Generator outlet circuit breaker contact defect intelligent identification method and system

Abstract

The invention relates to the technical field of image processing, in particular to a generator outlet circuit breaker contact defect intelligent identification method and system, and the method comprises the steps: carrying out the image collection of a generator outlet circuit breaker contact from a plurality of preset angles through an industrial camera, and carrying out the preprocessing of the collected image data; defect area labeling is carried out on the contact images in the obtained image sample set, meanwhile, defect types of defect areas are labeled in combination with a preset defect dictionary, and a contact defect feature library is generated; improving a deep learning model based on the constructed feature pyramid fusion structure, and identifying and classifying a defect region in combination with a region suggestion structure to obtain an intelligent contact defect identification model; and real-time image data obtained in real time is preprocessed and then input into the contact defect intelligent identification model to obtain defect positioning and classification, so that on-site complex interference can be overcome, and various defects of the generator outlet circuit breaker contact can be automatically and accurately identified.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for intelligent identification of defects in the contacts of a generator outlet circuit breaker. Background Technology

[0002] The generator output circuit breaker is a critical protection device between the generator set and the main transformer. The health of its contact system directly affects the safe and stable operation of the entire power generation system. During the long-term switching of large currents, the contacts may develop defects due to arc erosion, mechanical wear, overheating oxidation, etc., such as surface burns, welding defects, foreign matter adhesion, and geometric deformation.

[0003] Currently, the condition monitoring of generator outlet circuit breaker contacts mainly relies on the following two methods: Regular power outage maintenance: Circuit breakers are taken out of service according to a fixed maintenance cycle and disassembled by professional technicians. The contacts are then observed and measured visually or with simple tools. This method suffers from significant economic losses due to planned shutdowns, long inspection cycles, and is heavily influenced by the subjective experience of technicians, making it prone to overlooking or misjudging minor defects.

[0004] Online monitoring and basic image analysis: Some advanced solutions attempt to acquire contact images using industrial cameras deployed in the circuit breaker room and perform preliminary analysis using traditional image processing algorithms (such as edge detection and threshold segmentation). However, the complex on-site environment, with its uneven lighting, oil contamination, and variable shooting angles, results in poor robustness of traditional algorithms, making it difficult to reliably extract effective contact features from background noise. Furthermore, traditional methods rely on manually defined features (such as color and shape), failing to adaptively learn the deep, abstract features of contact defects, resulting in insufficient ability to identify early, subtle defects and low levels of intelligence.

[0005] Therefore, existing technologies lack an intelligent method that can overcome complex on-site interference and automatically and accurately identify various defects in the contacts of generator outlet circuit breakers. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent identification method and system for defects in the contacts of generator outlet circuit breakers, which can overcome complex interference on site and automatically and accurately identify various defects in the contacts of generator outlet circuit breakers.

[0007] To achieve the above objectives, in a first aspect, the present invention provides a method for intelligent identification of contact defects in generator outlet circuit breakers, comprising the following steps: Images of the generator outlet circuit breaker contacts are captured using an industrial camera from multiple preset angles. The captured image data is then aligned, fused, and preprocessed to obtain an image sample set. Defect regions are labeled in the contact images of the image sample set, and the defect types of the defect regions are marked in combination with a preset defect dictionary to generate a contact defect feature library. The deep learning model is improved based on the constructed feature pyramid fusion structure, and the defect region is identified and classified by combining the region proposal structure to obtain the intelligent identification model of contact defects. The real-time image data acquired in real time is preprocessed and then input into the intelligent identification model for contact defects to obtain defect location and classification.

[0008] This involves using an industrial camera to capture images of the generator outlet circuit breaker contacts from multiple preset angles, aligning and fusing the acquired image data, and then preprocessing it to obtain an image sample set, including: Under various lighting modes, the macroscopic observation camera and the microscopic detail camera are controlled to be triggered synchronously or sequentially according to instructions, so as to acquire the overall and local images of the contact at the same point in time. Based on the same timestamp, local images at different preset angles are aligned into the overall image, and images at the same angle under different lighting modes are fused to obtain a lighting balance feature map. The illumination equalization feature map is preprocessed to obtain an image sample set.

[0009] The illumination equalization feature map is preprocessed to obtain an image sample set, including: The illumination equalization feature map is processed based on image segmentation technology, and the resulting binarized contact region mask is subjected to adaptive histogram equalization processing. The processed binarized contact area mask is filtered and its size and grayscale are standardized. After data augmentation of the standardized image, an image sample set is obtained.

