Small sample defect diagnosis method, device and equipment with failure feature decoupling, and medium

Abstract

The application provides a small sample defect diagnosis method and device with decoupled failure characteristics, equipment and medium. The method comprises the following steps: constructing a failure physical knowledge base for storing physical law information of defects of new energy equipment; training a characteristic decoupling network according to the knowledge base, the network can extract content and attribute characteristics from equipment images, and a physical constraint loss function based on the physical law information of the knowledge base is introduced during training, which contains characteristic separation degree, physical law compliance and semantic consistency constraint; performing semantic enhancement based on the trained network and the knowledge base, generating enhanced samples by using a physical constraint interpolation algorithm, dynamically determining interpolation coefficients according to the physical rules of the knowledge base, and finally determining a defect diagnosis result according to the enhanced samples. The application generates physically reasonable enhanced samples under small samples by constructing a knowledge base, introducing a constraint training network and performing semantic enhancement, overcomes the problems of insufficient generalization ability and unreal samples of the prior art model, and improves the defect diagnosis accuracy and adaptability.

Description

Technical Field

[0001] This application belongs to the field of small sample defect diagnosis, and in particular relates to a method, apparatus, equipment and medium for small sample defect diagnosis with decoupled failure features. Background Technology

[0002] Currently, deep learning-based defect diagnosis for new energy power plants is being used for intelligent operation and maintenance. This aims to automatically identify equipment defects, such as microcracks, hot spots, and corrosion in photovoltaic panels, through a data-driven approach. This method has become a research hotspot, focusing on learning patterns from large amounts of data to improve diagnostic efficiency.

[0003] Deep learning includes a data-driven learning model, which involves training deep neural networks (such as convolutional neural networks) to extract features directly from device images and using data augmentation techniques (such as simple geometric transformations or image rotation and scaling under statistical rules) to expand the number of samples in order to improve the model's performance on limited data.

[0004] In the process of using data-driven and simple data augmentation, there are significant drawbacks due to the lack of in-depth understanding of the physical mechanisms of equipment failure (such as the formation mechanism and evolution process of defects): the model is difficult to effectively handle small sample learning problems, has insufficient generalization ability, cannot adapt to changes in different equipment types and operating conditions, and the physical irrationality of the generated samples restricts the promotion and application of intelligent diagnostic technology in actual industrial scenarios. Summary of the Invention

[0005] The purpose of this application is to overcome the deficiencies in the prior art and provide a method, apparatus, equipment and medium for small sample defect diagnosis with decoupled failure characteristics.

[0006] This application provides a small-sample defect diagnosis method with decoupled failure features, including: A failure physics knowledge base is constructed, which stores information on the physical laws governing defects in new energy equipment; Based on the failure physics knowledge base, a feature decoupling network is trained. The feature decoupling network is used to extract content features and attribute features from device images. During the training process, a physical constraint loss function is introduced. The physical constraint loss function is based on the physical law information of the failure physics knowledge base and includes feature separation degree constraint, physical law conformity degree constraint and semantic consistency constraint. Based on the trained feature decoupling network and the failure physics knowledge base, semantic enhancement is performed, including: generating enhanced samples using a physical constraint interpolation algorithm, wherein the physical constraint interpolation algorithm dynamically determines the interpolation coefficients according to the physical rules in the failure physics knowledge base; Based on the enhanced sample, the defect diagnosis result is determined.

[0007] Optionally, a failure physics knowledge base is constructed, which stores physical law information about defects in new energy equipment, including: Design a multi-layered structure for the failure physics knowledge base, the multi-layered structure including a representation layer, a formation mechanism layer, an evolution process layer, and an object of action layer; In the representation layer, defect visual features are defined, including fine linear textures of hidden cracks, local color difference variations of hot spots, and increased surface roughness of corrosion. The physical causes are defined in the formation mechanism layer, including microcracks originating from mechanical stress concentration, hot spots caused by localized increase in resistance, and corrosion caused by electrochemical reactions; The evolutionary process layer defines the development process, which includes the initial characteristics, intermediate manifestations and final morphology of the defect; Influencing components are defined in the target layer, including the photovoltaic panel's battery cells and connection circuits.

[0008] Optionally, a physical constraint loss function is introduced, which is based on the physical law information of the failure physics knowledge base and includes feature separation degree constraints, physical law conformity constraints, and semantic consistency constraints, including: The calculation of feature separation constraints includes evaluating the independence between the content features and attribute features; The physical law compliance constraint is calculated, which includes comparing the content features and attribute features with the physical laws in the failed physical knowledge base. Calculate semantic consistency constraints, which include matching the content features and attribute features with the semantic description.

