A power equipment metal component defect detection system

Abstract

The embodiment of the present application relates to the technical field of power equipment detection, and provides a kind of power equipment metal component defect detection system, comprising: data acquisition module, data preprocessing module, multistage feature extraction module, model training module, defect detection module, result output module and data storage module.The embodiment of the present application is extracted by the vertical feature extraction architecture of progressive, shallow convolution, deep convolution, hollow convolution and channel attention mechanism are combined, the shallow basic features, middle layer detailed features and deep key features are fully mined from the detection data, the core information of subtle defects of metal component can be effectively captured, the precision and recall rate of defect detection can be significantly improved, and the miss rate and false detection rate are reduced, meanwhile, the feature extraction process of each level is dependent on each other, gradually deepening, which avoids the independence defect of traditional single level feature extraction.

Description

Technical Field

[0001] This invention relates to the field of power equipment testing technology, and in particular to a defect detection system for metal components of power equipment. Background Technology

[0002] Metal components in power equipment such as transformers, circuit breakers, and disconnectors, including metal housings, connecting bolts, and busbars, are core components for the safe and stable operation of power systems. Their operating condition directly affects the reliability of the entire power system. During long-term operation, metal components are susceptible to various factors such as environmental corrosion, mechanical wear, electrical aging, and temperature changes, which can easily lead to defects such as cracks, corrosion, porosity, and deformation. If these defects are not detected and addressed in a timely manner, they can develop and cause the metal components to fail, resulting in power equipment malfunctions or even large-scale power outages, causing significant economic losses and safety hazards.

[0003] Currently, defect detection methods for metal components in power equipment mainly include traditional detection methods and conventional intelligent detection methods. Traditional detection methods, such as manual visual inspection, ultrasonic testing, and radiographic testing, rely on the experience of workers, resulting in low efficiency, high labor intensity, and low accuracy in detecting minute defects, with high rates of missed and false detections, failing to meet the high-efficiency inspection requirements for large-scale power equipment metal components. Conventional intelligent detection methods typically employ deep learning models for defect detection, training the model by extracting features from metal component inspection data to achieve automated detection. However, existing intelligent detection methods still have the following significant shortcomings: 1. Insufficient feature extraction: Existing models mostly use shallow feature extraction or single-level feature extraction methods, which cannot fully explore the deep features in the detection data. Shallow features are prone to losing key information about subtle defects in metal components, resulting in the model's inability to accurately identify subtle defects and low detection accuracy and recall. 2. Lack of progressiveness in feature extraction: In the existing feature extraction process, there is no clear vertical progressive relationship between features at different levels. Features at different levels are independent of each other, and it is impossible to achieve gradual refinement and deepening of features. This will lead to insufficient correlation and completeness of deep features, affecting the detection performance of the model. 3. Low utilization of historical data: Existing technologies do not preprocess historical inspection data comprehensively enough. They do not perform effective data cleaning, enhancement and standardization, which leads to noise and outliers in historical inspection data interfering with feature extraction and model training. At the same time, the sample diversity is insufficient and the model has poor generalization ability, which cannot meet the defect detection needs of metal components of different power equipment under different working conditions. 4. Incomplete defect detection: Existing methods often only target a single type of detection data, such as image data, without integrating multiple types of detection data (such as ultrasonic and infrared thermal imaging data). This results in incomplete acquisition of defect features, making it impossible to accurately assess the severity of defects. Furthermore, the lack of effective abnormal alarm and remote reminder mechanisms makes it difficult for staff to handle defects in a timely manner.

[0004] Therefore, how to overcome the aforementioned shortcomings of current technology, solve the problems existing in the current technology, and ensure the safe and stable operation of the power system has become an urgent problem to be solved in this field. Summary of the Invention

[0005] The present invention aims to solve at least one of the problems existing in the prior art by providing a defect detection system for metal components of power equipment, which solves the problems of insufficient feature extraction, lack of progressive feature extraction, low utilization of historical data, and incomplete defect detection in the prior art.

