Detection System and Method for Hydrogen Fuel Cell Flow Guide Bipolar Plate

Abstract

The present invention relates to the field of bipolar plate detection technology, and discloses a detection system and method for hydrogen fuel cell guide bipolar plates, the method comprising: performing defect recognition on collected image data based on a multi-scale feature fusion network, and performing airtightness detection on the bipolar plates based on a multi-station airtightness testing device. According to the defect data and the airtightness data, the comprehensive quality score of each bipolar plate is determined in combination with the attention mechanism and fuzzy logic. According to the relationship between the comprehensive quality score and the configured preset comprehensive quality score, the defect severity classification is determined. The confidence of the bipolar plate classification after the defect severity classification is evaluated based on the Monte Carlo Dropout method, and an output report is generated. The present invention improves the accuracy, stability and reliability of bipolar plate defect detection through technical means such as multi-scale feature fusion, airtightness testing, intelligent dynamic adjustment and uncertainty assessment, and realizes the accurate identification and comprehensive quality assessment of complex defects of bipolar plates.

Description

Technical Field

[0001] The present invention relates to the technical field of bipolar plate detection, and in particular to a detection system and method for a flow guide bipolar plate of a hydrogen fuel cell. Background Art

[0002] Hydrogen fuel cells are a crucial clean energy technology. Their bipolar plates play a crucial role in separating gases, conducting current, and providing structural support. The manufacturing precision and quality of bipolar plates directly impact the performance, lifespan, and safety of fuel cells. Therefore, high-precision and efficient testing of bipolar plates is crucial.

[0003] Currently, traditional inspection methods primarily rely on single-mode optical imaging or airtightness testing. However, when faced with complex defects (such as hidden cracks and micro-dents) and multi-factor comprehensive assessments, they suffer from limited recognition capabilities, high false positive rates, and difficulty accurately quantifying the impact of defects on overall quality. These methods cannot meet the fuel cell industry's demand for high-precision inspection. Furthermore, most inspection methods rely on single-modal visual inspection or independent physical testing, making it difficult to fully utilize multi-source data to improve inspection accuracy. For example, while optical inspection can identify surface scratches and dents, its ability to detect micro-cracks or internal defects is limited. Airtightness testing can determine leaks but struggles to directly locate defects. Furthermore, traditional defect assessment relies on fixed rules and lacks intelligent dynamic adjustment mechanisms, resulting in insufficient reliability and adaptability to diverse operating conditions. Furthermore, existing methods lack the ability to assess the uncertainty of inspection results, making it difficult to provide effective decision-making based on low-confidence classification results, impacting inspection reliability and stability.

[0004] Therefore, there is an urgent need to invent a detection technology for bipolar plates to solve the problems of limited recognition ability of existing detection methods when facing complex defects, difficulty in integrating multi-source data for accurate evaluation, lack of intelligent dynamic adjustment and uncertainty quantification mechanism, resulting in low detection accuracy and reliability of bipolar plates. Summary of the Invention

[0005] In view of this, the present invention proposes a detection system and method for hydrogen fuel cell guide bipolar plates, aiming to solve the problems in current technology such as limited recognition ability when facing complex defects, difficulty in integrating multi-source data for accurate evaluation, lack of intelligent dynamic adjustment and uncertainty quantification mechanism, which in turn leads to low bipolar plate detection accuracy and reliability.

[0006] The present invention proposes a method for detecting a flow-guiding bipolar plate of a hydrogen fuel cell, comprising:

[0007] Defect recognition is performed on the collected image data based on a multi-scale feature fusion network, and the air tightness test of the bipolar plates is performed using a multi-station air tightness test device;

[0008] Based on the defect data and airtightness data, the comprehensive quality score of each bipolar plate is determined by combining the attention mechanism and fuzzy logic;

[0009] Determine the defect severity level based on the relationship between the comprehensive quality score and the configured preset comprehensive quality score;

[0010] The confidence level of bipolar plate classification after defect severity classification is evaluated based on the Monte Carlo Dropout method, and an output report is generated.

[0011] Furthermore, when performing defect recognition on the collected image data based on the multi-scale feature fusion network, it includes:

[0012] Preprocess images collected under different lighting conditions to enhance image contrast and edge information;

[0013] Based on the multi-scale feature extraction technology, the local detail information and global structure information of different receptive fields in the pre-processed image are extracted, and weighted fusion is performed based on the eliminated multi-scale features to obtain the comprehensive features of the processed image;

[0014] Based on the comprehensive features of the fused image, a threshold comparison is performed on the characteristic responses of each area in the image to determine whether there are defects in the bipolar plate.