[0009] Optionally, a feature decoupling network is trained to extract content features and attribute features from a device image, including: After extracting the content features and attribute features, the content features and attribute features are fused using a cross-attention mechanism; The fusion process includes converting the content features into query vectors, converting the attribute features into key-value vectors, calculating an attention weight matrix, and then weighting and fusing the content features and attribute features based on the attention weight matrix.

[0010] Optionally, semantic enhancement is performed, including generating enhanced samples using a physical constraint interpolation algorithm, including: Based on the defect evolution laws in the aforementioned failure physics knowledge base, a feature recombination rule set is established to define physically reasonable feature combinations; The feature recombination rule set includes rules to prevent arbitrary combinations of hidden crack features and hot spot features.

[0011] Optionally, the step of generating enhanced samples using a physical constraint interpolation algorithm, wherein the physical constraint interpolation algorithm dynamically determines the interpolation coefficients based on the physical rules in the failure physics knowledge base, includes: Based on the physical rules in the aforementioned failure physics knowledge base, the interpolation coefficients are dynamically calculated; The dynamic calculation of interpolation coefficients includes adjusting the interpolation parameters based on the defect evolution rate.

[0012] Optionally, semantic enhancement is performed, including generating enhanced samples using a physical constraint interpolation algorithm, including: Before generating enhanced samples, the generated features are validated according to rules. The rule verification includes: performing physical law compliance checks, performing numerical simulation verification, and performing case comparison analysis.

[0013] This application also provides a small-sample defect diagnosis device with decoupled failure features, comprising: The module constructs a failure physics knowledge base, which stores information on the physical laws governing defects in new energy equipment. The training module trains a feature decoupling network based on the failure physics knowledge base. The feature decoupling network is used to extract content features and attribute features from device images. During the training process, a physical constraint loss function is introduced. The physical constraint loss function is based on the physical law information of the failure physics knowledge base and includes feature separation degree constraint, physical law conformity degree constraint and semantic consistency constraint. The semantic module performs semantic enhancement based on the trained feature decoupling network and the failure physics knowledge base, including: generating enhanced samples using a physical constraint interpolation algorithm, wherein the physical constraint interpolation algorithm dynamically determines the interpolation coefficients according to the physical rules in the failure physics knowledge base; The diagnostic module determines the defect diagnosis result based on the enhanced sample.

[0014] This application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program, controls the execution of the method as described above.

[0015] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to control the execution of the above-described method.

[0016] The beneficial effects of this application are: This application provides a small-sample defect diagnosis method with decoupled failure features, comprising: constructing a failure physics knowledge base, wherein the failure physics knowledge base stores physical law information of defects in new energy equipment; training a feature decoupling network based on the failure physics knowledge base, wherein the feature decoupling network is used to extract content features and attribute features from equipment images, wherein a physical constraint loss function is introduced during the training process, wherein the physical constraint loss function is based on the physical law information of the failure physics knowledge base and includes feature separation degree constraint, physical law conformity constraint, and semantic consistency constraint; performing semantic enhancement based on the trained feature decoupling network and the failure physics knowledge base, including: generating enhanced samples using a physical constraint interpolation algorithm, wherein the physical constraint interpolation algorithm dynamically determines the interpolation coefficients according to the physical rules in the failure physics knowledge base; and determining the defect diagnosis result based on the enhanced samples. This application, by constructing a failure physics knowledge base, introducing physical constraints to train the feature decoupling network, and performing semantic enhancement based on physical rules, achieves the generation of physically reasonable enhanced samples under small-sample conditions, effectively overcoming the shortcomings of insufficient generalization ability and unrealistic samples in existing technologies, and improving the accuracy and adaptability of defect diagnosis. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the small sample defect diagnosis process for failure feature decoupling in this application; Figure 2 This is a schematic diagram of the small sample defect diagnosis architecture for decoupling failure features in this application. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, the embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] Please refer to Figure 1 and Figure 2 As shown, this application provides a small-sample defect diagnosis method with decoupled failure features, applied to the field of new energy equipment defect diagnosis, to solve the problems of model overfitting and insufficient generalization ability under small sample conditions. The method includes: After the system starts, it loads the basic database, including equipment technical parameters, material property data, and historical defect records. The input equipment images undergo standardization processing, including adjusting image size, using bicubic interpolation to maintain image quality, and performing contrast-limited adaptive histogram equalization to enhance the visual effect of defect areas.

[0020] At the same time, a data acquisition pipeline was established to ensure that image data and corresponding physical parameters were acquired and stored synchronously.

[0021] S101. Construct a failure physics knowledge base, which stores physical law information about defects in new energy equipment.