[0010] Specifically, defect regions are labeled in the contact images of the image sample set, and the defect types of the defect regions are marked using a preset defect dictionary to generate a contact defect feature library, including: The defect regions in the image sample set are labeled using the least bounding polygon annotation method; The defect types in the defect area are matched using a preset defect dictionary at two levels, and the text tags of the secondary defect types are bound to the defect area. A structured annotation description file is generated based on the defect type and defect region, and a contact defect feature library is generated by combining the image sample set.

[0011] The process includes generating a structured annotation description file based on the defect type and defect region, and combining this with the image sample set to generate a contact defect feature library, including: By combining image identification information, defect type, and the geometric coordinates of the defect area, a structured annotation description file is generated; The structured annotation description file is integrated with the corresponding images in the image sample set to obtain the contact defect feature library.

[0012] Among them, the deep learning model is improved based on the constructed feature pyramid fusion structure, and the defect region is identified and classified by combining the region proposal structure, resulting in an intelligent identification model for contact defects, including: The deep feature map output by the deep learning model is upsampled, and the upsampled deep feature map is concatenated and fused with the shallow feature map output by the deep learning model. The constructed region proposal structure is used to scan the multi-level predicted feature map obtained by fusion, and the candidate regions are identified and classified to obtain an intelligent identification model for contact defects.

[0013] The method further includes: The intelligent identification model for contact defects is trained and optimized using the contact defect feature library.

[0014] The process of training and optimizing the intelligent identification model for contact defects using the contact defect feature library includes: All samples in the contact defect feature library are divided into training set and validation set, and the intelligent contact defect recognition model is initialized. The training set is input into the intelligent identification model for contact defects for prediction, and the multi-task joint loss function is calculated. At the same time, the backpropagation algorithm is used to calculate the total value of the multi-task joint loss function and update the parameters. The updated contact defect intelligent identification model is validated and optimized using the validation set.

[0015] The method further includes: The obtained defect locations and defect classifications are visualized, displayed, and stored, and the contact defect feature library is adjusted accordingly.

[0016] Secondly, the present invention provides an intelligent identification system for contact defects of generator outlet circuit breakers, applied to the intelligent identification method for contact defects of generator outlet circuit breakers as provided in the first aspect. The intelligent identification system for contact defects of generator outlet circuit breakers includes an image sample acquisition module, a defect feature library generation module, an identification model construction module, and an intelligent identification module. The image sample acquisition module is used to acquire images of the generator outlet circuit breaker contacts from multiple preset angles using an industrial camera, and to align and fuse the acquired image data before preprocessing to obtain an image sample set. The defect feature library generation module is used to annotate the defect areas of the contact images in the image sample set, and at the same time, mark the defect types of the defect areas in combination with a preset defect dictionary to generate a contact defect feature library. The identification model construction module is used to improve the deep learning model based on the constructed feature pyramid fusion structure, and at the same time combine the region proposal structure to identify and classify the defect region, so as to obtain the intelligent identification model of contact defect. The intelligent recognition module is used to input the real-time image data acquired in real time into the intelligent recognition model of contact defects after preprocessing, so as to obtain defect location and classification.

[0017] This invention discloses an intelligent identification method and system for generator outlet circuit breaker contact defects. The intelligent identification system for generator outlet circuit breaker contact defects includes an image sample acquisition module, a defect feature library generation module, an identification model construction module, and an intelligent identification module. The system acquires images of the generator outlet circuit breaker contacts from multiple preset angles using an industrial camera. The acquired image data is aligned and fused, then preprocessed to obtain an image sample set. Defect regions are labeled in the contact images within the image sample set, and the defect types of the defect regions are marked using a preset defect dictionary, generating a contact defect feature library. A deep learning model is improved based on a constructed feature pyramid fusion structure, and region suggestions are also incorporated. The structure identifies and classifies defective areas to obtain an intelligent contact defect identification model. Real-time image data acquired in real time is preprocessed and input into the intelligent contact defect identification model to obtain defect location and classification. By combining multiple illumination intensities and angles, the interference of on-site reflections or shadows under single illumination is solved. By utilizing the powerful feature learning capabilities of deep convolutional neural networks, key features related to defects can be adaptively extracted from complex backgrounds and interferences. The sensitivity of early and subtle defect identification is much higher than that of traditional image processing methods, and it replaces the traditional manual interpretation that relies on the subjective experience of technicians. It can overcome complex on-site interference and automatically and accurately identify various defects in the contacts of generator outlet circuit breakers. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0019] Figure 1This is a schematic diagram illustrating the steps of a method for intelligent identification of defects in the contacts of a generator outlet circuit breaker according to the first embodiment of the present invention.