[0015] Furthermore, preprocessing of images collected under different lighting conditions includes:

[0016] Obtaining the image contrast of each region in the collected image, and obtaining the image contrast mean of each region, and comparing the image contrast mean with the contrast of each image to obtain images of each region whose contrast is lower than the image contrast mean;

[0017] For each area of the image with a contrast value lower than the image mean, brightness normalization is performed based on histogram equalization;

[0018] Obtain the image information after brightness normalization processing, and perform denoising on the image based on non-local mean filtering.

[0019] Furthermore, obtaining the comprehensive features of the processed image includes:

[0020] Multi-scale feature extraction technology is used on the pre-processed image to simultaneously extract local detail information and global structural information in the image based on several preset receptive fields of different sizes;

[0021] The extracted multi-scale features are weighted and fused based on pre-set weights to generate comprehensive image features, where:

[0022] The weights of preset scale features are set according to historical data, and the extracted scale features are matched with the preset scale features to obtain the weights of the extracted scale features. Local details and global structural information are weighted and spliced according to the weights to generate comprehensive features.

[0023] Furthermore, based on the comprehensive features of the fused image, threshold comparison is performed on the feature responses of each region in the image to determine whether the bipolar plate has defects, including:

[0024] Based on historical data and experimental samples, establish the correlation between detection indicators and defect responses;

[0025] Obtain the distance metric between each correlation formula, and iteratively cluster each correlation formula based on the distance metric;

[0026] Based on the lonely forest tree algorithm, the outliers of each cluster correlation in the clustering results are obtained, and the normal features and defect features in the same area are determined according to the relationship between the outliers and the preset outliers.

[0027] Determine the defect judgment threshold based on the feature vector between the normal feature and the defect feature;

[0028] Obtain the characteristic vector of the characteristic response of the region in the image, and determine whether there is a defect in the region in the image based on the relationship between the characteristic vector of the characteristic response and the defect judgment threshold:

[0029] When the eigenvector of the characteristic response is lower than the defect judgment threshold, it is determined that there is no defect in the area in the image;

[0030] When the eigenvector of the feature response is higher than or equal to the defect judgment threshold, it is determined that there is a defect in the area in the image.

[0031] Furthermore, based on the defect data and airtightness data, combined with the attention mechanism and fuzzy logic, the comprehensive quality score of each bipolar plate is determined, including:

[0032] Perform feature extraction on the defect data and airtightness data of the bipolar plate in different modes;

[0033] Perform weighted analysis on defect data and airtightness data based on the attention mechanism to determine key feature information;

[0034] Dynamically assign weights to defect data and airtightness data based on fuzzy logic, and determine the contribution ratio of each data in quality assessment based on the confidence level of the data;

[0035] The comprehensive quality score of the bipolar plate is determined based on the weighted fusion defect score, air tightness score and morphology feature score.

[0036] Furthermore, when obtaining the defect score, air tightness score and morphology feature score of the bipolar plate, it includes:

[0037] Collect defect data of bipolar plates and perform defect identification based on computer vision algorithms to extract defect size, shape, and distribution characteristics and generate defect scores;

[0038] Obtain air tightness data and determine the air tightness score based on leakage rate and air pressure change parameters;

[0039] The surface morphology information of the bipolar plate is obtained based on three-dimensional imaging technology, and the morphology feature score is determined based on the surface flatness and deformation degree.

[0040] Furthermore, based on the relationship between the comprehensive quality score and the configured preset comprehensive quality score, the defect severity level is determined, including:

[0041] The defect severity level is determined based on the relationship between the comprehensive quality score and the configured first preset comprehensive quality score and second preset comprehensive quality score:

[0042] When the comprehensive quality score is lower than the first preset comprehensive quality score, the defect severity level is determined to be high;

[0043] When the comprehensive quality score is higher than or equal to the first preset comprehensive quality score and lower than the second preset comprehensive quality score, the defect severity level is determined to be medium;

[0044] When the comprehensive quality score is higher than or equal to the second preset comprehensive quality score, the defect severity level is determined to be low;

[0045] Among them, the first preset comprehensive quality score is less than the second preset comprehensive quality score, and the defect severity levels are ranked from high to low as high, medium and low.

[0046] Furthermore, the confidence level of the bipolar plate classification after defect severity classification is evaluated based on the Monte Carlo Dropout method, and an output report is generated, including:

[0047] Based on the Monte Carlo Dropout method, multiple forward propagations are performed in the inference phase to obtain multiple prediction results of defect classification;

[0048] Obtain the probability distribution characteristics of the defect classification prediction results and determine the classification confidence based on uncertainty measurement methods such as entropy or variance;

[0049] The output result is determined based on the relationship between the confidence level and the configured preset confidence level:

[0050] When the confidence level is lower than the preset confidence level, it is determined that a secondary evaluation of the comprehensive quality score of the bipolar plate is performed;

[0051] When the confidence level is higher than or equal to the preset confidence level, it is determined that the output is graded according to the severity of the defect of the bipolar plate.