[0006] This invention provides a defect detection system for metal components of power equipment, comprising: The data acquisition module is used to collect historical test data and data to be tested of metal components of power equipment. Both historical test data and data to be tested include data of multiple data types. The data preprocessing module is used to preprocess historical detection data to obtain standardized preprocessed data, divide the standardized preprocessed data into training set, validation set and test set, and preprocess the data to be detected to obtain standardized data to be detected. The multi-level feature extraction module is used to extract the deep features corresponding to the standardized preprocessed data and the standardized data to be detected in a vertically progressive manner through a sequential feature extraction architecture. The model training module is used to construct an improved deep learning model as a defect detection model based on the deep features corresponding to the standardized preprocessed data, and to train the defect detection model based on the training set, validation set and test set, optimize the model parameters, and obtain the trained defect detection model. The defect detection module is used to input the deep features corresponding to the standardized data to be detected into the trained defect detection model to obtain the defect detection results of the metal components. The result output module is used to receive the defect detection results of metal components, visualize the defect detection results of metal components and provide abnormal alarms, so as to remind staff to deal with defects in a timely manner. The data storage module is used to store various types of data during the defect detection process of metal components in power equipment, so as to facilitate subsequent data traceability, model optimization and defect analysis.

[0007] Optionally, the data can be of multiple data types, including image data, ultrasound detection data, and infrared thermal imaging data.

[0008] Optionally, the data preprocessing module is used to preprocess historical detection data to obtain standardized preprocessed data, and divides the standardized preprocessed data into training set, validation set, and test set, including: The data preprocessing module is specifically used for: Data cleaning: Outliers in historical detection data are removed using the 3σ principle, noise in image data and infrared thermal imaging data in historical detection data is removed using the median filtering algorithm, and missing values ​​in ultrasonic detection data in historical detection data are filled using linear interpolation, resulting in cleaned historical detection data; Data standardization: Normalize the cleaned historical test data; Data augmentation: Perform data augmentation on the normalized historical detection data to improve the diversity of training samples; Data partitioning: The historical detection data after data augmentation is divided into training set, validation set and test set according to a preset ratio. The training set is used for model training, the validation set is used for model parameter adjustment, and the test set is used for model performance verification.

[0009] Optionally, the multi-level feature extraction module includes a first-level feature extraction unit, a second-level feature extraction unit, and a third-level feature extraction unit; The first-level feature extraction unit includes shallow convolutional layers and pooling layers of the convolutional neural network, which are used to perform shallow feature extraction on the standardized preprocessed data and the standardized data to be detected, respectively, to obtain the corresponding first-level features. The first-level features include the contour features, surface texture features, basic grayscale features and preliminary signal amplitude features of the metal components. The second-level feature extraction unit includes a deep convolutional layer and a dilated convolutional layer, which are used to extract mid-level features based on the first-level features to obtain the second-level features. The second-level features include subtle changes in the contour edges of the metal components, gray-level differences in local areas, frequency features of ultrasonic signals, and temperature gradient features of infrared thermal imaging. The second-level features are a further refinement and deepening of the first-level features. The third-level feature extraction unit includes an attention mechanism fusion layer and a global pooling layer, used for deep feature extraction based on the second-level features to obtain deep features. The attention mechanism fusion layer uses a channel attention mechanism to assign weights to different channels of the second-level features, strengthening the weights of defect-related features and suppressing the interference of irrelevant features. The obtained channel weights are multiplied with the second-level features channel by channel to obtain attention-enhanced features. The global pooling layer performs a global average pooling operation on the attention-enhanced features to obtain fixed-dimensional deep features. The channel attention mechanism includes a squeezing operation and an activation operation. The squeezing operation is implemented through global average pooling, and the activation operation is implemented through a fully connected operation, ReLU activation, a fully connected operation, and sigmoid activation in sequence.

