A training method and system for 5G flexible circuit board defect detection model

Abstract

The present invention relates to the field of defect detection technology, and in particular to a training method and system for a 5G flexible circuit board defect detection model. The method comprises the following steps: acquiring multimodal 5G flexible circuit board scanning data, and performing feature enhancement based on the multimodal 5G flexible circuit board scanning data to obtain multimodal 5G flexible circuit board image enhancement data; performing cross-modal feature mapping on the multimodal 5G flexible circuit board image enhancement data to obtain a cross-modal feature fusion data set; clustering the cross-modal feature fusion data set to obtain a self-supervised feature map; performing flexible circuit board defect pattern recognition according to the self-supervised feature map to obtain flexible circuit board defect pattern data; training a 5G flexible circuit board defect detection model on the self-supervised feature map according to the flexible circuit board defect pattern data, thereby obtaining a 5G flexible circuit board defect detection model. The present invention helps to improve the accuracy of 5G flexible circuit board defect detection.

Description

Technical Field

[0001] The present invention relates to the field of defect detection technology, and in particular to a training method and system for a 5G flexible circuit board defect detection model. Background Art

[0002] With the rapid development of 5G communication technology, Flexible Printed Circuit Board (FPCB), as an important key component in 5G equipment, carries electronic components and circuits and is widely used in mobile phones, smart wearable devices, sensors and other fields. FPCB has the advantages of light weight, small space occupation and bendability, but due to its flexibility, quality control in the manufacturing process becomes more complicated, especially in high-precision production processes, various defects such as surface cracks, misalignment, circuit short circuit or open circuit are prone to occur. Traditional defect detection methods for flexible circuit boards usually rely on manual visual inspection or image processing technology based on machine vision. Although manual visual inspection is intuitive, it is limited by the recognition ability of the human eye and fatigue caused by long-term work, and the detection accuracy and efficiency are difficult to meet the needs of modern large-scale production. Although image processing technology can automate the detection process to a certain extent, due to the characteristics of flexible circuit boards such as bendability, transparency and complex circuit structure, traditional image processing methods often have difficulty in accurately identifying subtle defects. For example, surface cracks may not be effectively distinguished due to changes in the angle of illumination or the influence of background noise, and circuit misalignment and poor contact problems are also difficult to be clearly presented in the image. Summary of the invention

[0003] Based on this, it is necessary for the present invention to provide a training method and system for a 5G flexible circuit board defect detection model to solve at least one of the above technical problems.

[0004] To achieve the above object, a training method for a 5G flexible circuit board defect detection model comprises the following steps: Step S1: acquiring multimodal 5G flexible circuit board scanning data, and performing multimodal scanning data denoising based on the multimodal 5G flexible circuit board scanning data, thereby obtaining multimodal 5G flexible circuit board perception denoising data; performing multimodal image feature enhancement on the multimodal 5G flexible circuit board perception denoising data, thereby obtaining multimodal 5G flexible circuit board image enhancement data; Step S2: performing deep multimodal feature fusion on the multimodal 5G flexible circuit board image enhancement data to obtain 5G flexible circuit board feature fusion data, and performing cross-modal feature mapping based on the 5G flexible circuit board feature fusion data to obtain a cross-modal feature fusion data set; Step S3: performing self-supervised image feature extraction on the cross-modal feature fusion data set to obtain flexible circuit board image feature data, and performing self-supervised image feature clustering on the flexible circuit board image feature data to obtain a self-supervised feature map; Step S4: performing defect analysis on the flexible circuit board according to the self-supervised feature map to obtain defect feature data of the flexible circuit board, and performing defect pattern recognition on the defect feature data of the flexible circuit board to obtain defect pattern data of the flexible circuit board; Step S5: Performing a global defect assessment of the flexible circuit board on the self-supervised feature map according to the defect pattern data of the flexible circuit board, thereby obtaining the global defect data of the flexible circuit board, and training a 5G flexible circuit board defect detection model based on the global defect data of the flexible circuit board, thereby obtaining a 5G flexible circuit board defect detection model.

[0005] Optionally, step S1 specifically includes: Step S11: Acquire multi-modal 5G flexible circuit board scanning data and flexible circuit board scanning environment sensing data; Step S12: performing multi-modal scanning data denoising on the multi-modal 5G flexible circuit board scanning data according to the flexible circuit board scanning environment sensing data, thereby obtaining multi-modal scanning denoised data; Step S13: performing preliminary fusion of multimodal self-attention perception based on the multimodal scanning denoised data, thereby obtaining multimodal 5G flexible circuit board perception denoised data; Step S14: performing multimodal image visibility evaluation on the multimodal 5G flexible circuit board perception denoising data, thereby obtaining multimodal image visibility data; Step S15: Perform multimodal image feature enhancement on the multimodal 5G flexible circuit board perception denoising data according to the multimodal image visibility data, so as to obtain multimodal 5G flexible circuit board image enhancement data.

[0006] Optionally, step S12 is specifically: Performing data alignment on the multi-modal 5G flexible circuit board scanning data and the flexible circuit board scanning environment sensing data, thereby obtaining the multi-modal 5G flexible circuit board scanning alignment data and the flexible circuit board scanning environment sensing alignment data; Performing scale conversion on the multimodal 5G flexible circuit board scanning alignment data, thereby obtaining the multimodal 5G flexible circuit board scanning data to be processed, wherein the multimodal 5G flexible circuit board scanning data to be processed includes a 5G flexible circuit board camera image, a 5G flexible circuit board infrared image, and a 5G flexible circuit board ultrasonic image; Multimodal data noise source identification is performed based on the multimodal 5G flexible circuit board scanning data to be processed, thereby obtaining image noise source data, ultrasonic signal noise source data, and infrared noise source data; The influence of the scanning environment illumination is quantified according to the scanning environment sensor alignment data of the flexible circuit board and the image noise source data, so as to obtain the environmental illumination influence factor; The influence of the scanning environment temperature is quantified according to the scanning environment sensor alignment data of the flexible circuit board and the infrared noise source data, so as to obtain the environmental temperature influence factor; The impact of the scanning environment signal is quantified according to the flexible circuit board scanning environment sensor alignment data and the ultrasonic signal noise source data, so as to obtain the environmental signal impact factor; Based on the environmental lighting influence factor, the 5G flexible circuit board camera image is subjected to median filtering and denoising, thereby obtaining a 5G flexible circuit board camera denoised image; Based on the environmental temperature influencing factor, the infrared scanning wavelet transform is used to denoise the infrared image of the 5G flexible circuit board, thereby obtaining the infrared scanning denoised image of the 5G flexible circuit board; Based on the environmental signal influencing factors, the ultrasonic signal Kalman denoising is performed on the ultrasonic image of the 5G flexible circuit board, so as to obtain the denoised image of the ultrasonic detection of the 5G flexible circuit board; The 5G flexible circuit board camera denoised image, the 5G flexible circuit board infrared scanning denoised image and the 5G flexible circuit board ultrasonic detection denoised image are merged to obtain multi-modal scanning denoised data.

[0007] Optionally, step S15 is specifically: Step S151: screening the multimodal 5G flexible circuit board perception denoising data for low-visibility images according to the multimodal image visibility data, thereby obtaining a multimodal low-visibility image set; Step S152: performing local contrast enhancement on the multimodal low-visibility image set, thereby obtaining a multimodal low-visibility local contrast enhanced image set; Step S153: performing multimodal convolutional layer feature sharing on the multimodal low-visibility local contrast enhanced image set, thereby obtaining a multimodal feature sharing image set; Step S154: performing image noise suppression and edge enhancement according to the multimodal feature sharing image set, thereby obtaining a low-visibility enhanced image set; Step S155: Perform image enhancement feature coupling on the multimodal 5G flexible circuit board perception denoising data based on the low-visibility enhanced image set, so as to obtain multimodal 5G flexible circuit board image enhancement data.

[0008] Optionally, step S2 specifically includes: Step S21: performing modal depth feature extraction on the multi-modal 5G flexible circuit board image enhancement data, so as to obtain camera modal image feature data, thermal imaging modal image feature data and ultrasonic modal image feature data; Step S22: performing modal feature contribution evaluation according to the camera modal image feature data, the thermal imaging modal image feature data, and the ultrasonic modal image feature data, so as to obtain modal feature contribution data; Step S23: performing multi-modal data weighted fusion on the camera modal image feature data, the thermal imaging modal image feature data and the ultrasonic modal image feature data based on the modal feature contribution data, so as to obtain 5G flexible circuit board feature fusion data; Step S24: performing cross-modal feature space mapping on the 5G flexible circuit board feature fusion data, thereby obtaining a cross-modal mapping feature space; Step S25: performing mapping feature splicing according to the cross-modal mapping feature space to obtain a cross-modal feature fusion data set.

[0009] Optionally, step S22 is specifically: Step S221: performing Spearman rank correlation coefficient calculation on the imaging modality image feature data and the thermal imaging modality image feature data, thereby obtaining imaging feature-thermal imaging feature correlation coefficient data; performing Spearman rank correlation coefficient calculation on the imaging modality image feature data and the ultrasonic modality image feature data, thereby obtaining imaging feature-ultrasonic image feature correlation coefficient data; Step S222: sorting the correlation coefficients according to the imaging feature-thermal imaging feature correlation coefficient data and the imaging feature-ultrasound image feature correlation coefficient data, thereby obtaining multimodal image feature correlation coefficient sorting data; Step S223: performing image feature independence variance analysis based on the multimodal image feature correlation coefficient sorting data, thereby obtaining multimodal image feature independence data; Step S224: performing feature complementarity evaluation and feature redundancy evaluation based on the multimodal image feature independence data, thereby obtaining modal feature complementarity data and modal feature redundancy data; Step S225: Visualize the modal feature contribution according to the modal feature complementarity data and the modal feature redundancy data, so as to obtain the modal feature contribution data.