[0052] Compared with existing technologies, the present invention offers the following advantages: through multi-scale feature fusion, the system can extract key defect information from diverse image data, thereby identifying complex defects such as subtle cracks and micro-depressions. Furthermore, the integration of airtightness testing with image inspection enables a more comprehensive inspection process, enabling both surface defects identification and accurate assessment of internal structural integrity through airtightness testing. This synergistic effect of multimodal data significantly improves inspection accuracy and comprehensiveness, avoiding the potential for missed and false detection issues associated with traditional single-mode inspection methods. Furthermore, the integration of an attention mechanism and fuzzy logic enables more intelligent processing of defect and airtightness data, dynamically adjusting the weights of different data sources to generate a more accurate overall quality score. This mechanism effectively addresses variations in operating conditions and enhances the adaptive capabilities of the inspection system. Furthermore, the Monte Carlo Dropout method, used to assess the confidence level of defect severity, quantifies the uncertainty of inspection results, providing a reliable basis for final decision-making and ensuring the accuracy of grading results. Ultimately, the resulting quality report provides more objective and scientific data support for manufacturing and quality control, optimizing decision-making during the inspection process and improving the overall quality assurance of bipolar plates.

[0053] On the other hand, the present application also provides a detection system for a hydrogen fuel cell flow guide bipolar plate, comprising:

[0054] A composite optical imaging unit, configured with a ring-shaped LED array light source, a near-infrared imaging module, and a structured light module, is configured to perform defect recognition on the collected image data based on a multi-scale feature fusion network;

[0055] A multi-station airtightness testing device configured to perform airtightness testing on bipolar plates;

[0056] an evaluation module configured to determine a comprehensive quality score for each bipolar plate based on the defect data and the airtightness data in combination with an attention mechanism and fuzzy logic; the evaluation module is further configured to determine a defect severity grade based on a relationship between the comprehensive quality score and a configured preset comprehensive quality score;

[0057] The output module is configured to evaluate the confidence of the bipolar plate classification after defect severity classification based on the Monte Carlo Dropout method and generate an output report.

[0058] It is understandable that the detection system and method for a hydrogen fuel cell flow guide bipolar plate in the above-mentioned embodiments of the present invention have the same beneficial effects and will not be described in detail. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0060] Figure 1 A flowchart of a method for detecting a flow guide bipolar plate of a hydrogen fuel cell provided by an embodiment of the present invention;

[0061] Figure 2 A functional block diagram of a detection system for a hydrogen fuel cell flow guide bipolar plate provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, unless there is a conflict, the embodiments of the present disclosure and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0063] like Figure 1 As shown, in some embodiments of the present application, this embodiment provides a method for detecting a flow guide bipolar plate of a hydrogen fuel cell, comprising:

[0064] Step S100: Defect recognition is performed on the collected image data based on a multi-scale feature fusion network, and air tightness testing is performed on the bipolar plate based on a multi-station air tightness testing device.

[0065] Specifically, defect detection based on a multi-scale feature fusion network for captured image data involves preprocessing images captured under different lighting conditions to enhance image contrast and edge information. Using multi-scale feature extraction technology, local detail information and global structural information from different receptive fields within the preprocessed image are extracted, and weighted fusion is performed based on the eliminated multi-scale features to obtain comprehensive features of the processed image. Based on the fused comprehensive features of the image, threshold comparisons are performed on the feature responses of each region in the image to determine whether the bipolar plate is defective.

[0066] Specifically, preprocessing images captured under different lighting conditions involves obtaining the image contrast of each region in the captured image, obtaining the mean image contrast of each region, comparing the mean image contrast with the contrast of each image, and obtaining images of regions with contrast values below the mean image contrast. For each region with contrast values below the mean image contrast, brightness normalization is performed using histogram equalization. After obtaining brightness-normalized image information, the image is denoised using non-local means filtering.

[0067] Specifically, obtaining the comprehensive features of the processed image includes: applying a multi-scale feature extraction technique to the pre-processed image, and simultaneously extracting local detail information and global structural information from the image based on a number of preset receptive fields of different sizes. The extracted multi-scale features are weighted and fused based on pre-set weights to generate comprehensive image features, wherein: weights of preset scale features are set based on historical data, each extracted scale feature is matched with each preset scale feature, and the weights of each extracted scale feature are obtained. Local detail and global structural information are weightedly spliced based on the weights to generate comprehensive features.