[0010] Optionally, the defect detection model is an improved convolutional neural network and Transformer fusion model, which includes an input layer, a feature fusion layer, a fully connected layer, a dropout layer, and an output layer; wherein, The input layer is used to receive the deep features output by the multi-level feature extraction module; The feature fusion layer adopts the encoder structure of Transformer to perform global feature association learning on deep features and output the fused global deep features. The encoder structure includes a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism is used to capture the global dependencies between deep features, and the feedforward neural network is used to perform nonlinear transformation on the features output by the multi-head self-attention mechanism. The fully connected layer consists of multiple fully connected layers with different output dimensions, and each fully connected layer is followed by a batch normalization layer and a ReLU activation function; The dropout layer is placed after the second fully connected layer to prevent the model from overfitting. The output layer uses the softmax activation function to output the probability values ​​of various defects in the metal components of power equipment.

[0011] Optionally, the model training module is specifically used for: Initialize the parameters of the defect detection model using Xavier normal initialization. Input the deep features corresponding to the training set into the defect detection model to obtain the corresponding prediction output; The cross-entropy loss function is used to calculate the loss value between the predicted output and the corresponding label in the training set; The Adam optimizer is used to iteratively update the parameters of the defect detection model. The learning rate of the Adam optimizer is decayed using a cosine annealing strategy. In each iteration, a preset number of rounds are used to verify the performance of the defect detection model using a validation set. The validation accuracy and validation loss of the defect detection model are calculated. If the validation loss does not decrease in a specified number of consecutive rounds, an early stopping strategy is used to stop training in order to avoid model overfitting. After training stops, the defect detection model is tested using a test set to determine its performance. If the model performance meets the preset performance requirements, the trained defect detection model is obtained. If the model performance does not meet the preset performance requirements, the model parameters are adjusted and the defect detection model is retrained until the model performance meets the preset performance requirements.

[0012] Optionally, the defect detection module is specifically used for: The deep features corresponding to the standardized data to be detected are input into the trained defect detection model to obtain the probability values ​​of various defects of the metal components of power equipment output by the model. The defect category with the highest probability value is determined as the defect type of the metal component. At the same time, the location information and severity of the defect are output as the defect detection result of the metal component. If the defect detection result of the metal component is that there are no defects, a normal inspection report is generated and output to the result output module. If the defect detection results of the metal component indicate the presence of a defect, the defect type, location information, and severity of the defect are marked, and a defect detection report and abnormal signal are generated and output to the result output module.

[0013] Optionally, the result output module includes: A visual display is used to show the defect detection results, corresponding defect location annotation maps, defect detection reports, and historical detection data query results in real time. An alarm device is used to automatically trigger an alarm when the defect results of a metal component indicate the presence of a defect of a preset severity, so as to remind staff to handle the situation in a timely manner.

[0014] Optionally, the alarm device may employ an audible and visual alarm method.

[0015] Optionally, the alarm device is also used to send alarm information to the staff's mobile terminal via a wireless communication module.

[0016] Compared with the prior art, the present invention has the following advantages: Precise and efficient deep feature extraction: Through a vertical feature extraction architecture that progresses through the first, second, and third levels, combined with shallow convolution, deep convolution, dilated convolution, and channel attention mechanisms, it fully mines shallow basic features, mid-level refined features, and deep key features from the detection data, effectively capturing the core information of subtle defects in metal components, significantly improving the accuracy and recall of defect detection, and reducing the false negative rate and false positive rate. Highlighting feature correlation and completeness: The feature extraction process at each level is interdependent and progressively deepened, avoiding the independence defects of traditional single-level feature extraction, ensuring the correlation and completeness of deep features, providing comprehensive feature support for the model to accurately judge defects, and optimizing the model's detection performance; Make full use of historical data: Through the whole process of data cleaning, standardization, enhancement and segmentation, noise and outliers in historical detection data are removed, data format is unified, sample diversity is enriched, the effectiveness of historical data is improved, and the generalization ability of the model is enhanced, so that it can adapt to the defect detection needs of metal components of different types of power equipment under different working conditions. Comprehensive Defect Detection: It integrates image data, ultrasonic testing data, and infrared thermal imaging data of metal components to achieve all-round detection and evaluation of defect type, location, and severity; coupled with visualization, audible and visual alarms, and remote mobile terminal reminder mechanisms, it ensures that staff can obtain defect information in a timely manner and handle it quickly, preventing defects from escalating and causing equipment failure. Combining automation and practicality: The entire process of data acquisition, preprocessing, feature extraction, defect identification, and result output is automated, requiring minimal manual intervention, reducing the workload of staff, and improving testing efficiency. It can meet the high-efficiency testing needs of metal components in large-scale power equipment, providing a reliable guarantee for the safe and stable operation of the power system. Attached Figure Description