[0022] Building a failure physics knowledge base refers to creating a database specifically for typical defects in new energy equipment. This knowledge base systematically organizes key information such as the surface manifestations, formation mechanisms, evolution processes, and targets of common defects in photovoltaic panels, such as microcracks, hot spots, and corrosion.

[0023] The knowledge base adopts a multi-layered structure. The representation layer records the visual characteristics of different defects in detail, such as microcracks appearing as fine linear textures, hot spots showing local color differences, and corrosion showing increased surface roughness. The formation mechanism layer analyzes the physical causes of various defects, such as microcracks originating from mechanical stress concentration, hot spots caused by increased local resistance, and corrosion caused by electrochemical reactions. The evolution process layer describes the complete process of defects from initiation to development, including initial characteristics, intermediate manifestations, and final morphology. The target layer clarifies the equipment components and functional modules affected by the defects, such as the battery cells and connection circuits of photovoltaic panels.

[0024] The knowledge base transforms the inherent limitations of natural language descriptions into machine-understandable structured data by establishing a mapping between text descriptions and image features. The specific implementation process involves using the BERT model to extract semantic features from the text, while simultaneously using a CNN network to extract visual features from the image. Then, an attention mechanism is used to calculate the similarity between the two types of features, establishing a mapping relationship from the semantic space to the visual space.

[0025] This mapping ensures that the physical laws in the knowledge base can accurately guide the subsequent feature processing, providing physical support for feature decoupling and semantic enhancement.

[0026] The system automatically parses equipment technical documents, extracts key physical parameters, and imports historical defect images and their corresponding text descriptions. A cross-modal learning algorithm is used to establish a mapping relationship between text features and image features, forming a complete knowledge base system.

[0027] The knowledge base adopts a version management mechanism, supports online updates and incremental learning, and automatically starts the knowledge base update process when a new defect type appears.

[0028] S102. Based on the failure physics knowledge base, train a feature decoupling network. The feature decoupling network is used to extract content features and attribute features from device images. During the training process, a physical constraint loss function is introduced. The physical constraint loss function is based on the physical law information of the failure physics knowledge base and includes feature separation degree constraint, physical law conformity degree constraint, and semantic consistency constraint.

[0029] The training feature decoupling network is a dual-branch network that integrates failure physics knowledge. This network consists of a content encoder, an attribute encoder, a physical constraint module, and a feature fusion module.

[0030] The content encoder adopts a ResNet architecture, extracts the basic structural features of the device through four residual blocks, and focuses on capturing common information that does not change with defects; The attribute encoder employs a fully convolutional network design, focusing on extracting the unique morphological features of defects through eight convolutional layers.

[0031] The input image is simultaneously fed into the content encoder and the attribute encoder to obtain content features and attribute features, respectively. These two features are then fed into the physical constraint module.

[0032] The physical constraint module constrains the network by constructing a multi-task loss function, including adversarial loss, reconstruction loss, and physical constraint loss. The physical constraint loss comprises three components: feature separation constraint, which ensures that content features and attribute features are independent of each other by evaluating the independence between them; physical law conformity constraint, which ensures that features conform to the physical laws in the knowledge base by comparing content features and attribute features with the physical laws in the invalid physical knowledge base; and semantic consistency constraint, which ensures that features are consistent with the semantic description by matching content features and attribute features with the semantic description.

[0033] The feature fusion module effectively fuses content and attribute features, using a cross-attention mechanism to enable interaction between the two feature branches. Specifically, the content features are converted into query vectors, the attribute features are converted into key-value vectors, the attention weight matrix is ​​calculated, and then feature weighted fusion is performed based on the attention weights.

[0034] This interaction mechanism ensures the collaborative representation of content and attribute features while preserving their individual characteristics. The network ultimately outputs fused features that include both the device's basic structural information and the morphological features of defects, while conforming to physical constraints.

[0035] A phased strategy is adopted during training. First, the network is pre-trained using a large-scale dataset to enable the network to learn general feature representations. Then, physical constraints are introduced for fine-tuning. The network parameters are gradually adjusted through an alternating optimization strategy. Each training batch contains 32 samples, and the training cycle is 100 rounds. The learning rate is adjusted using a cosine annealing scheduler.

[0036] S103. Based on the trained feature decoupling network and the failure physics knowledge base, semantic enhancement is performed, including: generating enhanced samples using a physical constraint interpolation algorithm, wherein the physical constraint interpolation algorithm dynamically determines the interpolation coefficients according to the physical rules in the failure physics knowledge base.