[0068] Specifically, based on the comprehensive features of the fused image, a threshold comparison is performed on the feature responses of each area in the image to determine whether the bipolar plate has defects, including: establishing a correlation between the detection index and the defect response based on historical data and experimental samples. Obtaining the distance measurement between each correlation, and iteratively clustering the correlation based on the distance measurement. Obtaining the outliers of each cluster correlation in the clustering result based on the lonely forest tree algorithm, and determining the normal features and defect features in the same area based on the relationship between the outliers and the preset outliers. Determining the defect judgment threshold based on the feature vector between the normal features and the defect features. Obtaining the feature vector of the feature response of the area in the image, and determining whether the area in the image has defects based on the relationship between the feature vector of the feature response and the defect judgment threshold: when the feature vector of the feature response is lower than the defect judgment threshold, it is determined that there is no defect in the area in the image. When the feature vector of the feature response is higher than or equal to the defect judgment threshold, it is determined that there is a defect in the area in the image.

[0069] As can be seen, by analyzing image contrast, calculating the mean contrast of each region, and identifying areas with sub-average contrast, the image information in these areas is enhanced in subsequent processing. Brightness normalization (such as histogram equalization) and denoising (such as non-local means filtering) ensure effective noise removal in the image, making subsequent defect detection more accurate and eliminating the impact of uneven illumination or noise interference. Next, the image undergoes multi-scale feature extraction. During this process, the image is divided into multiple receptive fields for feature extraction. Each receptive field can extract local details and global structural information at different scales. Multi-scale feature extraction captures both details and macrostructure in the image, which is crucial for identifying tiny defects or complex cracks. By setting several preset receptive field scales, feature information at different scales is extracted, providing rich feature support for subsequent defect detection. After multi-scale feature extraction, the extracted features at each scale are weighted and fused according to the set weights. Historical data is used to set the weights for the pre-set scale features. These weights can be trained using a machine learning model based on existing annotated data to ensure that important features receive higher weights and less important features receive lower weights. After weighted fusion, the resulting comprehensive image features incorporate detailed information about both local details and global structure, enabling more accurate defect identification and classification in subsequent stages. Based on the fused image comprehensive features, threshold comparisons are performed on the feature responses of each region in the image. By establishing correlation equations based on historical data and experimental samples, the similarity of the feature responses is measured, and distance metrics between the features are calculated. This information is then used for cluster analysis. Cluster analysis effectively distinguishes normal from abnormal regions and identifies defective areas. Finally, the lonely forest tree algorithm is used to identify outliers in the clustering results. By comparing these outliers with preset outliers, the characteristic differences between defective and normal regions are determined, and a defect judgment threshold is set based on this difference. Finally, the presence of defects in the image region is determined by comparing the feature vector with the threshold. If the feature response of a region is below the judgment threshold, the region is considered defect-free; if the feature response is above or equal to the judgment threshold, the region is considered defective.

[0070] As can be understood, image preprocessing improves image quality by enhancing contrast and edge information for images acquired under varying lighting conditions. Since uneven lighting conditions can affect image clarity and detail during acquisition, image preprocessing eliminates these lighting issues through contrast analysis and brightness normalization (such as histogram equalization), thereby ensuring image accuracy and stability in subsequent analysis. De-noising further enhances image quality, making defect detection more reliable and reducing noise interference with image feature extraction. Secondly, multi-scale feature extraction technology simultaneously extracts both local detail and global structural information from the image. Setting different receptive fields enables detection to capture defect features of varying sizes and complexities at multiple scales. This approach enhances the detection of small cracks, internal defects, and other complex morphologies, resulting in more accurate detection, especially for small and hidden defects. Weighted fusion of multi-scale features rationally distributes the contribution of features at each scale based on historical data and preset weights. The weighted fusion process improves the ability to integrate different image features, effectively combining detailed information with global structural information to generate comprehensive image features. This not only enhances defect recognition accuracy but also enables adaptive adjustment of detection strategies based on different image features, improving overall detection effectiveness. Furthermore, the combination of threshold comparison and cluster analysis enables the establishment of correlation equations based on historical data and experimental samples, providing stronger theoretical support for defect determination. Distance measurement and cluster analysis effectively distinguish between normal and defective areas, improving identification accuracy. Outliers in the clustering results are effectively identified, providing a reliable basis for defect determination. Comparison with pre-set outlier values further enhances accuracy and avoids false or missed detections. Finally, the lonely forest tree algorithm optimizes the clustering analysis results, identifies outliers, and performs effective feature determination. By determining the defect determination threshold, each region in the image can be meticulously inspected and, based on the relationship between the feature response and the threshold, accurately determined whether a defect exists within the region. The effectiveness of this process ensures high reliability of the entire defect detection process and provides accurate feedback for bipolar plate quality assessment. This not only improves the accuracy of defect identification, but also provides a scientific basis for subsequent quality assessment, ensuring the comprehensiveness and efficiency of the detection process.

[0071] Step S200: Determine the comprehensive quality score of each bipolar plate based on the defect data and the airtightness data, combined with the attention mechanism and fuzzy logic.