[0017] One or more embodiments are illustrated by way of example with the corresponding pictures in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0018] Figure 1 This is a schematic diagram of a defect detection system for metal components of power equipment provided in one embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the various embodiments of the present invention to facilitate a better understanding of the invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and with various variations and modifications based on the following embodiments. The division of the various embodiments below is for ease of description and should not constitute any limitation on the specific implementation of the present invention. The various embodiments can be combined with and referenced by each other without contradiction.

[0020] One embodiment of the present invention provides a defect detection system for metal components of power equipment, such as... Figure 1 As shown, it includes a data acquisition module, a data preprocessing module, a multi-level feature extraction module, a model training module, a defect detection module, a result output module, and a data storage module.

[0021] The data acquisition module is used to collect historical test data and data to be tested of metal components of power equipment. Both historical test data and data to be tested include data of multiple data types.

[0022] The data preprocessing module performs preprocessing operations on historical detection data to remove noise and outliers, standardize data format, and enhance sample diversity, resulting in standardized preprocessed data. This standardized preprocessed data is then divided into training, validation, and test sets. Finally, it performs preprocessing operations on the data to be detected to obtain standardized data for testing. Specifically, the preprocessing operations on historical detection data can include data cleaning, data standardization, data augmentation, and data partitioning. The preprocessing operations on the data to be detected can include data cleaning and data standardization.

[0023] The multi-level feature extraction module is used to extract deep features from the standardized preprocessed data and the standardized data to be detected in a vertically progressive manner through a sequentially progressive feature extraction architecture. Specifically, the multi-level feature extraction module can achieve vertically progressive extraction of deep features through a sequentially progressive first-level, second-level, and third-level feature extraction architecture, where the second-level features are extracted based on the first-level features, and the third-level features are extracted based on the second-level features.

[0024] The model training module is used to construct an improved deep learning model as a defect detection model based on the deep features corresponding to the standardized preprocessed data obtained by the multi-level feature extraction module. The defect detection model is trained based on the training set, validation set, and test set to optimize the model parameters and obtain the trained defect detection model.

[0025] The defect detection module is used to input the deep features corresponding to the preprocessed and multi-level feature extraction data to be detected (i.e., the standardized data to be detected) into the trained defect detection model to realize the automated detection of defects in metal components and obtain the defect detection results of metal components. The defect detection results of metal components may include the defect type, defect location information and severity of the defect.

[0026] The results output module is used to receive the defect detection results of metal components, visualize the defect detection results and issue abnormal alarms, so as to promptly remind staff to handle the defects.

[0027] The data storage module is used to store various types of data during the entire operation of the power equipment metal component defect detection system, i.e., the defect detection process of power equipment metal components, so as to facilitate subsequent data traceability, model optimization and defect analysis.

[0028] For example, the data includes multiple data types, such as image data, ultrasonic testing data, and infrared thermal imaging data. In other words, both historical testing data and the data to be tested include image data, ultrasonic testing data, and infrared thermal imaging data of the metal components. By setting both historical testing data and the data to be tested to include image data, ultrasonic testing data, and infrared thermal imaging data of the metal components, a comprehensive assessment of defect type, location, and severity can be achieved by fusing these data, avoiding detection bias caused by incomplete defect feature acquisition.