[0010] Optionally, step S3 specifically includes: Step S31: performing sample division on the cross-modal feature fusion data, thereby obtaining similar sample data and heterogeneous sample data; Step S32: maximizing positive sample learning on sample data of the same class, thereby obtaining similarity data of samples of the same class; minimizing negative sample learning on sample data of different classes, thereby obtaining similarity data of samples of different classes; Step S33: performing feature representation learning based on the similarity data of samples of the same category and the similarity data of samples of different categories, thereby obtaining an image feature representation space; Step S34: extracting image features from the cross-modal feature fusion data set based on the image feature representation space, thereby obtaining image feature data of the flexible circuit board; Step S35: performing image feature clustering calculation on the flexible circuit board image feature data to obtain circuit board image cluster feature data, and performing self-supervised feature label visualization on the circuit board image cluster feature data to obtain a self-supervised feature map.

[0011] Optionally, step S4 specifically includes: Step S41: performing local outlier factor detection on the self-supervised feature map to obtain the local outlier factor of the circuit board; Step S42: integrating outlier factor features according to the local outlier factors of the circuit board, thereby obtaining local defect feature data of the circuit board; Step S43: performing linear discriminant analysis based on the local defect feature data of the circuit board, thereby obtaining main feature data of the local defects of the flexible circuit board; Step S44: localizing the defect area of ​​the flexible circuit board using the cross-modal feature fusion data according to the main feature data of the local defect of the flexible circuit board, thereby obtaining feature data of the local defect area of ​​the flexible circuit board; Step S45: integrating the main characteristic data of the local defect of the flexible circuit board and the characteristic data of the local defect area of ​​the flexible circuit board with the defect characteristics of the flexible circuit board, thereby obtaining the defect characteristic data of the flexible circuit board; Step S46: performing FPCB defect pattern recognition on the FPCB defect feature data, thereby obtaining FPCB defect pattern data.

[0012] Optionally, step S42 is specifically: Step S421: mapping the self-supervised feature map to an outlier factor score based on the local outlier factor of the circuit board, thereby obtaining outlier factor score data of the sample point; Step S422: performing feature combination on the sample point outlier factor score data and the self-supervised feature map to obtain comprehensive feature data of the flexible circuit board image; Step S423: performing outlier factor score feature enhancement according to the flexible circuit board image comprehensive feature data, thereby obtaining flexible circuit board image enhanced feature data; Step S424: performing a weighted evaluation of the local abnormality of the circuit board image features on the flexible circuit board image enhanced feature data, thereby obtaining the local abnormal value data of the circuit board image features; Step S425: integrating high outlier value local image features according to the local outlier value data of the circuit board image features, thereby obtaining the local defect feature data of the circuit board.

[0013] The present invention can not only effectively reduce the impact of environmental noise on image quality, but also improve the accuracy of subsequent image processing by acquiring multimodal 5G flexible circuit board scanning data and performing denoising on it. By performing multimodal image feature enhancement on the denoised data, the visibility of the image can be improved in a complex industrial environment, ensuring that the defect information can be clearly expressed. This process enhances the visibility and reliability of the data and provides high-quality input data for subsequent image feature fusion and analysis. In the deep multimodal feature fusion stage, the features of different modalities (such as camera images, thermal imaging images, and ultrasonic images) are fused, which can not only maximize the data advantages of different modalities, but also reduce the limitations of a single modality, thereby improving the accuracy and robustness of the overall detection system. In particular, through cross-modal feature mapping, a common feature space can be found in data of different modalities, which provides strong support for subsequent data processing and analysis, so that data of multiple modalities can be optimized in the same computing framework, further improving the efficiency of detection. Self-supervised image feature extraction and clustering are one of the key steps of this solution. Self-supervised learning can effectively extract potential and representative features from complex image data by learning a large amount of unlabeled data, especially in the scenario of large-scale data sets. This method can greatly reduce the cost and time of manual annotation. At the same time, self-supervised image feature clustering can help the system automatically identify the inherent rules in the data and cluster similar samples together, providing strong support for the identification and classification of defect patterns. This automated feature map generation method not only improves the efficiency of the system, but also enables the system to perform large-scale defect analysis without human intervention, improving the intelligence level of the system. Defect analysis based on self-supervised feature maps can perform more accurate defect identification based on these maps. This process extracts and analyzes the characteristics of the circuit board to identify potential defect areas, laying the foundation for subsequent defect pattern recognition. Through pattern recognition of defect features of flexible circuit boards, not only existing defects can be detected, but also some minor defects or potential problems that may not be easily discovered by traditional methods can be effectively identified. Through global defect evaluation of self-supervised feature maps, a comprehensive judgment of the overall health of the circuit board can be provided, possible global defects can be identified, and the defect detection model can be further trained based on these defect information to improve the generalization ability and prediction accuracy of the model. The defect detection model trained based on the global defect data of flexible circuit boards can automatically identify new defect types and adapt to a variety of industrial environments. This deep learning-based model training method enables the system to be continuously optimized over time, gradually improving the accuracy of the model by adding new training samples.Finally, the generated 5G flexible circuit board defect detection model has powerful real-time monitoring capabilities, and can achieve efficient defect detection on the production line, which not only improves production efficiency, but also effectively reduces manual intervention and improves the automation and intelligence level of the production process. Through the above series of steps, the present invention realizes the optimization of the entire chain from data acquisition to defect identification, which not only improves the accuracy, efficiency and robustness of the detection system, but also significantly reduces labor costs and dependence on data labeling through the application of deep learning technology and self-supervised learning, making the detection system more intelligent and automated. At the same time, the combination of cross-modal feature fusion and self-supervised learning enables the system to fully utilize the data advantages of multiple modalities, and in the process of continuous learning and optimization, it has the ability to adapt to complex environments and changing working conditions, greatly improving the detection capability of 5G flexible circuit boards.

[0014] Optionally, the present specification also provides a training system for a 5G flexible circuit board defect detection model, which is used to execute the training method for the 5G flexible circuit board defect detection model as described above, and the training system for the 5G flexible circuit board defect detection model includes: A multimodal image preprocessing module is used to obtain multimodal 5G flexible circuit board scanning data, and perform multimodal scanning data denoising based on the multimodal 5G flexible circuit board scanning data, so as to obtain multimodal 5G flexible circuit board perception denoising data; perform multimodal image feature enhancement on the multimodal 5G flexible circuit board perception denoising data, so as to obtain multimodal 5G flexible circuit board image enhancement data; A cross-modal feature fusion module is used to perform deep multi-modal feature fusion on the multi-modal 5G flexible circuit board image enhancement data, thereby obtaining 5G flexible circuit board feature fusion data, and perform cross-modal feature mapping based on the 5G flexible circuit board feature fusion data, thereby obtaining a cross-modal feature fusion data set; A self-supervised feature extraction module is used to perform self-supervised image feature extraction on the cross-modal feature fusion data set to obtain flexible circuit board image feature data, and perform self-supervised image feature clustering on the flexible circuit board image feature data to obtain a self-supervised feature map; A defect mode analysis module is used to perform defect analysis of the flexible circuit board according to the self-supervised feature map, thereby obtaining defect feature data of the flexible circuit board, and perform defect mode recognition of the flexible circuit board on the defect feature data of the flexible circuit board, thereby obtaining defect mode data of the flexible circuit board; The model training module is used to perform global defect evaluation of the flexible circuit board on the self-supervised feature map according to the defect pattern data of the flexible circuit board, so as to obtain the global defect data of the flexible circuit board, and to train the 5G flexible circuit board defect detection model based on the global defect data of the flexible circuit board, so as to obtain the 5G flexible circuit board defect detection model.

[0015] The 5G flexible circuit board defect detection system of the present invention can implement any one of the 5G flexible circuit board defect detection methods of the present invention, and is used to combine the operation and signal transmission medium between various modules to complete the 5G flexible circuit board defect detection method. The internal modules of the system cooperate with each other, thereby helping to improve the accuracy of 5G flexible circuit board defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments thereof made with reference to the following drawings: Figure 1 The figure is a flowchart of the steps of the training method for the 5G flexible circuit board defect detection model of the present invention; Figure 2 Detailed step flow diagram of step S1 in the present invention; Figure 3 Detailed step flow diagram of step S2 in the present invention; The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0017] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0018] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0019] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0020] To achieve this, please refer to Figures 1 to 3 The present invention provides a training method for a 5G flexible circuit board defect detection model, the method comprising the following steps: Step S1: acquiring multimodal 5G flexible circuit board scanning data, and performing multimodal scanning data denoising based on the multimodal 5G flexible circuit board scanning data, thereby obtaining multimodal 5G flexible circuit board perception denoising data; performing multimodal image feature enhancement on the multimodal 5G flexible circuit board perception denoising data, thereby obtaining multimodal 5G flexible circuit board image enhancement data; In this embodiment, multimodal 5G flexible circuit board scanning data is obtained, including image data collected by different sensors, such as visible light images, infrared images, and ultrasonic images. These image data are subjected to multimodal scanning data denoising to remove noise and interference during the scanning process, and image denoising algorithms, such as convolutional neural networks (CNN) or wavelet transforms, are used to reduce artifacts and unnecessary details in the image. Through this process, the obtained multimodal 5G flexible circuit board perception denoising data is clearer in visual effect and can effectively improve the accuracy of subsequent processing. Next, multimodal image feature enhancement is performed on these perception denoising data, and feature enhancement techniques such as contrast stretching or edge enhancement are used to make the key areas in the image more prominent, especially the detail enhancement of defects or abnormal areas, which is helpful for the accurate extraction of important features in subsequent analysis.