[0072] Specifically, based on defect data and airtightness data, combined with the attention mechanism and fuzzy logic, the comprehensive quality score of each bipolar plate is determined, including: feature extraction of data of different modalities for the defect data and airtightness data of the bipolar plate. Based on the attention mechanism, the defect data and airtightness data are weighted analyzed to determine key feature information. Based on fuzzy logic, the weights of the defect data and airtightness data are dynamically assigned, and the contribution ratio of each data in the quality assessment is determined based on the confidence level of the data. The comprehensive quality score of the bipolar plate is determined based on the weighted fusion of defect score, airtightness score and morphological feature score.

[0073] Specifically, obtaining the defect score, air tightness score, and topographic feature score for bipolar plates involves: collecting bipolar plate defect data and identifying defects using computer vision algorithms to extract the size, shape, and distribution characteristics of the defects and generate a defect score; obtaining air tightness data and determining the air tightness score based on leakage rate and air pressure change parameters; and obtaining the surface topography information of the bipolar plates using 3D imaging technology and determining the topographic feature score based on surface flatness and deformation.

[0074] As can be seen, defect data extraction is achieved through computer vision algorithms, which automatically identify defects in images and extract relevant size, morphology, and distribution features. This process effectively obtains detailed defect information from images and provides a basis for subsequent defect scoring. Tightness data is obtained by analyzing leakage rate and air pressure variation parameters, ensuring comprehensive tightness testing and reflecting the air tightness performance of the bipolar plate in actual use. Topographic feature scoring uses 3D imaging technology to obtain surface topography data of the bipolar plate. Based on an assessment of surface flatness and deformation, this ensures that the geometric quality of the bipolar plate is fully quantified. Secondly, an attention mechanism is used to perform a weighted analysis of defect data and tightness data. By assigning different importance weights to different data items, the attention mechanism highlights key features and automatically adjusts the influence of each data item on the final score. This technology identifies the most meaningful features from multiple information and focuses on the parameters most relevant to quality assessment, optimizing data utilization efficiency. Combined with fuzzy logic, this process further enhances the ability to handle fuzzy and uncertain data. Fuzzy logic can dynamically assign weights to defect and airtightness data, and adjust them in each evaluation link based on the confidence level of the data, so that the final quality score is more in line with the uncertainty and variability in the actual detection process. Finally, the comprehensive scoring method based on weighted fusion can integrate data from different modalities (such as defect scores, airtightness scores, and morphological feature scores). Each score is calculated independently based on different physical parameters, and after weighted fusion, a comprehensive quality score for the bipolar plate is generated. This process ensures that each data can fully reflect its actual contribution to the bipolar plate quality assessment in the final score, making the quality score more accurate and comprehensive. Through this multi-dimensional comprehensive analysis, the test can comprehensively evaluate the quality of the bipolar plate and ensure its reliability in various performance aspects.

[0075] It's understandable that by combining defect and airtightness data, and incorporating various techniques such as computer vision and 3D imaging, a comprehensive assessment of bipolar plates' defects, airtightness performance, and topographical characteristics can be achieved. Traditional inspection methods often focus on a single metric, but the integrated use of multiple data sources can more comprehensively and accurately reflect the overall quality of bipolar plates, avoiding the one-sidedness and shortcomings of traditional methods and improving inspection reliability and accuracy. Secondly, the use of an attention mechanism and fuzzy logic further enhances the intelligence of data processing. The attention mechanism performs weighted analysis on data from different modalities, automatically identifying key features. This gives significant weight to important defect characteristics or airtightness parameters in the quality score, effectively improving the relevance and effectiveness of inspection results. Fuzzy logic dynamically adjusts these weights based on data uncertainty and confidence, enhancing the system's adaptability to varying operating conditions and inspection environments, ensuring that evaluation results are more realistic and improving the robustness and flexibility of the inspection process. Finally, the application of a comprehensive scoring mechanism delivers a multi-dimensional, integrated assessment. By weightedly integrating defect scores, airtightness scores, and morphological feature scores, the advantages and information of each data point can be maximized, resulting in a more comprehensive and accurate quality score. This approach not only evaluates each performance factor individually but also considers the relationships between them in a comprehensive assessment, further improving the comprehensiveness and reliability of quality inspection results and providing data support for subsequent optimization and quality control of hydrogen fuel cell flow guide bipolar plates.

[0076] Step S300: Determine the defect severity level based on the relationship between the comprehensive quality score and the configured preset comprehensive quality score.

[0077] Specifically, when determining the defect severity level based on the relationship between the comprehensive quality score and the configured preset comprehensive quality score, it includes: determining the defect severity level based on the relationship between the comprehensive quality score and the configured first preset comprehensive quality score and second preset comprehensive quality score: when the comprehensive quality score is lower than the first preset comprehensive quality score, the defect severity level is determined to be high. When the comprehensive quality score is higher than or equal to the first preset comprehensive quality score, and the comprehensive quality score is lower than the second preset comprehensive quality score, the defect severity level is determined to be medium. When the comprehensive quality score is higher than or equal to the second preset comprehensive quality score, the defect severity level is determined to be low. Among them, the first preset comprehensive quality score is lower than the second preset comprehensive quality score, and the defect severity levels are, from high to low, high, medium, and low.