[0029] For example, the data preprocessing module is used to preprocess historical detection data to obtain standardized preprocessed data, and divides the standardized preprocessed data into training set, validation set, and test set, including: The data preprocessing module is specifically used for: Data cleaning: Outliers in historical detection data are removed using the 3σ principle, noise in image data and infrared thermal imaging data in historical detection data is removed using the median filtering algorithm, and missing values ​​in ultrasonic detection data in historical detection data are filled using linear interpolation, resulting in cleaned historical detection data; Data standardization: The cleaned historical detection data is normalized. Specifically, the pixel values ​​of the image data and infrared thermal imaging data in the cleaned historical detection data are normalized to the [0,1] interval, and the amplitude of the ultrasonic detection data in the cleaned historical detection data is normalized to the [0,1] interval. The normalization formula is as follows: ;in, Here, x represents the data after normalization, and 'x' represents the data before normalization. Let x be the minimum value. Let x be the maximum value; Data augmentation: Data augmentation is performed on the normalized historical detection data to enhance the diversity of training samples. Specifically, image data in the normalized historical detection data can be augmented using random rotation, horizontal flipping, vertical flipping, brightness adjustment, and random cropping. Ultrasonic detection data in the normalized historical detection data can be augmented by adding random Gaussian noise and signal shifting. Infrared thermal imaging data in the normalized historical detection data can be augmented using contrast adjustment and temperature threshold stretching to enhance the diversity of training samples. Data partitioning: The historical detection data after data augmentation is divided into training set, validation set and test set according to a preset ratio. The training set is used for model training, the validation set is used for model parameter adjustment, and the test set is used for model performance verification. The preset ratio can be set according to actual needs. For example, the preset ratio can be set to 7:2:1.

[0030] By utilizing the data preprocessing module to perform preprocessing operations on historical inspection data, including data cleaning, data standardization, data augmentation, and data partitioning, noise and outliers in the historical inspection data can be removed, data formats can be standardized, sample diversity can be enriched, the effectiveness of historical inspection data can be improved, and the generalization ability of the defect detection model can be enhanced, enabling it to adapt to the defect detection needs of metal components of different types of power equipment under different working conditions.

[0031] Specifically, the data preprocessing module performs preprocessing operations on the data to be tested, including data cleaning and data standardization, to improve the quality of the data and ensure it meets the input requirements of the defect detection model, thereby improving the accuracy of defect detection. Specifically, when cleaning the data, the method used for cleaning historical test data can be referenced. Outliers are removed using the 3σ principle, noise in the image data and infrared thermal imaging data is removed using a median filtering algorithm, and missing values ​​in the ultrasonic test data are filled using linear interpolation, resulting in cleaned data. Similarly, when standardizing the data, the method used for standardizing historical test data can be referenced. The pixel values ​​of the image data and infrared thermal imaging data in the cleaned data are normalized to the [0,1] interval, and the amplitude values ​​of the ultrasonic test data in the cleaned data are normalized to the [0,1] interval.

[0032] For example, the multi-level feature extraction module includes a first-level feature extraction unit, a second-level feature extraction unit, and a third-level feature extraction unit.

[0033] The first-level feature extraction unit comprises shallow convolutional layers and pooling layers of a convolutional neural network. These layers are used to extract shallow features from the standardized preprocessed data and the standardized data to be detected, respectively, to obtain the corresponding first-level features. These first-level features include the contour features, surface texture features, basic grayscale features, and preliminary signal amplitude features of the metal component. For example, the shallow convolutional layers may include three convolutional layers with a kernel size of 3×3, a stride of 1, and the same padding method. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. A max-pooling layer with a kernel size of 2×2 and a stride of 2 can then be connected after the shallow convolutional layers, thus forming the first-level feature extraction unit.

