied feature map; the arrows represent different operations.

[0041] The model mainly consists of two parts. The first part can be seen in the right picture. Its main function is feature extraction. The second part is the upsampling part. Since the network structure is U-shaped and the input data is a 3D slice image, it is called 3DUNet network. In the feature extraction part, each pooling layer has a scale, including the original image scale, there are a total of 5 scales. In the upsampling part, each upsampling is fused with the same scale as the number of channels corresponding to the feature extraction part, but it must be cropped before fusion. The fusion here is splicing.

[0042] The network can use valid annotated data more effectively from very few training images by relying on data enhancement. U-net uses a completely different feature fusion method: splicing. 3DUNet splices features together in the channel dimension to form thicker features. The addition of corresponding points used in FCN fusion does not form thicker features. The five pooling layers realize the network's multi-scale feature recognition of image features. The upsampling part will fuse the output of the feature extraction part. This actually fuses multi-scale features together. Taking the last upsampling as an example, its features come from both the output of the first convolution block (same-scale features) and the output of the upsampling (large-scale features). This connection runs through the entire network. It can be seen that there are four fusion processes in the above network, and the corresponding FCN network only performs fusion in the last layer.

[0043] In the final model, the 3DUNet network structure was further upgraded, and a residual network mechanism was added to each block to make the model fitting ability stronger to extract the image features of the three-dimensional model of the gas diffusion layer. Figure 3This is a schematic diagram of the upgraded 3DResUnet network structure.

[0044] 2. Compression value vectorization Since the compression values ​​used for training are fixed at 5%, 10%, 15%, 20%, 25%, and 30%, this prediction scheme aims to achieve a continuous compression ratio of random input, and can perform compression structure prediction. Therefore, in order to achieve vectorized encoding of compression values, custom Figure 3 As shown in the autoencoder. Since the compression values ​​used for training are fixed at 5%, 10%, 15%, 20%, 25%, and 30%, this prediction scheme aims to achieve a continuous compression ratio of random input, and can predict the compression structure. Therefore, in order to achieve vectorized encoding of the compression value, the following Autoencoder is customized: The specific encoder-decoder structure is as follows

[0045]

[0046] By analyzing all possible compression ratio values ​​in the data, that is, 0 to 0.30 with an accuracy of 0.01, training data with an interval of 0.01 from 0 to 0.3 is constructed.

[0047] The training parameters of the present invention are Adam, lr = 0.001, weight_decay = 1e-5,

[0048] epoch3000, Figure 6 This is a diagram of the training process.

[0049] The prior art experiments to study the compression performance of gas diffusion layer structures are accompanied by a lot of expensive material costs and time costs, and the experimental environment is harsh, and a set of tests need to be completed at multiple points. The present invention can predict the performance changes of the gas diffusion layer under different compression ratios in seconds, saving a lot of material costs, and assisting in exploring the improvement of the fuel cell stack assembly compression process; secondly, the prior art numerical calculations are extremely time-consuming and labor-intensive. The present invention can output various performances after compression in seconds, guide the fuel cell stack assembly compression process, greatly shorten the research cycle of the stack assembly, and quickly predict the quality of the assembly process.

[0050] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for predicting the compression performance of a gas diffusion layer assembly, characterized in that: The following steps are involved: Step 1, extracting features from the three-dimensional point cloud of the gas diffusion layer and establishing a data set; Step 2: Train a vectorization encoder to vectorize the compressed value to obtain a feature vector corresponding to the compressed value; Step 3: Output the data set after 3D feature extraction and the feature vector corresponding to the compression value to the 1D feature model for fusion, and use it as a new feature vector for deep learning model regression. Use the fused new feature vector as input and the corresponding performance parameter as output to train a new network model. Step 4: input the new gas diffusion layer three-dimensional point cloud data and compression value to be predicted into the trained new network model for prediction, and finally obtain performance parameters under multiple different compression values.

2. A method for predicting the compression performance of a gas diffusion layer assembly as claimed in claim 1, characterized in that: Feature extraction is performed through the 3DUnet deep learning model to establish a data set.

3. A method for predicting the compression performance of a gas diffusion layer assembly as claimed in claim 1, characterized in that: A vectorized encoder is trained via the Deep AE network.

4. A method for predicting the compression performance of a gas diffusion layer assembly as claimed in claim 1, characterized in that: The 1D feature model is a 1D-resnet deep learning model.

5. A method for predicting the compression performance of a gas diffusion layer assembly as claimed in claim 2, characterized in that: The 3DUnet deep learning model mainly consists of two parts: the first part is the feature extraction part; the second part is the upsampling part; Among them, for the feature extraction part, each pooling layer is a scale, including the original image scale, there are a total of 5 scales; for the upsampling part, each upsampling is fused with the same scale as the number of channels corresponding to the feature extraction part.

6. A method for predicting the compression performance of a gas diffusion layer assembly as claimed in claim 1, characterized in that: The new network model further upgrades the 3DUNet network structure and adds a residual network mechanism within each module.

7. A method for predicting the compression performance of a gas diffusion layer assembly as claimed in claim 1, characterized in that: The performance parameters include vertical electrical conductivity, in-plane electrical conductivity, vertical thermal conductivity, in-plane thermal conductivity, vertical diffusivity, and in-plane diffusivity.

Images