[0078] Step S400: Evaluate the confidence level of the bipolar plate classification after defect severity classification based on the Monte Carlo Dropout method, and generate an output report.

[0079] Specifically, when evaluating the confidence of the bipolar plate grading after defect severity grading based on the Monte Carlo Dropout method and generating an output report, it includes: performing multiple forward propagations in the inference phase based on the Monte Carlo Dropout method to obtain multiple prediction results of the defect grading. Obtain the probability distribution characteristics of the defect grading prediction results, and determine the grading confidence based on uncertainty measurement methods such as entropy or variance. According to the relationship between the confidence and the configured preset confidence, the output result is determined: when the confidence is lower than the preset confidence, it is determined to conduct a secondary evaluation of the comprehensive quality score of the bipolar plate. When the confidence is higher than or equal to the preset confidence, it is determined to output according to the defect severity grading of the bipolar plate.

[0080] As can be seen, the Monte Carlo Dropout method is used to address the model's prediction uncertainty and assess the confidence level of defect severity based on the probability distribution characteristics obtained through multiple forward propagations. Dropout is a common regularization technique in deep learning models, often used to prevent overfitting. During the inference phase, Dropout randomly "drops out" certain nodes in the neural network during each forward propagation, resulting in different model outputs and thus simulating model uncertainty. By performing multiple forward propagations, multiple predictions are obtained, generating a probability distribution for each defect classification result. Secondly, by calculating the probability distribution characteristics of multiple predictions, uncertainty metrics can be obtained. These metrics, such as entropy or variance, quantify prediction uncertainty. For example, a higher entropy value indicates greater uncertainty in the model's defect classification, indicating significant disagreement in the model's judgment of a specific defect; while variance measures the degree of dispersion between different predictions. Using these metrics, each defect classification result can be assigned a confidence level, reflecting the reliability and stability of the model in making defect judgments. Finally, based on the calculated confidence level, it is compared with the preset confidence threshold to determine whether a secondary evaluation of the bipolar plate is required. Specifically, when the confidence level is lower than the preset threshold, it indicates that the model has a large uncertainty in a certain classification. In this case, further evaluation or adjustment may be required to ensure the reliability of the test results. When the confidence level is higher than or equal to the preset threshold, it indicates that the model's prediction has a high degree of certainty and can be directly output. This method, by introducing a confidence level assessment, helps improve the decision-making accuracy of the test results, especially when dealing with defect classification tasks with high uncertainty.

[0081] It's understandable that the Monte Carlo Dropout method effectively addresses model uncertainty. In defect severity classification tasks, traditional models often generate varying degrees of uncertainty due to the complexity and diversity of the data. Through multiple forward propagations, Dropout provides multiple samples for each prediction, forming a probability distribution. This method allows the model to not only provide predictions but also quantify the uncertainty of these results, thereby improving the reliability and stability of defect classification. This technique enables more accurate risk assessment of model predictions, avoiding misjudgments caused by a single prediction. Secondly, uncertainty metrics such as entropy or variance can further enhance the model's decision-making capabilities in complex scenarios. Entropy and variance provide effective ways to quantify the uncertainty of model predictions. When the model's judgments on certain defects differ significantly, high entropy or variance can alert the decision-making system that further analysis is required. This approach eliminates the need for evaluation results based on a single prediction and, through uncertainty metrics, allows for flexible adaptation to inspection tasks of varying complexity, ensuring a more reliable decision-making process. Finally, by comparing the predictions against a preset confidence threshold, the model can adaptively adjust whether a second evaluation is necessary. This mechanism enhances the system's intelligence and flexibility, triggering further verification when confidence levels are low, effectively avoiding erroneous decisions resulting from low-confidence predictions. Furthermore, when confidence levels exceed a threshold, the system can directly output based on the existing classification results, improving overall inspection efficiency. This mechanism optimizes the inspection process and improves the overall accuracy and efficiency of bipolar plate inspections, particularly when faced with complex defects or data, providing a more accurate basis for final quality assessment.