[0034] The second-level feature extraction unit includes a deep convolutional layer and a dilated convolutional layer, used to extract mid-level features based on the first-level features to obtain second-level features. These second-level features include subtle changes in the contour edges of the metal component, gray-level differences in local areas, frequency characteristics of ultrasonic signals, and temperature gradient characteristics of infrared thermal imaging. The second-level features further refine and deepen the first-level features. Specifically, the deep convolutional layer can use a grouped convolution method, with a kernel size of 5×5, a stride of 1, a padding method of "same," and a group number of 4. Following the deep convolutional layer is a batch normalization layer and a ReLU activation function, followed by two dilated convolutional layers with kernel sizes of 3×3 and dilation rates of 2 and 3, respectively, to expand the receptive field and capture a wider range of feature associations.

[0035] The third-level feature extraction unit includes an attention mechanism fusion layer and a global pooling layer, used for deep feature extraction based on the second-level features to obtain deep features. The attention mechanism fusion layer uses a channel attention mechanism to assign weights to different channels of the second-level features, strengthening the weights of defect-related features and suppressing interference from irrelevant features. The obtained channel weights are then multiplied channel-by-channel with the second-level features to obtain attention-enhanced features. The global pooling layer performs global average pooling on the attention-enhanced features to obtain fixed-dimensional deep features. The channel attention mechanism includes squeezing and activation operations. The squeezing operation is implemented through global average pooling, and the activation operation is implemented through a fully connected layer + ReLU activation + fully connected layer + sigmoid activation, i.e., a fully connected operation, ReLU activation, fully connected operation, and sigmoid activation performed sequentially.

[0036] Through a vertical feature extraction architecture that progressively advances through first-level, second-level, and third-level feature extraction units, combined with shallow convolution, deep convolution, dilated convolution, and channel attention mechanisms, the system can fully extract shallow basic features, mid-level refined features, and deep key features from the detection data. This effectively captures the core information of subtle defects in metal components, significantly improving the accuracy and recall of defect detection while reducing the false negative and false positive rates. Furthermore, the feature extraction processes at each level are interdependent and progressively deepened, avoiding the independence inherent in traditional single-level feature extraction. This ensures the correlation and completeness of deep features, providing comprehensive feature support for the model to accurately identify defects and optimizing model detection performance.

[0037] For example, the defect detection model is an improved convolutional neural network and Transformer fusion model, which includes an input layer, a feature fusion layer, a fully connected layer, a dropout layer, and an output layer.

[0038] The input layer is used to receive the deep features output by the multi-level feature extraction module.

[0039] The feature fusion layer employs a Transformer encoder structure to perform global feature association learning on deep features, outputting fused global deep features. The encoder structure includes a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism captures global dependencies between deep features, while the feedforward neural network performs non-linear transformations on the features output by the multi-head self-attention mechanism. For example, the multi-head self-attention mechanism may include four attention heads.

[0040] A fully connected layer consists of multiple fully connected layers with different output dimensions, each followed by a batch normalization layer and a ReLU activation function. For example, a fully connected layer may include two fully connected layers with different output dimensions, where the output dimension of the first fully connected layer can be set to 512 and the output dimension of the second fully connected layer can be set to 128.

[0041] The dropout layer is placed after the second fully connected layer to prevent the model from overfitting. For example, the dropout probability of the dropout layer can be set to 0.5.

[0042] The output layer uses the softmax activation function to output the probability values ​​of various defects in the metal components of the power equipment. The output dimension of the output layer is N, where N is the total number of defect categories in the metal components of the power equipment. Defect categories can include cracks, corrosion, porosity, deformation, and no defects. The defect category with the highest probability value is the detected defect type of the metal component.

[0043] By setting the defect detection model as a fusion model of an improved convolutional neural network and a Transformer, the model's detection accuracy can be further improved, thus enhancing the precision of defect detection.