[0021] Step S2: performing deep multimodal feature fusion on the multimodal 5G flexible circuit board image enhancement data to obtain 5G flexible circuit board feature fusion data, and performing cross-modal feature mapping based on the 5G flexible circuit board feature fusion data to obtain a cross-modal feature fusion data set; In this embodiment, deep multimodal feature fusion is performed on the multimodal 5G flexible circuit board image enhancement data. Using deep learning algorithms such as convolutional neural networks (CNN) or bidirectional long short-term memory networks (Bi-LSTM), data from different sensors are fused into a unified feature vector. This fusion not only takes into account the characteristics of each modality, but also can utilize the complementary information between different modalities to enhance the perception of surface defects and deformations of the flexible circuit board. Through these fused features, feature fusion data of the 5G flexible circuit board is constructed. Subsequently, cross-modal feature mapping is performed based on these data, and a multimodal mapping network (such as an autoencoder or a deep generative model) is used to learn the mapping relationship between different modalities, and finally a cross-modal feature fusion data set is generated, which can process data from different sources at the same time and obtain comprehensive analysis results.

[0022] Step S3: performing self-supervised image feature extraction on the cross-modal feature fusion data set to obtain flexible circuit board image feature data, and performing self-supervised image feature clustering on the flexible circuit board image feature data to obtain a self-supervised feature map; In this embodiment, self-supervised image feature extraction is performed based on a cross-modal feature fusion data set. Through self-supervised learning technology, representative features are automatically extracted from image data without manual annotation. These features include low-level features such as color, texture, shape, and high-level features extracted based on deep learning, such as structural information of local areas. By performing self-supervised image feature clustering on the flexible circuit board image feature data, a clustering algorithm (such as K-means or DBSCAN) is applied to divide the features into several groups, each group contains similar image features, and the clustering results of these features form a self-supervised feature map. The map can show the feature distribution of different areas in the image and provide a basis for subsequent defect identification. Through this map, the system can identify different image feature patterns, thereby effectively classifying and detecting defects.

[0023] Step S4: performing defect analysis on the flexible circuit board according to the self-supervised feature map to obtain defect feature data of the flexible circuit board, and performing defect pattern recognition on the defect feature data of the flexible circuit board to obtain defect pattern data of the flexible circuit board; In this embodiment, defect analysis of the flexible circuit board is performed based on the self-supervised feature map. By analyzing the distribution of each feature in the map, characteristic patterns representing different defects, such as cracks, scratches, deformations, etc., are identified. These patterns are classified and trained using image classification algorithms (such as convolutional neural networks CNN or support vector machines SVM) so that the model can learn to identify different types of defects. During the training process, the algorithm will learn how to distinguish different types of defects from image features, and eventually form an efficient defect recognition system. In addition, the defect pattern of the flexible circuit board will be identified by a pattern recognition algorithm (such as a multi-classification network of a deep neural network), so that the system can automatically determine the type and severity of the defect.

[0024] Step S5: Performing a global defect assessment of the flexible circuit board on the self-supervised feature map according to the defect pattern data of the flexible circuit board, thereby obtaining the global defect data of the flexible circuit board, and training a 5G flexible circuit board defect detection model based on the global defect data of the flexible circuit board, thereby obtaining a 5G flexible circuit board defect detection model.

[0025] In this embodiment, the self-supervised feature map is further evaluated for global defects based on the defect pattern data of the flexible circuit board. Through the global evaluation method, the local defect information is summarized and the overall health status of the circuit board is considered. The local defect information is integrated using image fusion algorithms, such as weighted average or maximum pooling, to generate global defect data. Based on these global defect data, the quality and status of the entire circuit board can be evaluated, and high-quality training data can be provided for subsequent defect detection model training. Finally, these global defect data are used to train the 5G flexible circuit board defect detection model, and technologies such as convolutional neural networks (CNN) or deep generative adversarial networks (GAN) are used to train a deep learning model that can automatically identify and detect different types of defects. The trained model will have high accuracy and robustness, and will be able to detect defects on flexible circuit boards in real time in practical applications, and effectively locate and classify them.

[0026] Optionally, step S1 specifically includes: Step S11: Acquire multi-modal 5G flexible circuit board scanning data and flexible circuit board scanning environment sensing data; In this embodiment, the 5G flexible circuit board is scanned by multimodal sensor devices (such as high-resolution cameras, infrared thermal imagers, and ultrasonic sensors). These devices collect visible light images, thermal imaging images, and ultrasonic signals for waveform decomposition and imaging, thereby obtaining ultrasonic images, ensuring comprehensive perception of different dimensions of the circuit board (such as surface defects, temperature changes, and structural integrity), and then the ultrasonic signal. In addition, combined with the flexible circuit board scanning environment sensors (such as light sensors, temperature and humidity sensors, etc.), data related to the scanning environment is obtained to ensure that subsequent processing can take into account the impact of environmental factors, especially the impact on image quality and signal accuracy. These multimodal data and environmental data will be used in subsequent processing steps to improve the accuracy of defect detection.

[0027] Step S12: performing multi-modal scanning data denoising on the multi-modal 5G flexible circuit board scanning data according to the flexible circuit board scanning environment sensing data, thereby obtaining multi-modal scanning denoised data; In this embodiment, the analysis of the flexible circuit board scanning environment sensor data is combined with the multimodal 5G flexible circuit board scanning data, and a denoising algorithm (such as an autoencoder network or filter technology based on deep learning) is applied to denoise the scanned data. The denoising algorithm automatically adjusts the denoising degree according to parameters such as light and temperature in the scanning environment to remove invalid data caused by environmental interference or sensor noise. This process can not only effectively reduce the impact of noise on image and signal quality, but also improve the accuracy of subsequent feature extraction and defect analysis, making the data more consistent with the actual situation.

[0028] Step S13: performing preliminary fusion of multimodal self-attention perception based on the multimodal scanning denoised data, thereby obtaining multimodal 5G flexible circuit board perception denoised data; In this embodiment, a multimodal self-attention mechanism is used to perform preliminary fusion of the denoised data. By designing a multimodal self-attention network, the network can perform adaptive weighting and feature fusion for different modalities (such as visible light images, infrared images, ultrasonic images, etc.). The key to this process is the self-attention mechanism, which can dynamically adjust the weights of different modal features so that important modal features dominate the final output. Specifically, through the self-attention module, the network can automatically focus on key areas in the image, such as cracks or deformation areas, while suppressing irrelevant background noise, thereby achieving intelligent fusion of multimodal data and improving the recognition effect of flexible circuit board defects.

[0029] Step S14: performing multimodal image visibility evaluation on the multimodal 5G flexible circuit board perception denoising data, thereby obtaining multimodal image visibility data; In this embodiment, the perceptual denoising data is analyzed using multimodal image visibility assessment technology. Specifically, based on the image's brightness, contrast, texture and other features, an algorithm combined with image quality assessment (such as the structural similarity index SSIM or the perceptual quality assessment network) is used to evaluate the visibility of each image. This evaluation process helps the algorithm identify which image areas have higher visibility and which areas may be affected by environmental interference or blur by comprehensively scoring the quality of images in different modalities. This data can provide a basis for subsequent image enhancement steps to ensure that the key areas of enhancement are those that are critical to defect detection.

[0030] Step S15: Perform multimodal image feature enhancement on the multimodal 5G flexible circuit board perception denoising data according to the multimodal image visibility data, so as to obtain multimodal 5G flexible circuit board image enhancement data.

[0031] In this embodiment, based on the multimodal image visibility data, the 5G flexible circuit board perception denoising data is feature enhanced. Specifically, an image processing-based enhancement method, such as histogram equalization, contrast stretching, and edge enhancement, is used in combination with a deep learning method (such as the enhancement module in the convolutional neural network CNN) to enhance the image features. This enhancement process pays special attention to low-visibility areas, such as cracks that may be obscured, tiny welding defects, etc. Through feature enhancement, the defect features in the image can be enhanced, making these enhanced features more prominent in the subsequent defect detection steps, thereby helping to improve the accuracy and robustness of detection.

[0032] Optionally, step S12 is specifically: Performing data alignment on the multi-modal 5G flexible circuit board scanning data and the flexible circuit board scanning environment sensing data, thereby obtaining the multi-modal 5G flexible circuit board scanning alignment data and the flexible circuit board scanning environment sensing alignment data; In this embodiment, multimodal data is obtained from multiple sensors, including cameras, infrared sensors, ultrasonic sensors, and environmental sensors (such as temperature and humidity, light intensity, airflow, etc.). These data come from different devices, so they need to be aligned in time and space. By using the alignment algorithm of the timestamp and sensor coordinate system, it is ensured that the data of different modalities are matched at the same time and space position. For example, the camera image and the infrared image must correspond to the data taken at the same time, and the ultrasonic image also needs to be synchronized with the first two. After the alignment is completed, we obtained the multimodal 5G flexible circuit board scanning alignment data and the flexible circuit board scanning environmental sensing alignment data as the basis for subsequent data processing.

[0033] Performing scale conversion on the multimodal 5G flexible circuit board scanning alignment data, thereby obtaining the multimodal 5G flexible circuit board scanning data to be processed, wherein the multimodal 5G flexible circuit board scanning data to be processed includes a 5G flexible circuit board camera image, a 5G flexible circuit board infrared image, and a 5G flexible circuit board ultrasonic image; In this embodiment, a unified conversion is performed for the scale differences of different modal data. Taking camera images, infrared images, and ultrasonic images as examples, the image resolutions and acquisition scales of these three are different. Use an image interpolation algorithm (such as bilinear interpolation or bicubic interpolation) to uniformly convert each modality image to the same size and resolution. For example, the camera image is converted to the same resolution as the infrared image, or the size of the ultrasonic image is unified to the size of the camera image. After the scale conversion, the multi-modal 5G flexible circuit board scanning data to be processed is obtained, which includes camera images, infrared images, and ultrasonic images. These data have consistent size and resolution and are suitable for subsequent processing.