[0020] Figure 2 This is a flowchart illustrating an intelligent identification method for contact defects in generator outlet circuit breakers provided by the present invention.

[0021] Figure 3 This is a complete flowchart illustrating the intelligent identification method for contact defects in generator outlet circuit breakers provided by the present invention.

[0022] Figure 4 This is a schematic diagram of the structure of an intelligent identification system for contact defects of a generator outlet circuit breaker according to the second embodiment of the present invention.

[0023] Figure 5 This is a schematic diagram of the electronic device of the present invention.

[0024] In the diagram: 101 - Image sample acquisition module, 102 - Defect feature library generation module, 103 - Recognition model construction module, 104 - Intelligent recognition module. Detailed Implementation

[0025] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0026] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0027] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0028] The first embodiment of this application is as follows: Please see Figures 1-3This invention provides an intelligent identification method for defects in the contacts of a generator outlet circuit breaker, comprising the following steps: S101. Use an industrial camera to acquire images of the generator outlet circuit breaker contacts from multiple preset angles, and then align and fuse the acquired image data before preprocessing to obtain an image sample set.

[0029] Specifically, to ensure the accuracy of the acquired image data, a standardized image acquisition environment was constructed in the maintenance compartment of the generator outlet circuit breaker: Controllable lighting system: Adjustable brightness, multi-angle LED cold light sources are arranged in a ring around the circuit breaker contacts. This is intended to eliminate interference from changes in natural light, ensure consistent lighting conditions for each acquisition, and highlight the microscopic textures (such as minor burns or wear) on the contact surface by adjusting the light angle.

[0030] Fixed multi-view image acquisition device: This device consists of at least two industrial camera modules and is mechanically fixed to the circuit breaker housing to ensure that the shooting position is absolutely repeatable each time.

[0031] Macroscopic observation camera: Equipped with a standard focal length lens, it captures the overall appearance of the contact from a frontal angle to assess large-area defects (such as large-area welding or severe deformation).

[0032] Microscopic detail camera: Equipped with a high-resolution macro lens and small-angle pitch adjustment, it is specifically designed to capture minute features of the contact surface of the contactor at close range, such as point arc ablation, micro cracks, foreign object adhesion, etc.

[0033] After the circuit breaker is shut down and the contacts are safely exposed, an automated acquisition sequence is executed: First, a set of initial-state images is acquired without additional lighting or assistance to record ambient background noise. Then, a controllable lighting system is activated, switching between different lighting angles and brightness combinations according to a preset program for image capture. For example, low-angle grazing light best highlights surface irregularities; uniform front lighting facilitates color and overall contour reproduction. This set of images provides rich raw information for subsequent processing. Under each lighting mode, the macroscopic observation camera and the microscopic detail camera are triggered synchronously or sequentially according to instructions, acquiring overall and partial images of the contacts at the same point in time. All images are automatically timestamped, labeled with the viewing angle and lighting mode.

[0034] The acquired image data is transmitted to the central computing server in real time via the local area network, and is automatically named and stored according to the rule of "device number-acquisition date-serial number" to form a structured raw image database, which is ready for subsequent processing and analysis.

[0035] For images acquired at the same timestamp or at the same time point but from different perspectives (macro and micro), a registration algorithm based on feature points (such as SIFT features) is used to accurately align the micro-detail images to the corresponding regions in the macro-observation images, establishing a pixel-level coordinate mapping relationship between the two. Then, images acquired from the same perspective under different lighting conditions are fused. A weighted average or pyramid-based fusion algorithm is used to synthesize a single, equally lit feature map. This map integrates features revealed under different lighting angles; for example, it includes both clear texture information under grazing light and true color information under frontal light, resulting in a fused image with a much greater amount of information than a single image.

[0036] Based on image segmentation techniques (such as the GrabCut algorithm or training a lightweight U-Net segmentation network), the core contact working surface (i.e., the actual contact and defect-prone area) is precisely separated from the background (such as metal supports and insulating materials) from the fused illumination equalization feature map, generating a binary contact region mask. Then, adaptive histogram equalization is performed within the contact working surface defined by the contact region mask. This operation greatly enhances the dynamic range of gray levels on the contact surface, making defect features such as brightness differences caused by ablation and color changes caused by oxidation extremely vivid, while effectively protecting the background area from over-enhancement interference.