[0082] In the above-mentioned embodiment, through multi-scale feature fusion, the system can extract key defect information from different image data, thereby identifying complex defects such as hidden cracks and tiny dents. Furthermore, the combination of airtightness testing and image inspection enables a more comprehensive inspection, capable of not only identifying surface defects but also accurately assessing the integrity of the internal structure through airtightness testing. This synergistic effect of multimodal data significantly improves the accuracy and comprehensiveness of inspection, avoiding the problems of missed and false detections that can occur with traditional single-mode inspection methods. Furthermore, the combination of attention mechanisms and fuzzy logic enables more intelligent processing of defect and airtightness data, dynamically adjusting the weights of different data sources to generate a more accurate overall quality score. This mechanism effectively adapts to changes in different operating conditions and enhances the adaptive capabilities of the inspection system. Furthermore, the confidence level of defect severity assessed using the Monte Carlo Dropout method can quantify the uncertainty of the inspection results, providing a reliable basis for final decision-making and ensuring the accuracy of the grading results. Ultimately, the output quality report provides more objective and scientific data support for manufacturing and quality control, optimizes decision-making during the inspection process, and improves the overall quality assurance of bipolar plates.

[0083] In another preferred embodiment based on the above embodiment, Figure 2 As shown, this embodiment provides a detection system for a hydrogen fuel cell flow guide bipolar plate, comprising:

[0084] The composite optical imaging unit is equipped with a ring-shaped LED array light source, a near-infrared imaging module and a structured light module. The composite optical imaging unit is configured to perform defect recognition on the collected image data based on a multi-scale feature fusion network.

[0085] The multi-station air tightness testing device is configured to perform air tightness testing on bipolar plates.

[0086] The evaluation module is configured to determine a comprehensive quality score for each bipolar plate based on the defect data and the airtightness data, using an attention mechanism and fuzzy logic. The evaluation module is also configured to determine a defect severity classification based on a relationship between the comprehensive quality score and a configured preset comprehensive quality score.

[0087] The output module is configured to evaluate the confidence of the bipolar plate classification after defect severity classification based on the Monte Carlo Dropout method and generate an output report.

[0088] It is understandable that the detection system and method for a hydrogen fuel cell flow guide bipolar plate in the above-mentioned embodiments of the present invention have the same beneficial effects and will not be described in detail.

[0089] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0091] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0092] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for detecting a flow-guiding bipolar plate of a hydrogen fuel cell, characterized in that: include: Defect recognition is performed on the collected image data based on a multi-scale feature fusion network, and the air tightness test of the bipolar plates is performed using a multi-station air tightness test device; Based on the defect data and airtightness data, the comprehensive quality score of each bipolar plate is determined by combining the attention mechanism and fuzzy logic; Determine the defect severity level based on the relationship between the comprehensive quality score and the configured preset comprehensive quality score; The confidence level of the bipolar plate classification after defect severity classification is evaluated based on the Monte Carlo Dropout method, and an output report is generated, where: Based on the Monte Carlo Dropout method, multiple forward propagations are performed in the inference phase to obtain multiple prediction results of defect classification; Obtain the probability distribution characteristics of the defect classification prediction results and determine the classification confidence based on the entropy value or variance uncertainty measurement method; The output result is determined based on the relationship between the confidence level and the configured preset confidence level: When the confidence level is lower than the preset confidence level, it is determined that a secondary evaluation of the comprehensive quality score of the bipolar plate is performed; When the confidence level is higher than or equal to the preset confidence level, it is determined that the output is graded according to the severity of the defect of the bipolar plate; When performing defect recognition on collected image data based on a multi-scale feature fusion network, it includes: Preprocess images collected under different lighting conditions to enhance image contrast and edge information; Based on the multi-scale feature extraction technology, the local detail information and global structure information of different receptive fields in the pre-processed image are extracted, and weighted fusion is performed based on the eliminated multi-scale features to obtain the comprehensive features of the processed image; Based on the comprehensive features of the fused image, a threshold comparison is performed on the characteristic responses of each area in the image to determine whether there are defects in the bipolar plate.

2. The method for detecting a flow-guiding bipolar plate for a hydrogen fuel cell according to claim 1, wherein: Preprocessing of images collected under different lighting conditions includes: Obtaining the image contrast of each region in the collected image, and obtaining the image contrast mean of each region, and comparing the image contrast mean with the contrast of each image to obtain images of each region whose contrast is lower than the image contrast mean; For each area of the image with a contrast value lower than the image mean, brightness normalization is performed based on histogram equalization; Obtain the image information after brightness normalization processing, and perform denoising on the image based on non-local mean filtering.

3. The method for detecting a flow-guiding bipolar plate for a hydrogen fuel cell according to claim 2, wherein: When obtaining the comprehensive features of the processed image, it includes: Multi-scale feature extraction technology is used on the pre-processed image to simultaneously extract local detail information and global structural information in the image based on several preset receptive fields of different sizes; The extracted multi-scale features are weighted and fused based on pre-set weights to generate comprehensive image features, where: The weights of preset scale features are set according to historical data, and the extracted scale features are matched with the preset scale features to obtain the weights of the extracted scale features. Local details and global structural information are weighted and spliced according to the weights to generate comprehensive features.