[0034] Multimodal data noise source identification is performed based on the multimodal 5G flexible circuit board scanning data to be processed, thereby obtaining image noise source data, ultrasonic signal noise source data, and infrared noise source data; In this embodiment, when processing the multimodal data to be processed, the noise sources in each modal image are identified. By analyzing the texture, brightness change and other information of the image, the illumination noise in the camera image, the temperature change noise in the infrared image and the signal attenuation or echo noise in the ultrasonic image can be identified. The specific method includes using image processing algorithms such as edge detection, frequency domain filtering and other technologies to extract noise features from the image. Ultrasonic image noise source identification can be completed by signal amplitude analysis and time window filtering technology. After identifying these noise sources, image noise source data, ultrasonic signal noise source data and infrared noise source data are obtained, which provides necessary information for the next step of environmental impact quantification and denoising.

[0035] The influence of the scanning environment illumination is quantified according to the scanning environment sensor alignment data of the flexible circuit board and the image noise source data, so as to obtain the environmental illumination influence factor; In this embodiment, the lighting data provided by the environmental sensor is used to quantify the lighting effect on the camera image. Lighting has a significant impact on the quality of the camera image. Too strong or too weak lighting will cause the image to be overexposed or underexposed. By comparing and analyzing the environmental sensor data (such as light intensity, angle, etc.), the lighting effect factor on the image is calculated. Specifically, a lighting effect model can be established based on the relationship between the brightness and contrast of the image and the ambient light intensity, thereby quantifying the ambient light effect factor. This factor will be used for subsequent denoising of the camera image, which will help reduce the interference of uneven lighting.

[0036] The influence of the scanning environment temperature is quantified according to the scanning environment sensor alignment data of the flexible circuit board and the infrared noise source data, so as to obtain the environmental temperature influence factor; In this embodiment, the temperature noise in the infrared image needs to be analyzed based on the temperature data provided by the environmental sensor. Infrared images are usually affected by changes in ambient temperature. Temperature fluctuations may cause unclear infrared imaging or deviations in the temperature distribution of the image. By matching the temperature sensor data with the infrared image, the impact of temperature changes on the infrared image can be quantified. In specific implementation, a temperature calibration method can be used to establish a mapping relationship between ambient temperature and image temperature, thereby calculating the temperature influence factor. This factor will be used in subsequent infrared image denoising to ensure that the final image can truly reflect the temperature distribution of the target area.

[0037] The impact of the scanning environment signal is quantified according to the flexible circuit board scanning environment sensor alignment data and the ultrasonic signal noise source data, so as to obtain the environmental signal impact factor; In this embodiment, the quality of the ultrasonic image is often affected by the propagation characteristics of the sound waves in the scanning environment. The stray signals, reflections, attenuation, etc. in the environment will cause image noise. By analyzing the signal data (such as air density, humidity, etc.) provided by the environmental sensor and the ultrasonic signal noise source data, the impact of the environmental signal on the ultrasonic image can be quantified. Specifically, by establishing an environmental signal propagation model, the attenuation and interference of each ultrasonic signal under specific environmental conditions can be calculated to obtain the environmental signal impact factor. This factor will be used for subsequent ultrasonic image denoising.

[0038] Based on the environmental lighting influence factor, the 5G flexible circuit board camera image is subjected to median filtering and denoising, thereby obtaining a 5G flexible circuit board camera denoised image; In this embodiment, in the denoising process of the camera image, the ambient light influence factor of the camera image is used to guide the median filtering operation. Median filtering is a commonly used denoising method that can effectively remove salt and pepper noise in the image. On this basis, by adjusting the window size and weight of the filter, the filtering process is optimized according to the ambient light influence factor to eliminate the noise introduced by insufficient or excessive light. After such processing, the obtained camera image is clearer, the details are more prominent, and the image quality is enhanced.

[0039] Based on the environmental temperature influencing factor, the infrared scanning wavelet transform is used to denoise the infrared image of the 5G flexible circuit board, thereby obtaining the infrared scanning denoised image of the 5G flexible circuit board; In this embodiment, infrared image denoising uses wavelet transform technology to decompose the image into frequency components of different scales, and remove noise according to the ambient temperature influencing factor. Wavelet transform can effectively extract detail information from the image and suppress noise. In practical applications, the number of decomposition layers and thresholds of wavelet transform are adjusted according to the temperature influencing factor to better remove the noise caused by ambient temperature changes. After denoising, the temperature distribution of the infrared image is smoother and clearer.

[0040] Based on the environmental signal influencing factors, the ultrasonic signal Kalman denoising is performed on the ultrasonic image of the 5G flexible circuit board, so as to obtain the denoised image of the ultrasonic detection of the 5G flexible circuit board; In this embodiment, Kalman filtering technology is used to denoise the ultrasonic image. Kalman filtering can dynamically estimate and correct the ultrasonic signal according to the environmental signal influencing factors, and eliminate the noise caused by environmental interference. In practical applications, the parameters of the Kalman filter are adjusted according to the propagation characteristics of the ultrasonic signal and the environmental influencing factors, so that the filtering process is more accurate and a clear ultrasonic image is finally obtained.

[0041] The 5G flexible circuit board camera denoised image, the 5G flexible circuit board infrared scanning denoised image and the 5G flexible circuit board ultrasonic detection denoised image are merged to obtain multi-modal scanning denoised data.

[0042] In this embodiment, after all modal images are denoised, the three images are merged. The specific method can use weighted fusion technology to fuse the key information of camera images, infrared images and ultrasonic images. By reasonably adjusting the weight of each image, it is ensured that the useful information of each modality is fully retained. Finally, the merged multi-modal scanning denoised data provides complete and accurate data support for subsequent defect detection and analysis.

[0043] Optionally, step S15 is specifically: Step S151: screening the multimodal 5G flexible circuit board perception denoising data for low-visibility images according to the multimodal image visibility data, thereby obtaining a multimodal low-visibility image set; In this embodiment, images with low visibility are screened out based on the obtained multimodal image visibility data. First, the image contrast, brightness value and detail information are used to quantify the visibility of the image. For example, the visual quality of each image is evaluated by calculating the global brightness mean and local contrast of each image. If the contrast of the image is lower than the set threshold, it is determined to be a low-visibility image and these images are screened out. The screening process can also be adjusted in combination with the image noise level. When the noise is high and the details are blurred, it will also be classified as a low-visibility image. Finally, a multimodal low-visibility image set is obtained, covering multiple modalities such as camera images, infrared images and ultrasonic images.

[0044] Step S152: performing local contrast enhancement on the multimodal low-visibility image set, thereby obtaining a multimodal low-visibility local contrast enhanced image set; In this embodiment, the local contrast enhancement technology is used to process the selected low-visibility image set. Specifically, the Local Histogram Equalization (LHE) method is used to enhance the details in the image. By calculating the contrast of each local area and adaptively adjusting it, the dark and bright details of the image are made more prominent. For example, in infrared images, some areas appear blurred due to insufficient light. Through local contrast enhancement, the temperature differences in these areas can be highlighted and the image quality can be improved. When processing ultrasonic images, by enhancing the contrast of local areas, the details of the echo signal are clearer, which is convenient for subsequent analysis and detection. After processing, the resulting image set will have higher local visibility and richer details.

[0045] Step S153: performing multimodal convolutional layer feature sharing on the multimodal low-visibility local contrast enhanced image set, thereby obtaining a multimodal feature sharing image set; In this embodiment, for the low-visibility image set (images after local contrast enhancement processing), a multimodal convolutional neural network (CNN) is used for feature sharing. By using a shared convolutional layer structure, images of multiple modalities can be feature extracted under the same network structure, thereby effectively extracting common features in each modality image. For example, camera images, infrared images, and ultrasonic images can share the convolution kernel in the convolution layer to learn some universal low-level features, such as edges, textures, and shapes. By sharing the convolution layer, not only can the efficiency of feature extraction be improved, but also the number of network parameters can be reduced to avoid overfitting. In addition, the shared weights in the convolution layer enable data of different modalities to be better jointly learned, thereby improving the fusion effect between multimodal data. Finally, the obtained multimodal feature sharing image set contains the fusion features of multiple modal images, providing more comprehensive information.

[0046] Step S154: performing image noise suppression and edge enhancement according to the multimodal feature sharing image set, thereby obtaining a low-visibility enhanced image set; In this embodiment, based on the multimodal feature sharing image set, the image is first subjected to noise suppression processing to reduce high-frequency noise in low-visibility images. Noise suppression can be achieved by using techniques such as denoising convolutional networks or Gaussian blur filtering to remove unnecessary noise components in the image. During the processing, special attention is paid to retaining the edge information of the image, because the edge is crucial for subsequent image analysis and defect detection. Next, an edge enhancement algorithm, such as the Laplace operator or Sobel edge detection, is used to amplify the edge information in the image to make the image outline clearer. This processing helps to clearly distinguish the target area and the background in the image, improve the readability of the image and the accuracy of subsequent processing. The image set after noise suppression and edge enhancement processing will show clearer and more refined features, providing a more reliable data basis for subsequent detection and analysis.

[0047] Step S155: Perform image enhancement feature coupling on the multimodal 5G flexible circuit board perception denoising data based on the low-visibility enhanced image set, so as to obtain multimodal 5G flexible circuit board image enhancement data.