[0037] After enhancement, image noise may also be amplified, thus requiring intelligent filtering, employing anisotropic diffusion filtering. This algorithm can smooth in flat areas to suppress noise based on the image's gradient information, while reducing smoothing in edge areas (such as the boundary between defects and normal areas) to preserve sharpness. This achieves perfect preservation of the key contour information of defects while reducing noise.

[0038] All processed contact surface images were cropped using a binarized contact region mask and scaled to the uniform size required by the model. Simultaneously, grayscale normalization was performed, scaling pixel values ​​to the [0,1] range to accelerate subsequent model training convergence. To ensure model robustness, the standardized images underwent data augmentation using conformal transformations such as small-angle random rotations (within ±5°), horizontal / vertical flipping, and slight brightness and contrast adjustments. These transformations simulated the minute changes in perspective and lighting that might occur during actual shooting, effectively increasing data diversity without distorting the essence of the defects. This resulted in a high-quality, standardized sample set of contact defect images. Each image in this sample set focuses on the key contact region, exhibiting distinct features and controlled noise, providing a solid data foundation for training a high-precision intelligent contact defect recognition model.

[0039] S102. Defect regions are labeled in the contact images of the image sample set, and the defect types of the defect regions are marked in combination with a preset defect dictionary to generate a contact defect feature library.

[0040] Specifically, before commencing the actual labeling work, a standardized dictionary of generator outlet circuit breaker contact defects must first be established. This dictionary serves as the fundamental basis for all subsequent labeling work. Its content is based on equipment mechanisms and historical maintenance data, and was developed by domain experts. It primarily includes: Defect type hierarchical definition: Defects are divided into first-level and second-level types.

[0041] Primary defect types: These refer to major categories of defects, such as "electrolytic erosion defects," "mechanical defects," and "contamination defects."

[0042] Secondary defect types: These are refinements and specificities of primary defect types. For example: Electrical erosion defects -> Arc burns Electrolytic corrosion defects -> localized fusion welding Mechanical defects -> Contact surface deformation Contamination defects -> Conductive foreign matter Contamination defects -> Oil buildup Defect Morphology and Attribute Description: For each type of secondary defect, describe in detail its typical visual morphological characteristics (such as color, texture, and shape), common occurrence locations, and potential impact on equipment function. For example, "arc burn" is described as "irregular pits or pockmarks appearing on the contact surface, usually darker than the surrounding normal area, appearing as dark gray or black, with irregular edges."

[0043] Unified labeling rules: It is clearly stipulated that multiple defects of different types appearing in the same image must be labeled separately, and merging is not allowed. For large-area continuous defects, they should be labeled as a whole; for discretely distributed defects of the same type, the labeling should be determined based on their spacing to determine whether to label them as a single region or multiple independent regions.

[0044] For each image in the image sample set, the least bounding polygon annotation method is used for annotation. That is, point cloud data is generated based on the precise contour of the obtained defect area, and the point cloud data is connected to generate a polygon that tightly wraps around the defect. For irregular defects such as "arc burns", this method can delineate their boundaries with extreme precision; for "contact surface deformation", it can delineate the precise polygon of the deformation area.

[0045] For small defect clusters that are scattered but belong to the same type in the image (such as multiple tiny "conductive foreign objects"), calculate the farthest straight-line distance and the shortest straight-line distance of multiple polygons. If the farthest straight-line distance is less than the set selection threshold, then use a polygon to select the entire cluster. If the shortest straight-line distance is greater than the selection threshold, then use multiple polygons to independently label the clusters.

[0046] After a polygon is drawn, the first-level defect type, such as "electrolytic erosion defect," is matched according to the defect dictionary. Then, the corresponding second-level defect type, such as "arc burn," is matched. Finally, the text label of the second-level defect type is bound to the polygon area.

[0047] Each meticulously annotated contact image generates a unique corresponding annotation description file. This file is a structured data document whose core function is to record the precise information and semantic attributes of all annotated defects in the image in a machine-readable and logically clear manner. This file mainly contains the following parts: Part 1: Image Identification Information. This part is used to uniquely identify the contact image corresponding to this description file and to record the basic attributes of the image.

[0048] Image file name: Records the full name of the corresponding preprocessed contact image file, which is the core link between the description file and the image file.

[0049] Image size information: Records the width and height in pixels of the image. This information provides a spatial reference for the coordinate positioning of subsequent defect annotations.