[0037] Semantic enhancement refers to performing physically meaningful recombination operations in the feature space to generate new samples that conform to the real defect distribution, in order to solve the problem of data scarcity under small sample conditions.

[0038] The specific implementation process includes three stages: First, based on the defect evolution law in the knowledge base, a feature recombination rule set is established to clarify which feature combinations are physically reasonable. For example, the feature of microcracks cannot be arbitrarily combined with the color change feature of hot spots. The feature recombination rule set is used to define physically reasonable feature combinations. Secondly, a physical constraint feature interpolation algorithm is designed to perform physically based interpolation between the feature vectors of the source samples. The interpolation coefficients are dynamically determined by the physical rules in the knowledge base. The dynamic calculation of the interpolation coefficients includes adjusting the interpolation parameters based on the defect evolution rate. Finally, the generated features are physically validated by a rule validation engine. Rule validation includes physical law compliance checks, numerical simulation checks, and case comparison analysis.

[0039] The rule validation engine employs a three-level validation process: The first stage performs a rapid physical law compliance check to verify whether the generated features violate basic physical principles. The second level involves numerical simulation verification, which evaluates the rationality of the feature combination using a simplified physical model. The third level involves case comparison, which analyzes the similarity between the generated features and real cases in the knowledge base.

[0040] Only features that pass all verifications will proceed to the sample generation stage.

[0041] Finally, the validated features are decoded into enhanced samples by a decoder based on the U-Net architecture. The decoder extracts semantic information of the features through the encoding path and reconstructs image details through the decoding path. Skip connections ensure the effective fusion of features at different levels.

[0042] The generated new samples not only retain the statistical characteristics of the original data but also conform to physical laws, providing high-quality augmented data for subsequent meta-learning training.

[0043] Sample pairs are randomly selected from the training set, and their feature representations are extracted by the network. All generated samples undergo quality evaluation, and only those that meet the requirements are added to the augmented training set.

[0044] S104. Determine the defect diagnosis result based on the enhanced sample.

[0045] Determining the defect diagnosis result involves training a multi-task meta-learning framework using augmented samples, ultimately generating a physically plausible defect diagnosis model with strong generalization ability. The multi-task meta-learning framework comprises four execution steps: The first step is task construction, which involves building a meta-learning task set based on the defect classification in the knowledge base. Each task contains a support set and a query set. The second step is rapid adaptation. In the inner loop, the model performs gradient updates on the support set and learns the feature representation of the new task. The third step is physical rationality verification. Physical constraint loss is introduced into the outer loop. The physical consistency assessment module performs physical law conformity test on the model prediction results based on the failure physics knowledge base. The physical consistency score is obtained by calculating the matching degree between the model prediction and the physical laws in the knowledge base. The fourth step is multi-task optimization, which coordinates the optimization of three loss functions, including the main task loss, physical constraint loss, and domain adaptation loss. The three losses are coordinated and optimized through an adaptive weight balancing mechanism, and the weight coefficients are dynamically adjusted according to the gradient norm of each task.

[0046] This multi-task design allows the framework to maintain rapid adaptability while significantly improving the physical rationality and domain generalization of the model.

[0047] The meta-learning framework directly utilizes the results of pre-sequence feature decoupling and semantic enhancement to ultimately generate a defect diagnosis model with strong generalization ability and physical rationality. This model can quickly adapt to new defect types under small sample conditions while ensuring the physical credibility of the diagnosis results.

[0048] A learning strategy is employed, gradually transitioning from simple to complex tasks. After training, the optimal model parameters are saved, and performance is validated.

[0049] This application also provides a small-sample defect diagnosis method for failure feature decoupling, including: The module constructs a failure physics knowledge base, which stores information on the physical laws governing defects in new energy equipment. The training module trains a feature decoupling network based on the failure physics knowledge base. The feature decoupling network is used to extract content features and attribute features from device images. During the training process, a physical constraint loss function is introduced. The physical constraint loss function is based on the physical law information of the failure physics knowledge base and includes feature separation degree constraint, physical law conformity degree constraint and semantic consistency constraint. The semantic module performs semantic enhancement based on the trained feature decoupling network and the failure physics knowledge base, including: generating enhanced samples using a physical constraint interpolation algorithm, wherein the physical constraint interpolation algorithm dynamically determines the interpolation coefficients according to the physical rules in the failure physics knowledge base; The diagnostic module determines the defect diagnosis result based on the enhanced sample.

[0050] This application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program, controls the execution of the method as described above.

[0051] This application also provides a computer-readable storage medium storing a computer program that, when executed in a computer, causes the computer to control the execution of the above-described method.