[0048] In this embodiment, image enhancement feature coupling is performed based on a low-visibility enhanced image set. First, low-level features (such as edges and textures) and high-level features (such as shapes and object structures) in the image are combined, and feature fusion methods such as weighted averaging or feature splicing are used to effectively fuse the data and enhanced features of the multimodal image. By coupling the image enhancement features, not only can the image quality be further improved, but also the complementarity between images of different modalities can be enhanced. For example, in an ultrasonic image, the clarity of the signal may be affected by environmental noise, but by coupling with the features of a camera image or an infrared image, the missing details can be filled in and the overall quality of the image can be improved. Finally, the obtained multimodal 5G flexible circuit board image enhancement data contains rich feature information, which can provide more refined and clear data support for subsequent detection tasks (such as defect detection, structural analysis, etc.).

[0049] Optionally, step S2 specifically includes: Step S21: performing modal depth feature extraction on the multi-modal 5G flexible circuit board image enhancement data, so as to obtain camera modal image feature data, thermal imaging modal image feature data and ultrasonic modal image feature data; In this embodiment, a deep convolutional neural network (CNN) is used to extract features from the multimodal 5G flexible circuit board image enhancement data. Specifically, pre-trained deep network models are used to extract features for camera images, thermal imaging images, and ultrasonic images. For example, for camera images, ResNet or VGG networks can be used for feature extraction to capture the edges, textures, and high-level shape information in the image; for thermal imaging images, a convolutional network that specifically processes infrared images is used to extract the changing features of the temperature distribution; and for ultrasonic images, an improved U-Net structure can be used to extract abnormal patterns in the echo signal. The results of these deep feature extractions will generate camera modality image feature data, thermal imaging modality image feature data, and ultrasonic modality image feature data, respectively, providing detailed feature information of different modalities for subsequent data analysis.

[0050] Step S22: performing modal feature contribution evaluation according to the camera modal image feature data, the thermal imaging modal image feature data, and the ultrasonic modal image feature data, so as to obtain modal feature contribution data; In this embodiment, based on the camera modality image feature data, thermal imaging modality image feature data and ultrasonic modality image feature data, a contribution evaluation method is used to quantify the features of different modality images. Specifically, a weighted feature importance analysis technique, such as a feature importance evaluation based on a random forest, is used to measure the contribution of each modality to the final goal (such as defect detection, structural recognition, etc.). By comparing the features of each modality and evaluating its contribution to accuracy, the modality feature contribution data is obtained. For example, if ultrasonic images have a greater weight in defect detection, the contribution of this modality will be higher. The evaluation results help to reasonably weight the features of different modalities in the subsequent fusion step.

[0051] Step S23: performing multi-modal data weighted fusion on the camera modal image feature data, the thermal imaging modal image feature data and the ultrasonic modal image feature data based on the modal feature contribution data, so as to obtain 5G flexible circuit board feature fusion data; In this embodiment, the image features of different modalities are weighted fused according to the feature contribution data of each modality. In specific implementation, the camera modality image feature data, thermal imaging modality image feature data, and ultrasonic modality image feature data can be fused into a comprehensive feature set by weighted averaging or weighted summing based on the obtained modality feature contribution data. For example, if the camera modality contribution is low, the features of the camera image can be given a lower weight during fusion, while the features of the ultrasonic image and the thermal imaging image will be given a higher weight. The fused data will be able to better represent the dominant features of each modality and improve the recognition ability of multimodal images. This weighted fusion feature will serve as the basis for subsequent processing and can enhance the accurate perception and detection capabilities of 5G flexible circuit boards.

[0052] Step S24: performing cross-modal feature space mapping on the 5G flexible circuit board feature fusion data, thereby obtaining a cross-modal mapping feature space; In this embodiment, cross-modal feature space mapping is performed based on the obtained 5G flexible circuit board feature fusion data. Specifically, feature space mapping technology, such as principal component analysis (PCA) or t-SNE and other methods, is used to project the fusion features of each modality into a unified high-dimensional feature space. When mapping, a cross-modal alignment algorithm is used to ensure that the feature data of different modalities can be in the same semantic space after mapping, thereby maintaining the effective relevance of each modal data. For example, after mapping the features of camera images, thermal imaging images, and ultrasonic images into a unified space, the differences between the modalities will be compressed to a minimum, making subsequent image matching, defect recognition and other tasks more accurate and efficient. Feature space mapping can not only reduce redundant information between modalities, but also improve the effect of cross-modal fusion.

[0053] Step S25: performing mapping feature splicing according to the cross-modal mapping feature space to obtain a cross-modal feature fusion data set.

[0054] In this embodiment, feature splicing is performed based on the output of the cross-modal mapping feature space. Feature splicing is to splice the mapping results of different modalities (camera images, thermal imaging images, and ultrasonic images) in a unified feature space to form a richer set of feature vectors. Specifically, by horizontally splicing the mapped feature vectors, the high-dimensional feature data of each modality is connected into a long vector. For example, if the feature space dimension of the camera image is 100, the feature space dimension of the thermal imaging image is 80, and the feature space dimension of the ultrasonic image is 120, then the dimension of the final spliced ​​feature vector is 300. These spliced ​​features will be able to reflect the commonalities and differences between images of different modalities, provide high-quality data input for subsequent analysis and classification, and improve the accuracy and efficiency of 5G flexible circuit board detection. Ultimately, the obtained cross-modal feature fusion data set will contain the comprehensive features of multi-modal images, which is convenient for subsequent model training and defect identification.

[0055] Optionally, step S22 is specifically: Step S221: performing Spearman rank correlation coefficient calculation on the imaging modality image feature data and the thermal imaging modality image feature data, thereby obtaining imaging feature-thermal imaging feature correlation coefficient data; performing Spearman rank correlation coefficient calculation on the imaging modality image feature data and the ultrasonic modality image feature data, thereby obtaining imaging feature-ultrasonic image feature correlation coefficient data; In this embodiment, the Spearman Rank Correlation Coefficient is used to calculate the feature data of the camera modality image and the feature data of the thermal imaging modality image to evaluate the correlation between the two modalities. In specific implementation, the scipy.stats.spearmanr() function in Python can be used to input the feature data of each modality and calculate the rank correlation coefficient between the two feature data. For example, if the feature of the camera image is the edge detection result of the image, and the feature of the thermal imaging image is the temperature distribution data, the Spearman rank correlation coefficient can reflect the nonlinear correlation between the two features. If the coefficient value is close to 1, it means that the two features are highly correlated; if it is close to 0, it means that the correlation between the two is low. Similarly, the rank correlation coefficient between the camera feature and the ultrasound image feature also needs to be calculated in a similar way to obtain the correlation data between the camera feature and the ultrasound image feature. This provides a basis for subsequent feature fusion and analysis.

[0056] Step S222: sorting the correlation coefficients according to the imaging feature-thermal imaging feature correlation coefficient data and the imaging feature-ultrasound image feature correlation coefficient data, thereby obtaining multimodal image feature correlation coefficient sorting data; In this embodiment, the Spearman rank correlation coefficient data of the camera feature and the thermal imaging feature, and the camera feature and the ultrasonic image feature calculated in step S221 are sorted. Specifically, the correlation coefficient data of the camera feature and the thermal imaging feature and the correlation coefficient data of the camera feature and the ultrasonic image feature are first sorted in order from high to low for further analysis and processing. In practical applications, these correlation coefficient data can be imported into a data processing platform, and a sorting algorithm such as quick sort or merge sort can be used to arrange the data in ascending or descending order. For example, if the correlation coefficient between the camera image feature and the thermal imaging image feature is 0.85, and the correlation coefficient between the camera image feature and the ultrasonic image feature is 0.65, the correlation coefficients are sorted first, and the higher correlation coefficients are placed in front. This can help give priority to those modal features with strong correlation in subsequent steps to optimize the feature fusion effect.

[0057] Step S223: performing image feature independence variance analysis based on the multimodal image feature correlation coefficient sorting data, thereby obtaining multimodal image feature independence data; In this embodiment, the independence analysis of variance (ANOVA) of image features is performed based on the obtained correlation coefficient sorting data. Specifically, the independence of different modal features is evaluated by the variance analysis method, aiming to identify and quantify the statistical independence between the modalities. When implemented, the ANOVA method can be used, for example, using the scipy.stats.f_oneway() function in Python to perform a one-way variance analysis, using video, thermal imaging, and ultrasonic features as inputs to evaluate the variance of image features of different modalities. If the variance between certain features is small, it means that they have a high overlap in information expression, so their independence is low; if the variance is large, it means that these features are relatively independent in information content. Through this analysis, the independence between the features of each modality can be quantified, providing a theoretical basis for feature fusion.

[0058] Step S224: performing feature complementarity evaluation and feature redundancy evaluation based on the multimodal image feature independence data, thereby obtaining modal feature complementarity data and modal feature redundancy data; In this embodiment, based on the obtained multimodal image feature independence data, the complementarity and redundancy of the features are evaluated. Indicators such as mutual information (MI) are used to evaluate the complementarity between different modal features. Features with high complementarity help to enhance the performance of the model, while features with high redundancy lead to a decrease in the performance of the model. In actual operation, tools such as sklearn.feature_selection.mutual_info_classif are used to calculate the mutual information between features and quantify the complementarity of features. For example, if the edge features of the camera image and the temperature distribution features of the thermal imaging image have large mutual information for the defect recognition task, the two features have strong complementarity. At the same time, a technique similar to the correlation coefficient is used to evaluate feature redundancy, and features with high redundancy contribute less to the model and need to be appropriately removed. The evaluation results provide guidance for subsequent feature selection and weighted fusion.

[0059] Step S225: Visualize the modal feature contribution according to the modal feature complementarity data and the modal feature redundancy data, so as to obtain the modal feature contribution data.

[0060] In this embodiment, the contribution of each modal feature is displayed by visualization means in combination with the obtained feature complementarity data and redundant data. Specifically, the contribution of different modal features to the final goal can be intuitively presented using heat maps, radar maps, etc. For example, a heat map is generated using Python's matplotlib or seaborn library to display the complementarity and redundancy of each modal feature with other features. Through these graphical displays, it is possible to clearly see which modal features contribute more to defect identification and which features are redundant, helping analysts to prioritize and select features. This visualization operation can not only help further optimize the feature fusion strategy, but also provide a basis for subsequent model training and improve overall prediction performance.