[0050] Part Two: Defect Annotation Entries. This part is an ordered list that records detailed information about each annotated defect instance in the image. Each entry in the list is a defect annotation unit, fully describing an independent defect.

[0051] Each defect annotation unit contains the following key information: Defect instance number: Within the current image file range, each labeled defect is assigned a unique sequential number to distinguish multiple defects in the same image.

[0052] Defect type hierarchy information: This directly references the contents of the predefined defect dictionary.

[0053] Level 1 Defect Type: Indicates the higher-level category to which the defect belongs, such as "electrolytic erosion defect".

[0054] Secondary defect type: Specifies the specific name of the defect, such as "arc burn". This level of label is the direct target for subsequent model training and recognition.

[0055] Defect Severity Rating: This is an optional but recommended field to record the judgment of the severity of the defect based on established criteria, such as "mild", "moderate" or "severe".

[0056] Defect region geometric coordinates: The result, generated using the minimum bounding polygon annotation method, is an ordered sequence of coordinate points. This sequence sequentially records the two-dimensional coordinates of each vertex of the polygon that outlines the precise contour of the defect. Using this series of coordinates, the actual shape and coverage area of ​​the defect can be completely reproduced in the image.

[0057] Part Three: Labeling metadata for quality traceability and management.

[0058] Personnel Identification: Records the identity information of the engineer or expert who performed this annotation task.

[0059] Completion Date: Record the specific date on which this description file was generated.

[0060] Mark the quality review status: Record whether the marking result has passed the second quality inspection review, for example, mark it as "reviewed" or "pending review".

[0061] A unique correspondence is established between the description file and the image file by labeling the image file name. Within the description file, each defect instance number represents an independent defect instance, and its defect region's geometric coordinates precisely define the spatial location and extent of the defect in the image pixel coordinate system. The secondary defect type defines the semantic information of this region.

[0062] The structured annotation description file is integrated with the corresponding images in the image sample set to obtain the contact defect feature library. The contact defect feature library includes a main sample table, a defect record sub-table, and a dictionary table. First, the main sample table stores the core information of each image sample, such as sample ID, corresponding device number, acquisition time, image file path, and annotation description file path. The defect record sub-table is associated with the main sample table to store information for each specific defect instance, including its sample ID, defect type (level 1 and level 2), severity, and geometric coordinates of the defect area. The dictionary table stores the complete contents of the defect dictionary. The constructed contact defect feature library is no longer just a collection of images, but a structured knowledge base rich in precise spatial and semantic information. It not only provides high-quality training data for supervised learning, but its internal defect dictionary and hierarchical annotation system also enable subsequent intelligent recognition models to learn to understand and locate complex contact defects from an expert perspective, laying a solid foundation for ultimately achieving high-precision, interpretable intelligent diagnosis. S103. Based on the constructed feature pyramid fusion structure, the deep learning model is improved, and at the same time, the defect region is identified and classified by combining the region proposal structure to obtain the intelligent identification model of contact defects.

[0063] Specifically, since existing deep learning models cannot automatically and accurately identify contact defects, improvements are made to existing deep learning models to obtain an intelligent contact defect identification model. The specific improvement method is as follows: The intelligent contact defect recognition model comprises a deep feature perception encoding subnetwork, a multi-scale feature adaptive fusion subnetwork, and a defect region recognition and classification decision subnetwork. The deep feature perception encoding subnetwork acts as the model's "sensory organ," responsible for abstractly extracting visual features from the standardized input contact image layer by layer, from low-level to high-level. It consists of a series of cascaded "convolution-activation-pooling" basic modules. Primary feature perception layers: Located in the first few layers of the network. They use small convolutional kernels to scan every local region of the image, focusing on perceiving and extracting basic visual elements such as edges, corners, and color patches. For contact images, this layer effectively captures the texture of normal metal surfaces and the boundaries of defect regions. High-level feature abstraction layers: As the network deepens, subsequent layers receive increasingly larger receptive fields. They combine primary features into more complex patterns. At this stage, the subnetwork begins to abstract features related to defect semantics, such as the irregular dark spot texture unique to "arc burns," the outline of raised areas formed by "local welding," or the material difference features between "conductive foreign objects" and the metal body. The final output will be a series of feature maps at different scales. Shallow feature maps have high resolution, contain rich details and location information but weak semantics; deep feature maps have low resolution, strong semantics but coarse location information.