4. The method for detecting a flow-guiding bipolar plate for a hydrogen fuel cell according to claim 3, wherein: Based on the comprehensive features of the fused image, threshold comparison is performed on the feature responses of each area in the image to determine whether the bipolar plate has defects, including: Based on historical data and experimental samples, establish the correlation between detection indicators and defect responses; Obtain the distance metric between each correlation formula, and iteratively cluster each correlation formula based on the distance metric; Based on the lonely forest tree algorithm, the outliers of each cluster correlation in the clustering results are obtained, and the normal features and defect features in the same area are determined according to the relationship between the outliers and the preset outliers. Determine the defect judgment threshold based on the feature vector between the normal feature and the defect feature; Obtain the characteristic vector of the characteristic response of the region in the image, and determine whether there is a defect in the region in the image based on the relationship between the characteristic vector of the characteristic response and the defect judgment threshold: When the eigenvector of the characteristic response is lower than the defect judgment threshold, it is determined that there is no defect in the area in the image; When the eigenvector of the feature response is higher than or equal to the defect judgment threshold, it is determined that there is a defect in the area in the image.

5. The method for detecting a flow-guiding bipolar plate for a hydrogen fuel cell according to claim 1, wherein: Based on the defect data and airtightness data, combined with the attention mechanism and fuzzy logic, the comprehensive quality score of each bipolar plate is determined, including: Perform feature extraction on the defect data and airtightness data of the bipolar plate in different modes; Perform weighted analysis on defect data and airtightness data based on the attention mechanism to determine key feature information; Dynamically assign weights to defect data and airtightness data based on fuzzy logic, and determine the contribution ratio of each data in quality assessment based on the confidence level of the data; The comprehensive quality score of the bipolar plate is determined based on the weighted fusion defect score, air tightness score and morphology feature score.

6. The method for detecting a flow-guiding bipolar plate for a hydrogen fuel cell according to claim 5, wherein: Obtaining defect scores, air tightness scores, and morphology score of bipolar plates includes: Collect defect data of bipolar plates and perform defect identification based on computer vision algorithms to extract defect size, shape, and distribution characteristics and generate defect scores; Obtain air tightness data and determine the air tightness score based on leakage rate and air pressure change parameters; The surface morphology information of the bipolar plate is obtained based on three-dimensional imaging technology, and the morphology feature score is determined based on the surface flatness and deformation degree.

7. The method for detecting a flow-guiding bipolar plate for a hydrogen fuel cell according to claim 6, wherein: Determine the defect severity level based on the relationship between the comprehensive quality score and the configured preset comprehensive quality score, including: The defect severity level is determined based on the relationship between the comprehensive quality score and the configured first preset comprehensive quality score and second preset comprehensive quality score: When the comprehensive quality score is lower than the first preset comprehensive quality score, the defect severity level is determined to be high; When the comprehensive quality score is higher than or equal to the first preset comprehensive quality score and lower than the second preset comprehensive quality score, the defect severity level is determined to be medium; When the comprehensive quality score is higher than or equal to the second preset comprehensive quality score, the defect severity level is determined to be low; Among them, the first preset comprehensive quality score is less than the second preset comprehensive quality score, and the defect severity levels are ranked from high to low as high, medium and low.

8. The method for detecting a flow-guiding bipolar plate for a hydrogen fuel cell according to claim 1, wherein: The confidence level of the bipolar plate classification after defect severity classification is evaluated based on the Monte Carlo Dropout method, and an output report is generated, including: Based on the Monte Carlo Dropout method, multiple forward propagations are performed in the inference phase to obtain multiple prediction results of defect classification; Obtain the probability distribution characteristics of the defect classification prediction results and determine the classification confidence based on uncertainty measurement methods such as entropy or variance; The output result is determined based on the relationship between the confidence level and the configured preset confidence level: When the confidence level is lower than the preset confidence level, it is determined that a secondary evaluation of the comprehensive quality score of the bipolar plate is performed; When the confidence level is higher than or equal to the preset confidence level, it is determined that the output is graded according to the severity of the defect of the bipolar plate.

9. A detection system for a hydrogen fuel cell guide bipolar plate, applicable to a detection method for a hydrogen fuel cell guide bipolar plate according to any one of claims 1 to 8, characterized in that: include: A composite optical imaging unit, configured with a ring-shaped LED array light source, a near-infrared imaging module, and a structured light module, is configured to perform defect recognition on the collected image data based on a multi-scale feature fusion network; A multi-station airtightness testing device configured to perform airtightness testing on bipolar plates; an evaluation module configured to determine a comprehensive quality score for each bipolar plate based on the defect data and the airtightness data in combination with an attention mechanism and fuzzy logic; the evaluation module is further configured to determine a defect severity grade based on a relationship between the comprehensive quality score and a configured preset comprehensive quality score; The output module is configured to evaluate the confidence of the bipolar plate classification after defect severity classification based on the Monte Carlo Dropout method and generate an output report.

Images