[0061] Optionally, step S3 specifically includes: Step S31: performing sample division on the cross-modal feature fusion data, thereby obtaining similar sample data and heterogeneous sample data; In this embodiment, the cross-modal feature fusion data is divided into sample data of the same type and heterogeneous sample data. Specifically, for multimodal data such as images of flexible circuit boards, infrared images, and ultrasonic images, each group of samples is labeled and classified and divided according to defect types or detection tasks. Similar samples refer to samples with similar defects or features, while heterogeneous samples include samples with different defect types or features. During implementation, the data can be preliminarily divided through clustering algorithms such as K-means or DBSCAN, or the data can be manually labeled with pre-defined category labels. Through this division, it can be ensured that the subsequent learning process focuses on the similarity learning of similar samples and the difference learning of heterogeneous samples, thereby optimizing the effect of classification or regression tasks.

[0062] Step S32: maximizing positive sample learning on sample data of the same class, thereby obtaining similarity data of samples of the same class; minimizing negative sample learning on sample data of different classes, thereby obtaining similarity data of samples of different classes; In this embodiment, similarity calculation is performed on sample data of the same type by maximizing positive sample learning, aiming to obtain similarity data of samples of the same type by enhancing the similarity between samples of the same type. Specifically, a contrastive learning method (such as a triplet loss function or a cross entropy loss function) is used to map samples of the same type into the same feature space to enhance their similarity. Assuming that samples of the same type include two images from the same defect type, they should be clustered together in the feature space after model training, and the distance between them should be minimized. In the implementation process, a convolutional neural network (CNN) can be used to extract image features, and the optimal feature representation can be learned through the network. At the same time, negative sample learning is minimized for heterogeneous samples to increase the distance between them and samples of the same type and enhance the model's ability to distinguish. This will effectively improve the recognition accuracy of the model when facing samples of the same type and reduce classification or regression errors.

[0063] Step S33: performing feature representation learning based on the similarity data of samples of the same category and the similarity data of samples of different categories, thereby obtaining an image feature representation space; In this embodiment, feature representation learning is performed based on the obtained similarity data of similar samples and similarity data of heterogeneous samples. Feature representation learning automatically learns and optimizes the low-dimensional feature representation of data through methods such as deep neural networks (DNN) or autoencoders, thereby obtaining a more accurate image feature space. For samples of the same type, their similarity is maximized, and the distance between heterogeneous samples is pulled apart, thereby forming a clearer feature representation space. For example, the Triplet Loss function (including Anchor, Positive, and Negative samples) can be used to train the model so that the distance between samples of the same type (Positive) is minimized, while the distance between heterogeneous samples (Negative) is maximized. The goal of this step is to automatically extract the high-dimensional features of the sample and optimize its structure through unsupervised learning or semi-supervised learning, so that different modal data can share the same feature representation space and improve the effect of multimodal learning.

[0064] Step S34: extracting image features from the cross-modal feature fusion data set based on the image feature representation space, thereby obtaining image feature data of the flexible circuit board; In this embodiment, image feature extraction is performed on the cross-modal feature fusion data set based on the obtained image feature representation space. In the specific implementation, a deep convolutional neural network (CNN) model, such as ResNet or EfficientNet, is used to input image data and extract key features in the image. Through the mapping of the image feature representation space, data of different modalities (such as video, thermal imaging, ultrasound, etc.) are converted into a unified feature space. This enables features of different modalities to share the same feature space structure, which is conducive to the fusion and joint analysis of multimodal data. In actual operation, the data input to the network can be an image after multimodal fusion or separate data for each modality, and the extracted features will be further used for classification, detection or regression tasks. Ultimately, the obtained image feature data will provide effective feature input for subsequent tasks.

[0065] Step S35: performing image feature clustering calculation on the flexible circuit board image feature data to obtain circuit board image cluster feature data, and performing self-supervised feature label visualization on the circuit board image cluster feature data to obtain a self-supervised feature map.

[0066] In this embodiment, the obtained flexible circuit board image feature data is subjected to image feature clustering calculation. The image features are grouped by a clustering algorithm (such as K-means clustering or spectral clustering), and samples with similar features are clustered together. The purpose of clustering is to classify images with similar defects or similar appearance features into one category, so as to facilitate subsequent defect diagnosis and classification. For example, the clustering result may classify all images with the same type of defects into one category, while the other category includes images without defects or other types of defective images. When implemented, clustering can be performed by calculating the Euclidean distance or cosine similarity between image features, and the clustering algorithm is used to further refine the feature performance of each category, thereby improving the efficiency and accuracy of subsequent image analysis. The clustered image data can also be used as labels for model training to improve the effect of supervised learning. The image features in the clustering results are visualized by self-supervised feature labels. Through visualization tools (such as t-SNE or PCA), the high-dimensional image features are reduced to two-dimensional or three-dimensional data, and marked by color, shape, etc., so that the feature labels obtained by self-supervised learning are intuitively displayed. When implementing, first perform dimensionality reduction on the features obtained by clustering to reduce the dimension of the data, and then use visualization methods to display the distribution of different clusters. The purpose of this step is to allow the model trainer or analyst to intuitively understand the feature distribution of each type of image, and further verify the clustering effect and the rationality of the features. This visualization can not only provide feedback on model training, but also help to discover potential misclassification or feature confusion problems, and ultimately form a high-quality self-supervised feature map.

[0067] Optionally, step S4 specifically includes: Step S41: performing local outlier factor detection on the self-supervised feature map to obtain the local outlier factor of the circuit board; In this embodiment, the self-supervised feature map is analyzed by local outlier factor detection (LOF) to identify abnormal or irregular data points in the map, thereby obtaining the local outlier factor of the circuit board. In specific implementation, the high-dimensional features of each image sample in the self-supervised feature map are first extracted, and the local outlier factor score of each sample is calculated using the LOF algorithm. The LOF algorithm determines whether each sample is an outlier by evaluating the density of each sample relative to its neighborhood samples. Samples with higher outlier factor scores indicate that they are far away from other samples in the feature space and have anomalies or defects. Select appropriate neighborhood size and distance metrics (such as Euclidean distance or Manhattan distance) to ensure the accuracy of outlier factor detection. For example, setting K to 20 means that each sample is compared with its 20 nearest neighbor samples.

[0068] Step S42: integrating outlier factor features according to the local outlier factors of the circuit board, thereby obtaining local defect feature data of the circuit board; In this embodiment, the obtained local outlier factor data is feature integrated to generate more discriminative local defect feature data. Specifically, for the outlier factor score of each sample, it is combined with other relevant image features (such as texture, shape and color features) and integrated into a multidimensional feature vector through weighted averaging or principal component analysis (PCA). These integrated features will further reveal the specific nature of defects or anomalies, such as tiny cracks, bubbles or other irregularities on the surface of the circuit board. For each feature sample, a classification model (such as support vector machine SVM or random forest) can be applied to evaluate whether these local features indicate potential defects and provide input data for the next step of defect diagnosis and classification.

[0069] Step S43: performing linear discriminant analysis based on the local defect feature data of the circuit board, thereby obtaining main feature data of the local defects of the flexible circuit board; In this embodiment, linear discriminant analysis (LDA) is performed based on the obtained local defect feature data of the circuit board to extract the main features and reduce the data dimension. LDA is a supervised dimensionality reduction method that finds the best projection direction by maximizing the inter-class distance and minimizing the intra-class distance, and then maps the high-dimensional features into a low-dimensional space. In practical applications, samples of different defect categories (such as cracks, corrosion, peeling, etc.) can be used as different category labels, and the LDA algorithm can be used to learn the differences between these categories. In the LDA process, the feature axis that can best distinguish the defect categories is found by calculating the mean and covariance matrix of each type of data. Ultimately, the main feature data obtained will help identify and classify defect types and improve the accuracy of defect detection.

[0070] Step S44: localizing the defect area of ​​the flexible circuit board using the cross-modal feature fusion data according to the main feature data of the local defect of the flexible circuit board, thereby obtaining feature data of the local defect area of ​​the flexible circuit board; In this embodiment, local defect areas are located based on the main characteristic data of local defects of the flexible circuit board. Specifically, the image is processed by combining the characteristic data with image processing technology (such as edge detection, image segmentation or region growing) to locate the specific defect area. For example, in the image of the flexible circuit board, the image is scanned by the sliding window method, and the similarity between the characteristic data in each window and the local defect characteristics is calculated to identify the defective area. The characteristic data of the local defect area will include information such as the location, size, shape, etc. of the defect, which can be used for further analysis and repair of the subsequent defects. In addition, image segmentation can be combined with a deep learning model (such as U-Net or Mask R-CNN) to accurately determine the defect area and generate an accurate bounding box.

[0071] Step S45: integrating the main characteristic data of the local defect of the flexible circuit board and the characteristic data of the local defect area of ​​the flexible circuit board with the defect characteristics of the flexible circuit board, thereby obtaining the defect characteristic data of the flexible circuit board; In this embodiment, defect feature integration is performed by combining the obtained local defect main feature data and the obtained local defect area feature data. The spatial information of the local defect is integrated with its visual features through feature fusion technology (such as weighted average, principal component analysis PCA or multimodal learning) to form a comprehensive defect feature vector. For example, the geometric features of the defect (such as shape, edge information) are combined with the texture features (such as surface roughness, brightness change) to further improve the identifiability of the defect. The integrated defect feature data will provide a reliable basis for subsequent defect classification, positioning and repair, and support efficient flexible circuit board defect detection and diagnosis.