Claims

1. A small-sample defect diagnosis method with decoupled failure features, characterized in that, include: A failure physics knowledge base is constructed, which stores information on the physical laws governing defects in new energy equipment; According to the aforementioned failure physics knowledge base A feature decoupling network is trained to extract content features and attribute features from device images. During training, a physical constraint loss function is introduced. The physical constraint loss function is based on the physical law information of the failure physics knowledge base and includes feature separation degree constraint, physical law compliance degree constraint and semantic consistency constraint. Based on the trained feature decoupling network and the failure physics knowledge base, semantic enhancement is performed, including: generating enhanced samples using a physical constraint interpolation algorithm, wherein the physical constraint interpolation algorithm dynamically determines the interpolation coefficients according to the physical rules in the failure physics knowledge base; Based on the enhanced sample, the defect diagnosis result is determined.

2. The method according to claim 1, characterized in that, A failure physics knowledge base is constructed, which stores information on the physical laws governing defects in new energy equipment, including: Design a multi-layered structure for the failure physics knowledge base, the multi-layered structure including a representation layer, a formation mechanism layer, an evolution process layer, and an object of action layer; In the representation layer, defect visual features are defined, including fine linear textures of hidden cracks, local color difference variations of hot spots, and increased surface roughness of corrosion. The physical causes are defined in the formation mechanism layer, including microcracks originating from mechanical stress concentration, hot spots caused by localized increase in resistance, and corrosion caused by electrochemical reactions; The evolutionary process layer defines the development process, which includes the initial characteristics, intermediate manifestations and final morphology of the defect; Influencing components are defined in the target layer, including the photovoltaic panel's battery cells and connection circuits.

3. The method according to claim 1, characterized in that, A physical constraint loss function is introduced, which is based on the physical law information of the failure physics knowledge base and includes feature separation degree constraint, physical law conformity degree constraint, and semantic consistency constraint, including: The calculation of feature separation constraints includes evaluating the independence between the content features and attribute features; The physical law compliance constraint is calculated, which includes comparing the content features and attribute features with the physical laws in the failed physical knowledge base. Calculate semantic consistency constraints, which include matching the content features and attribute features with the semantic description.

4. The method according to claim 1, characterized in that, Training a feature decoupling network, which is used to extract content features and attribute features from a device image, including: After extracting the content features and attribute features, the content features and attribute features are fused using a cross-attention mechanism; The fusion process includes converting the content features into query vectors, converting the attribute features into key-value vectors, calculating an attention weight matrix, and then weighting and fusing the content features and attribute features based on the attention weight matrix.

5. The method according to claim 1, characterized in that, Semantic enhancement includes generating enhanced samples using physical constraint interpolation algorithms, including: Based on the defect evolution laws in the aforementioned failure physics knowledge base, a feature recombination rule set is established to define physically reasonable feature combinations; The feature recombination rule set includes rules to prevent arbitrary combinations of hidden crack features and hot spot features.

6. The method according to claim 1, characterized in that, The step of generating enhanced samples using a physical constraint interpolation algorithm, wherein the physical constraint interpolation algorithm dynamically determines interpolation coefficients based on physical rules in the failure physics knowledge base, includes: Based on the physical rules in the aforementioned failure physics knowledge base, the interpolation coefficients are dynamically calculated; The dynamic calculation of interpolation coefficients includes adjusting the interpolation parameters based on the defect evolution rate.

7. The method according to claim 1, characterized in that, Semantic enhancement includes generating enhanced samples using physical constraint interpolation algorithms, including: Before generating enhanced samples, the generated features are validated according to rules. The rule verification includes: performing physical law compliance checks, performing numerical simulation verification, and performing case comparison analysis.

8. A small-sample defect diagnosis device with decoupled failure characteristics, characterized in that, include: The module constructs a failure physics knowledge base, which stores information on the physical laws governing defects in new energy equipment. The training module trains a feature decoupling network based on the failure physics knowledge base. The feature decoupling network is used to extract content features and attribute features from device images. During the training process, a physical constraint loss function is introduced. The physical constraint loss function is based on the physical law information of the failure physics knowledge base and includes feature separation degree constraint, physical law conformity degree constraint and semantic consistency constraint. The semantic module performs semantic enhancement based on the trained feature decoupling network and the failure physics knowledge base, including: generating enhanced samples using a physical constraint interpolation algorithm, wherein the physical constraint interpolation algorithm dynamically determines the interpolation coefficients according to the physical rules in the failure physics knowledge base; The diagnostic module determines the defect diagnosis result based on the enhanced sample.

9. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor, when executing the computer program, controls the execution of the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed in a computer, causes the computer to control the execution of the method according to any one of claims 1-7.

Images