[0064] The multi-scale feature adaptive fusion sub-network employs a feature pyramid fusion structure to achieve adaptive fusion of multi-scale features. The specific fusion process involves upsampling the deepest, most semantically strongest feature map output from the deep feature-aware coding sub-network, enlarging its size. This upsampled deep feature map is then concatenated and fused with shallow feature maps of the same scale from the deep feature-aware coding sub-network. Shallow features provide detailed information such as "where the precise edge of this defect is," while deep features provide semantic information such as "this area is likely an arc burn." Through this operation, a series of novel, fused feature maps are obtained. Each fused feature map simultaneously incorporates good semantic information and spatial detail suitable for its scale. These feature maps are uniformly named multi-level prediction feature maps, and they will be responsible for detecting defects of different sizes.

[0065] The defect region identification and classification decision sub-network, based on the fused multi-level predicted feature maps, completes the two final tasks of defect localization and classification. First, a lightweight region proposal network quickly scans each multi-level predicted feature map, identifying numerous candidate regions that may contain defects and generating an initial bounding box for each region. Then, these candidate regions and their corresponding features are fed into a more refined detection head. This part performs two parallel operations: Bounding box regression: Fine-tuning the initial bounding box by making precise coordinate offsets to better match the true contour of the defect (i.e., the bounding rectangle of the smallest bounding polygon used in annotation).

[0066] Defect type classification: Simultaneously, the features within this region are processed to calculate the probability that it belongs to each secondary defect type (such as "arc burn" or "local fusion welding"). Finally, the type with the highest probability is selected as the defect identification result.

[0067] To efficiently and reliably integrate the knowledge from the contact defect feature library into the aforementioned intelligent contact defect recognition model, all samples in the feature library are randomly but evenly divided into a training set and a validation set in an approximately 7:3 ratio. During this division, it is ensured that each defect type appears in approximately equal proportions in both the training and validation sets, avoiding the absence of any particular defect type in the training set. The training set is used directly to adjust the model parameters. The model learns the general patterns of contact defects from these samples. The validation set is not involved in parameter adjustment and is used to objectively evaluate the model's generalization ability during training, preventing overfitting.

[0068] All weight parameters of the model are not initialized from zero, but rather using a strategy called "He normal distribution initialization." This strategy randomly selects initial values ​​from a mathematically optimized normal distribution based on the network layer structure. This provides the model with a high-starting-point "initial intuition," significantly accelerating subsequent convergence and improving final performance.

[0069] The adaptive moment estimation algorithm is selected for model optimization. It not only records the gradient (learning direction) but also the historical momentum of gradient changes, enabling intelligent adjustment of the learning step size (i.e., learning rate) for each parameter. More importantly, a "step-like dynamic learning rate decay" strategy is implemented: in the early stages of training, a large learning rate is used to quickly approach the optimal solution region; when training reaches a certain stage and the validation set loss no longer decreases significantly, the learning rate is automatically reduced by an order of magnitude (e.g., to one-tenth of its original value), allowing the model to hover around the optimal solution with finer step sizes, eventually stabilizing at the optimal point.

[0070] A small batch of contact image samples is extracted from the model training set and input into the current model. The data flows sequentially through the deep feature perception encoding subnetwork and the multi-scale feature adaptive fusion subnetwork. Finally, the defect region identification and classification decision subnetwork outputs two prediction results: the coordinates of each prediction box and the probability distribution of the defect belonging to each category within each prediction box.

[0071] After inputting the training set into the intelligent contact defect recognition model for prediction, the multi-task joint loss function is calculated. The multi-task joint loss function consists of a weighted sum of three parts: Bounding box regression loss: Calculates the positional error between the defect bounding box predicted by the model and the true bounding box of the defect annotation cell in the annotation file. A smooth L1 loss function is used, which is insensitive to small errors but strongly penalizes large errors, allowing the model to robustly learn the accurate location of the boxes.

[0072] Defect classification loss: This calculates the difference between the defect category probability distribution predicted by the model and the true category labels. During training, it monitors in real time the difficulty (manifested as low classification probability) and frequency of each defect category (e.g., "arc burn," "local fusion welding") in the current batch. A dynamic weight is automatically assigned to the calculation results for this defect category. For defect categories that occur infrequently but are difficult to identify, the system assigns a higher loss weight. This forces the model to give equal importance to rare but critical defects during training, rather than focusing all its efforts on identifying common defects, thus significantly improving the model's sensitivity to identifying "long-tailed" defects.

[0073] Consistency Enhancement Regularization Loss: For the same training image, two different, randomly augmented versions are generated (e.g., different brightness, slight rotation). These two versions are simultaneously input into the model, resulting in two predictions. This loss term penalizes the inconsistency between these two predictions regarding the same defect's bounding box location and classification.