[0072] Step S46: performing FPCB defect pattern recognition on the FPCB defect feature data, thereby obtaining FPCB defect pattern data.

[0073] In this embodiment, defect pattern recognition is performed based on the obtained defect feature data of the flexible circuit board. In specific implementation, a deep learning model (such as a convolutional neural network CNN or an integrated learning method) is used to train and test the integrated feature data to identify different types of defect patterns. For example, a convolutional neural network can be used to perform multi-level convolution and pooling operations on the feature data to automatically extract deeper features and map them to corresponding defect categories through a classifier (such as a Softmax layer). Ultimately, the model will output a defect category label indicating the type of defect detected in the image (such as cracks, contamination, short circuits, etc.). In addition, in order to improve the accuracy and robustness of the model, data enhancement, regularization techniques, and transfer learning can also be used to further enhance the generalization ability of the model. Through this process, the diagnosis and repair of flexible circuit board defects will become more automated and efficient.

[0074] Optionally, step S42 is specifically: Step S421: mapping the self-supervised feature map to an outlier factor score based on the local outlier factor of the circuit board, thereby obtaining outlier factor score data of the sample point; In this embodiment, a local outlier factor (LOF) is calculated for each sample point in the self-supervised feature map to obtain outlier factor score data of the sample point. This process uses the LOF algorithm to evaluate the degree of abnormality of each sample point in the map relative to its neighborhood sample points. The outlier factor score of each sample point represents the density difference between the point and the surrounding samples in the high-dimensional feature space. For the processing of the self-supervised feature map, the feature data of each image sample in the map is first extracted, and the outlier factor score is obtained by calculating the relative density difference with the neighboring sample points. If an image sample has a low density in the feature space, its outlier factor score is high, indicating that the point may be an abnormal sample. These outlier factor score data will provide iconic features for defect detection in subsequent steps.

[0075] Step S422: performing feature combination on the sample point outlier factor score data and the self-supervised feature map to obtain comprehensive feature data of the flexible circuit board image; In this embodiment, the obtained sample point outlier factor score data and the self-supervised feature map are feature combined to generate comprehensive feature data of the flexible circuit board image. Feature combination can be achieved in a variety of ways, such as dimensionality reduction through feature splicing, weighted summation or principal component analysis (PCA). The outlier factor score data is integrated with the visual feature data in the self-supervised map to generate a unified feature vector. These comprehensive features include abnormal information of the sample point and local visual features of the image, such as shape, texture, color, etc., which can improve the accuracy of subsequent defect recognition after combination. For example, the outlier factor score of the sample point can be weighted proportionally with the texture features and shape features of the image to form a more comprehensive feature description to better identify potential defect areas.

[0076] Step S423: performing outlier factor score feature enhancement according to the flexible circuit board image comprehensive feature data, thereby obtaining flexible circuit board image enhanced feature data; In this embodiment, the outlier factor score feature is enhanced based on the obtained comprehensive feature data of the flexible circuit board image. In specific implementation, weighted reconstruction, feature enhancement and other technologies can be used to increase the weight of the outlier factor score in the feature vector to highlight the discrimination ability of abnormal samples. In this way, the outlier factor score of the abnormal sample will increase the influence on the final feature vector, thereby more significantly reflecting its potential defects in the subsequent detection process. When implemented, the enhancement strategy may include using an adaptive learning algorithm to adjust the weight of the score according to the contribution of different features. For example, a convolutional neural network (CNN) is used to process the comprehensive features, and the influence of the outlier factor score is automatically adjusted during the training process to more accurately identify the defect mode in the flexible circuit board.

[0077] Step S424: performing a weighted evaluation of the local abnormality of the circuit board image features on the flexible circuit board image enhanced feature data, thereby obtaining the local abnormal value data of the circuit board image features; In this embodiment, the image features are subjected to a weighted evaluation of local abnormalities in combination with the obtained flexible circuit board image enhancement feature data. In specific implementation, the image features can be regionally weighted by a local weighted evaluation method to enhance those areas where defects may exist. For example, using a weighting method based on gradient or image details, the abnormal values ​​in the local area will be more strongly reflected in the image features than other areas. Through weighted evaluation, a specific weight can be assigned to each local area in the image to identify the significance of the highly abnormal area. This process can adjust the degree of influence of the feature through a weighted algorithm (such as weighted average, weighted regression, etc.), making the abnormal area in the image more prominent, thereby providing an important basis for subsequent defect diagnosis.

[0078] Step S425: integrating high outlier value local image features according to the local outlier value data of the circuit board image features, thereby obtaining the local defect feature data of the circuit board.

[0079] In this embodiment, the high outlier value local image features are integrated based on the obtained local outlier value data of the circuit board image features. During implementation, sample points with higher outlier values ​​are first screened out from the feature local outlier value data, and these sample points represent potential defect areas. Then, the feature data of these high outlier value areas are merged with the features of other areas through a feature integration algorithm (such as weighted summation, principal component analysis PCA), thereby forming a comprehensive local defect feature data set. This data set can be used for subsequent defect analysis and diagnosis, and the accuracy of defect identification can be improved by centralized processing of high outliers. For example, for areas where cracks or poor welding may exist, the integrated features will help to calibrate these abnormal areas and provide a reference for repair, thereby improving the production quality and detection efficiency of the circuit board.

[0080] Optionally, the present specification also provides a training system for a 5G flexible circuit board defect detection model, which is used to execute the training method for the 5G flexible circuit board defect detection model as described above, and the training system for the 5G flexible circuit board defect detection model includes: A multimodal image preprocessing module is used to obtain multimodal 5G flexible circuit board scanning data, and perform multimodal scanning data denoising based on the multimodal 5G flexible circuit board scanning data, so as to obtain multimodal 5G flexible circuit board perception denoising data; perform multimodal image feature enhancement on the multimodal 5G flexible circuit board perception denoising data, so as to obtain multimodal 5G flexible circuit board image enhancement data; A cross-modal feature fusion module is used to perform deep multi-modal feature fusion on the multi-modal 5G flexible circuit board image enhancement data, thereby obtaining 5G flexible circuit board feature fusion data, and perform cross-modal feature mapping based on the 5G flexible circuit board feature fusion data, thereby obtaining a cross-modal feature fusion data set; A self-supervised feature extraction module is used to perform self-supervised image feature extraction on the cross-modal feature fusion data set to obtain flexible circuit board image feature data, and perform self-supervised image feature clustering on the flexible circuit board image feature data to obtain a self-supervised feature map; A defect mode analysis module is used to perform defect analysis of the flexible circuit board according to the self-supervised feature map, thereby obtaining defect feature data of the flexible circuit board, and perform defect mode recognition of the flexible circuit board on the defect feature data of the flexible circuit board, thereby obtaining defect mode data of the flexible circuit board; The model training module is used to perform global defect evaluation of the flexible circuit board on the self-supervised feature map according to the defect pattern data of the flexible circuit board, so as to obtain the global defect data of the flexible circuit board, and to train the 5G flexible circuit board defect detection model based on the global defect data of the flexible circuit board, so as to obtain the 5G flexible circuit board defect detection model.

Claims

1. A training method for a 5G flexible circuit board defect detection model, characterized in that: The following steps are involved: Step S1: acquiring multimodal 5G flexible circuit board scanning data, and performing multimodal scanning data denoising based on the multimodal 5G flexible circuit board scanning data, thereby obtaining multimodal 5G flexible circuit board perception denoising data; performing multimodal image feature enhancement on the multimodal 5G flexible circuit board perception denoising data, thereby obtaining multimodal 5G flexible circuit board image enhancement data; Step S2: performing deep multimodal feature fusion on the multimodal 5G flexible circuit board image enhancement data to obtain 5G flexible circuit board feature fusion data, and performing cross-modal feature mapping based on the 5G flexible circuit board feature fusion data to obtain a cross-modal feature fusion data set; Step S3: performing self-supervised image feature extraction on the cross-modal feature fusion data set to obtain flexible circuit board image feature data, and performing self-supervised image feature clustering on the flexible circuit board image feature data to obtain a self-supervised feature map; Step S4: performing defect analysis on the flexible circuit board according to the self-supervised feature map to obtain defect feature data of the flexible circuit board, and performing defect pattern recognition on the defect feature data of the flexible circuit board to obtain defect pattern data of the flexible circuit board; Step S5: Performing a global defect assessment of the flexible circuit board on the self-supervised feature map according to the defect pattern data of the flexible circuit board, thereby obtaining the global defect data of the flexible circuit board, and training a 5G flexible circuit board defect detection model based on the global defect data of the flexible circuit board, thereby obtaining a 5G flexible circuit board defect detection model.

2. The training method for a 5G flexible circuit board defect detection model according to claim 1, characterized in that: Step S1 is specifically as follows: Step S11: Acquire multi-modal 5G flexible circuit board scanning data and flexible circuit board scanning environment sensing data; Step S12: performing multi-modal scanning data denoising on the multi-modal 5G flexible circuit board scanning data according to the flexible circuit board scanning environment sensing data, thereby obtaining multi-modal scanning denoised data; Step S13: performing preliminary fusion of multimodal self-attention perception based on the multimodal scanning denoised data, thereby obtaining multimodal 5G flexible circuit board perception denoised data; Step S14: performing multimodal image visibility evaluation on the multimodal 5G flexible circuit board perception denoising data, thereby obtaining multimodal image visibility data; Step S15: Perform multimodal image feature enhancement on the multimodal 5G flexible circuit board perception denoising data according to the multimodal image visibility data, so as to obtain multimodal 5G flexible circuit board image enhancement data.

3. The training method for a 5G flexible circuit board defect detection model according to claim 2, characterized in that: Step S12 is specifically as follows: Performing data alignment on the multi-modal 5G flexible circuit board scanning data and the flexible circuit board scanning environment sensing data, thereby obtaining the multi-modal 5G flexible circuit board scanning alignment data and the flexible circuit board scanning environment sensing alignment data; Performing scale conversion on the multimodal 5G flexible circuit board scanning alignment data, thereby obtaining the multimodal 5G flexible circuit board scanning data to be processed, wherein the multimodal 5G flexible circuit board scanning data to be processed includes a 5G flexible circuit board camera image, a 5G flexible circuit board infrared image, and a 5G flexible circuit board ultrasonic image; Multimodal data noise source identification is performed based on the multimodal 5G flexible circuit board scanning data to be processed, thereby obtaining image noise source data, ultrasonic signal noise source data, and infrared noise source data; The influence of the scanning environment illumination is quantified according to the scanning environment sensor alignment data of the flexible circuit board and the image noise source data, so as to obtain the environmental illumination influence factor; The influence of the scanning environment temperature is quantified according to the scanning environment sensor alignment data of the flexible circuit board and the infrared noise source data, so as to obtain the environmental temperature influence factor; The impact of the scanning environment signal is quantified according to the flexible circuit board scanning environment sensor alignment data and the ultrasonic signal noise source data, so as to obtain the environmental signal impact factor; Based on the environmental lighting influence factor, the 5G flexible circuit board camera image is subjected to median filtering and denoising, thereby obtaining a 5G flexible circuit board camera denoised image; Based on the environmental temperature influencing factor, the infrared scanning wavelet transform is used to denoise the infrared image of the 5G flexible circuit board, thereby obtaining the infrared scanning denoised image of the 5G flexible circuit board; Based on the environmental signal influencing factors, the ultrasonic signal Kalman denoising is performed on the ultrasonic image of the 5G flexible circuit board, so as to obtain the denoised image of the ultrasonic detection of the 5G flexible circuit board; The 5G flexible circuit board camera denoised image, the 5G flexible circuit board infrared scanning denoised image and the 5G flexible circuit board ultrasonic detection denoised image are merged to obtain multi-modal scanning denoised data.

4. The training method for a 5G flexible circuit board defect detection model according to claim 2, characterized in that: Step S15 is specifically as follows: Step S151: screening the multimodal 5G flexible circuit board perception denoising data for low-visibility images according to the multimodal image visibility data, thereby obtaining a multimodal low-visibility image set; Step S152: performing local contrast enhancement on the multimodal low-visibility image set, thereby obtaining a multimodal low-visibility local contrast enhanced image set; Step S153: performing multimodal convolutional layer feature sharing on the multimodal low-visibility local contrast enhanced image set, thereby obtaining a multimodal feature sharing image set; Step S154: performing image noise suppression and edge enhancement according to the multimodal feature sharing image set, thereby obtaining a low-visibility enhanced image set; Step S155: Perform image enhancement feature coupling on the multimodal 5G flexible circuit board perception denoising data based on the low-visibility enhanced image set, so as to obtain multimodal 5G flexible circuit board image enhancement data.

5. The training method for a 5G flexible circuit board defect detection model according to claim 1, characterized in that: Step S2 is specifically as follows: Step S21: performing modal depth feature extraction on the multi-modal 5G flexible circuit board image enhancement data, so as to obtain camera modal image feature data, thermal imaging modal image feature data and ultrasonic modal image feature data; Step S22: performing modal feature contribution evaluation according to the camera modal image feature data, the thermal imaging modal image feature data, and the ultrasonic modal image feature data, so as to obtain modal feature contribution data; Step S23: performing multi-modal data weighted fusion on the camera modal image feature data, the thermal imaging modal image feature data and the ultrasonic modal image feature data based on the modal feature contribution data, so as to obtain 5G flexible circuit board feature fusion data; Step S24: performing cross-modal feature space mapping on the 5G flexible circuit board feature fusion data, thereby obtaining a cross-modal mapping feature space; Step S25: performing mapping feature splicing according to the cross-modal mapping feature space to obtain a cross-modal feature fusion data set.

6. The training method for a 5G flexible circuit board defect detection model according to claim 5, characterized in that: Step S22 is specifically as follows: Step S221: performing Spearman rank correlation coefficient calculation on the imaging modality image feature data and the thermal imaging modality image feature data, thereby obtaining imaging feature-thermal imaging feature correlation coefficient data; performing Spearman rank correlation coefficient calculation on the imaging modality image feature data and the ultrasonic modality image feature data, thereby obtaining imaging feature-ultrasonic image feature correlation coefficient data; Step S222: sorting the correlation coefficients according to the imaging feature-thermal imaging feature correlation coefficient data and the imaging feature-ultrasound image feature correlation coefficient data, thereby obtaining multimodal image feature correlation coefficient sorting data; Step S223: performing image feature independence variance analysis based on the multimodal image feature correlation coefficient sorting data, thereby obtaining multimodal image feature independence data; Step S224: performing feature complementarity evaluation and feature redundancy evaluation based on the multimodal image feature independence data, thereby obtaining modal feature complementarity data and modal feature redundancy data; Step S225: Visualize the modal feature contribution according to the modal feature complementarity data and the modal feature redundancy data, so as to obtain the modal feature contribution data.

7. The training method for a 5G flexible circuit board defect detection model according to claim 1, characterized in that: Step S3 is specifically as follows: Step S31: performing sample division on the cross-modal feature fusion data, thereby obtaining similar sample data and heterogeneous sample data; Step S32: maximizing positive sample learning for sample data of the same type, thereby obtaining similarity data of samples of the same type; Minimize negative sample learning on heterogeneous sample data to obtain heterogeneous sample similarity data; Step S33: performing feature representation learning based on the similarity data of samples of the same category and the similarity data of samples of different categories, thereby obtaining an image feature representation space; Step S34: extracting image features from the cross-modal feature fusion data set based on the image feature representation space, thereby obtaining image feature data of the flexible circuit board; Step S35: performing image feature clustering calculation on the flexible circuit board image feature data to obtain circuit board image cluster feature data, and performing self-supervised feature label visualization on the circuit board image cluster feature data to obtain a self-supervised feature map.

8. The training method for a 5G flexible circuit board defect detection model according to claim 1, characterized in that: Step S4 is specifically as follows: Step S41: performing local outlier factor detection on the self-supervised feature map to obtain the local outlier factor of the circuit board; Step S42: integrating outlier factor features according to the local outlier factors of the circuit board, thereby obtaining local defect feature data of the circuit board; Step S43: performing linear discriminant analysis based on the local defect feature data of the circuit board, thereby obtaining main feature data of the local defects of the flexible circuit board; Step S44: localizing the defect area of ​​the flexible circuit board using the cross-modal feature fusion data according to the main feature data of the local defect of the flexible circuit board, thereby obtaining feature data of the local defect area of ​​the flexible circuit board; Step S45: integrating the main characteristic data of the local defect of the flexible circuit board and the characteristic data of the local defect area of ​​the flexible circuit board with the defect characteristics of the flexible circuit board, thereby obtaining the defect characteristic data of the flexible circuit board; Step S46: performing FPCB defect pattern recognition on the FPCB defect feature data, thereby obtaining FPCB defect pattern data.

9. The training method for a 5G flexible circuit board defect detection model according to claim 8, characterized in that: Step S42 is specifically as follows: Step S421: mapping the self-supervised feature map to an outlier factor score based on the local outlier factor of the circuit board, thereby obtaining outlier factor score data of the sample point; Step S422: performing feature combination on the sample point outlier factor score data and the self-supervised feature map to obtain comprehensive feature data of the flexible circuit board image; Step S423: performing outlier factor score feature enhancement according to the flexible circuit board image comprehensive feature data, thereby obtaining flexible circuit board image enhanced feature data; Step S424: performing a weighted evaluation of the local abnormality of the circuit board image features on the flexible circuit board image enhanced feature data, thereby obtaining the local abnormal value data of the circuit board image features; Step S425: integrating high outlier value local image features according to the local outlier value data of the circuit board image features, thereby obtaining the local defect feature data of the circuit board.

10. A training system for a 5G flexible circuit board defect detection model, characterized in that: Used to execute the training method for the 5G flexible circuit board defect detection model as claimed in claim 1, the training system for the 5G flexible circuit board defect detection model comprises: A multimodal image preprocessing module is used to obtain multimodal 5G flexible circuit board scanning data, and perform multimodal scanning data denoising based on the multimodal 5G flexible circuit board scanning data, so as to obtain multimodal 5G flexible circuit board perception denoising data; perform multimodal image feature enhancement on the multimodal 5G flexible circuit board perception denoising data, so as to obtain multimodal 5G flexible circuit board image enhancement data; A cross-modal feature fusion module is used to perform deep multi-modal feature fusion on the multi-modal 5G flexible circuit board image enhancement data, thereby obtaining 5G flexible circuit board feature fusion data, and perform cross-modal feature mapping based on the 5G flexible circuit board feature fusion data, thereby obtaining a cross-modal feature fusion data set; A self-supervised feature extraction module is used to perform self-supervised image feature extraction on the cross-modal feature fusion data set to obtain flexible circuit board image feature data, and perform self-supervised image feature clustering on the flexible circuit board image feature data to obtain a self-supervised feature map; A defect mode analysis module is used to perform defect analysis of the flexible circuit board according to the self-supervised feature map, thereby obtaining defect feature data of the flexible circuit board, and perform defect mode recognition of the flexible circuit board on the defect feature data of the flexible circuit board, thereby obtaining defect mode data of the flexible circuit board; The model training module is used to perform global defect evaluation of the flexible circuit board on the self-supervised feature map according to the defect pattern data of the flexible circuit board, so as to obtain the global defect data of the flexible circuit board, and to train the 5G flexible circuit board defect detection model based on the global defect data of the flexible circuit board, so as to obtain the 5G flexible circuit board defect detection model.

Images