[0074] Based on the calculated total value of the multi-task joint loss function, the backpropagation algorithm is used to calculate the gradient of each trainable parameter with respect to the total loss along the computational graph of the model. The gradient indicates the direction and magnitude of adjustment for each parameter. The adaptive moment estimation algorithm calculates a unique update amount for each parameter based on the calculated gradient, historical gradient momentum, and the current dynamic learning rate, and then performs a batch update operation.

[0075] After each training epoch completes a full iteration through all training data, a comprehensive evaluation of the current model is performed using a validation set. The model's average accuracy and joint loss on the validation set are recorded, and the performance of the validation set is continuously monitored. Training automatically terminates immediately once the validation set performance no longer produces new best results over several consecutive training epochs (e.g., 10 epochs). At this point, the system does not save the model from the last iteration but automatically backtracks and loads the model version that performed best on the validation set throughout the entire training process. This mechanism ensures that the final delivered model is the version with the strongest generalization ability, thus obtaining the required intelligent contact defect recognition model.

[0076] S104. The real-time image data acquired in real time is preprocessed and then input into the intelligent identification model of the contact defect to obtain defect location and classification.

[0077] Specifically, when a condition assessment of a generator outlet circuit breaker needs to be performed, real-time or offline image data of its contacts is acquired using an industrial camera, and then subjected to the same enhancement and filtering processing as in step S101. The preprocessed image is then input into a trained and optimized intelligent contact defect recognition model. The model automatically performs feature extraction, defect localization, and classification, and outputs structured recognition results. These results include at least: the presence of a defect, the type of defect, the location of the defect, and the confidence level of the recognition. These results can be visualized through a human-machine interface and stored in a database for generating maintenance reports and equipment condition trend analysis, and for feedback to adjust the contact defect feature library, facilitating the updating of the parameters of the intelligent contact defect recognition model.

[0078] The second embodiment of this application is as follows: Please see Figure 4 This invention provides an intelligent identification system for defects in the contacts of a generator outlet circuit breaker, applied to an intelligent identification method for defects in the contacts of a generator outlet circuit breaker as provided in the first embodiment. The intelligent identification system for defects in the contacts of a generator outlet circuit breaker includes an image sample acquisition module 101, a defect feature library generation module 102, an identification model construction module 103, and an intelligent identification module 104. The image sample acquisition module 101 is used to acquire images of the generator outlet circuit breaker contacts from multiple preset angles using an industrial camera, and to align and fuse the acquired image data before preprocessing to obtain an image sample set. The defect feature library generation module 102 is used to annotate the defect areas of the contact images in the image sample set, and at the same time, mark the defect types of the defect areas in combination with a preset defect dictionary to generate a contact defect feature library. The identification model construction module 103 is used to improve the deep learning model based on the constructed feature pyramid fusion structure, and at the same time combine the region proposal structure to identify and classify the defect region, so as to obtain the intelligent identification model of the contact defect. The intelligent recognition module 104 is used to input the real-time image data acquired in real time into the intelligent recognition model of contact defects after preprocessing, so as to obtain defect location and classification.

[0079] Regarding the system in the above embodiments, the specific manner in which each module performs its operations has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0080] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0081] Accordingly, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; and, when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the intelligent identification method for generator outlet circuit breaker contact defects as described above. Figure 5 The diagram shown is a hardware structure diagram of any device with data processing capabilities, used in an intelligent identification system for faulty contacts of a generator outlet circuit breaker according to an embodiment of the present invention. (Except for...) Figure 5 In addition to the processor, memory, and network interface shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0082] Accordingly, this application also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the intelligent identification method for contact defects of generator outlet circuit breakers as described above. The computer-readable storage medium can be an internal storage unit of any data processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0083] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0084] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for intelligent identification of contact defects in generator outlet circuit breakers, characterized in that, Includes the following steps: Images of the generator outlet circuit breaker contacts are captured using an industrial camera from multiple preset angles. The captured image data is then aligned, fused, and preprocessed to obtain an image sample set. Defect regions are labeled in the contact images of the image sample set, and the defect types of the defect regions are marked in combination with a preset defect dictionary to generate a contact defect feature library. The deep learning model is improved based on the constructed feature pyramid fusion structure, and the defect region is identified and classified by combining the region proposal structure to obtain the intelligent identification model of contact defects. It includes: upsampling the deep feature map output by the deep learning model, concatenating and fusing the upsampled deep feature map with the shallow feature map output by the deep learning model; scanning the multi-level predicted feature map obtained by the fusion using the constructed region proposal structure, and identifying and classifying the obtained candidate regions to obtain a contact defect intelligent recognition model; The intelligent identification model for contact defects is trained and optimized using the contact defect feature library. The real-time image data acquired in real time is preprocessed and then input into the intelligent identification model for contact defects to obtain defect location and defect classification.

2. The intelligent identification method for generator outlet circuit breaker contact defects as described in claim 1, characterized in that, Images of the generator outlet circuit breaker contacts are acquired using an industrial camera from multiple preset angles. The acquired image data is then aligned, fused, and preprocessed to obtain an image sample set, including: Under various lighting modes, the macroscopic observation camera and the microscopic detail camera are controlled to be triggered synchronously or sequentially according to instructions, so as to acquire the overall and local images of the contact at the same point in time. Based on the same timestamp, local images at different preset angles are aligned into the overall image, and images at the same angle under different lighting modes are fused to obtain a lighting balance feature map. The illumination equalization feature map is preprocessed to obtain an image sample set.

3. The intelligent identification method for generator outlet circuit breaker contact defects as described in claim 2, characterized in that, The illumination equalization feature map is preprocessed to obtain an image sample set, including: The illumination equalization feature map is processed based on image segmentation technology, and the resulting binarized contact region mask is subjected to adaptive histogram equalization processing. The processed binarized contact area mask is filtered and its size and grayscale are standardized. After data augmentation of the standardized image, an image sample set is obtained.

4. The intelligent identification method for generator outlet circuit breaker contact defects as described in claim 2, characterized in that, Defect regions are labeled in the contact images of the image sample set, and the defect types of the defect regions are marked using a preset defect dictionary to generate a contact defect feature library, including: The defect regions in the image sample set are labeled using the least bounding polygon annotation method; The defect types in the defect area are matched using a preset defect dictionary at two levels, and the text tags of the secondary defect types are bound to the defect area. A structured annotation description file is generated based on the defect type and defect region, and a contact defect feature library is generated by combining the image sample set.

5. The intelligent identification method for generator outlet circuit breaker contact defects as described in claim 4, characterized in that, A structured annotation description file is generated based on the defect type and defect region, and a contact defect feature library is generated by combining the image sample set, including: By combining image identification information, defect type, and the geometric coordinates of the defect area, a structured annotation description file is generated; The structured annotation description file is integrated with the corresponding images in the image sample set to obtain the contact defect feature library.

6. The intelligent identification method for generator outlet circuit breaker contact defects as described in claim 1, characterized in that, The intelligent identification model for contact defects is trained and optimized using the aforementioned contact defect feature library, including: All samples in the contact defect feature library are divided into training set and validation set, and the intelligent contact defect recognition model is initialized. The training set is input into the intelligent identification model for contact defects for prediction, and the multi-task joint loss function is calculated. At the same time, the backpropagation algorithm is used to calculate the total value of the multi-task joint loss function and update the parameters. The updated contact defect intelligent identification model is validated and optimized using the validation set.

7. The intelligent identification method for generator outlet circuit breaker contact defects as described in claim 1, characterized in that, The method further includes: The obtained defect locations and defect classifications are visualized, displayed, and stored, and the contact defect feature library is adjusted accordingly.

8. A smart identification system for contact defects of generator outlet circuit breakers, applied to the smart identification method for contact defects of generator outlet circuit breakers as described in claim 1, characterized in that, The intelligent identification system for defects in the contacts of the generator outlet circuit breaker includes an image sample acquisition module, a defect feature library generation module, an identification model construction module, and an intelligent identification module. The image sample acquisition module is used to acquire images of the generator outlet circuit breaker contacts from multiple preset angles using an industrial camera, and to align and fuse the acquired image data before preprocessing to obtain an image sample set. The defect feature library generation module is used to annotate the defect areas of the contact images in the image sample set, and at the same time, mark the defect types of the defect areas in combination with a preset defect dictionary to generate a contact defect feature library. The identification model construction module is used to improve the deep learning model based on the constructed feature pyramid fusion structure, and at the same time combine the region proposal structure to identify and classify the defect region, so as to obtain the intelligent identification model of contact defect. The intelligent recognition module is used to input the real-time image data acquired in real time into the intelligent recognition model of contact defects after preprocessing, so as to obtain defect location and classification.

Images