[0044] For example, the model training module is specifically used for: Initialize the parameters of the defect detection model using Xavier normal initialization. The parameters of the initialized defect detection model may include convolution kernel parameters, fully connected layer weights, and Transformer encoder parameters. The deep features corresponding to the training set are input into the defect detection model to obtain the corresponding prediction output. The deep features corresponding to the training set are obtained by extracting features from the standardized preprocessed data corresponding to the training set using a multi-level feature extraction module. The cross-entropy loss function is used to calculate the loss between the predicted output and the corresponding labels in the training set. The cross-entropy loss function can be expressed as: ; in, Here, M is the loss value, M is the number of samples in the training set, and N is the total number of defect categories in the metal components of the power equipment. Let be the true label of the i-th sample. If the i-th sample belongs to the j-th type of defect, then... ,otherwise , This is the probability value of the i-th sample in the predicted output belonging to the j-th type of defect; The Adam optimizer is used to iteratively update the parameters of the defect detection model. The learning rate of the Adam optimizer is decayed using a cosine annealing strategy. For example, the initial learning rate of the Adam optimizer can be set to 0.001, the weight decay coefficient can be set to 0.0001, the batch size can be set to 32, and the number of iterations can be set to 100 rounds. Each iteration has a preset number of rounds, such as 10 rounds. The performance of the defect detection model is verified using a validation set. The validation accuracy and validation loss of the defect detection model are calculated. If the validation loss does not decrease for a specified number of rounds, such as 15 rounds, an early stopping strategy is used to stop training in order to avoid model overfitting. After training stops, the defect detection model is tested using a test set to determine its performance. If the model performance meets the preset performance requirements, the trained defect detection model is considered complete. If the model performance does not meet the preset performance requirements, the model parameters, such as the number of convolutional kernels, the number of attention heads, and the learning rate, are adjusted, and training is repeated until the model performance meets the preset performance requirements. For example, model performance can be reflected by detection precision, recall, F1 score, and false negative rate. If the detection precision, recall, F1 score, and false negative rate meet the corresponding thresholds, such as detection precision ≥ 98%, recall ≥ 97%, F1 score ≥ 97.5%, and false negative rate ≤ 1%, then the model performance is considered to meet the preset performance requirements; otherwise, the model performance is considered not to meet the preset performance requirements.

[0045] This implementation method, by training the defect detection model, can further enhance the model's generalization ability and improve its detection accuracy and precision.

[0046] For example, the defect detection module is specifically used to: input the deep features corresponding to the standardized data to be detected into the trained defect detection model, obtain the probability values ​​of various defects of the metal components of the power equipment output by the model, determine the defect category with the highest probability value as the defect type of the metal component, and output the location information and severity of the defect as the defect detection result of the metal component; if the defect detection result of the metal component shows no defects, a normal detection report is generated and output to the result output module; if the defect detection result of the metal component shows that defects exist, the defect type, location information and severity of the defect are marked, and a defect detection report and abnormal signal are generated and output to the result output module.

[0047] Based on this, the defect detection process for metal components of power equipment specifically includes the following steps: The first step is to use a data acquisition module to collect the data to be tested from the metal components of the power equipment. The data to be tested includes real-time image data, ultrasonic test data, and infrared thermal imaging data. The second step is to input the data to be tested into the data preprocessing module for data cleaning and data standardization to obtain standardized data to be tested. The third step is to input the standardized data to be detected into the multi-level feature extraction module, which sequentially passes through the first-level feature extraction unit, the second-level feature extraction unit, and the third-level feature extraction unit to obtain the depth features to be detected. The fourth step involves inputting the depth features to be detected into the trained defect detection model to obtain the probability values ​​of various defects in the metal components of the power equipment. The defect category with the highest probability value is determined as the defect type of the metal component. At the same time, the location information and severity of the defect are output as the defect detection result of the metal component. Among them, the location information can be based on coordinate positioning based on image data; the severity can be comprehensively evaluated based on ultrasonic signal amplitude, infrared temperature gradient and defect area, such as being divided into three levels: slight, moderate and severe. Fifth step: If the defect detection result of the metal component is that there is no defect, a normal inspection report is generated and output to the result output module; if the defect detection result of the metal component is that there is a defect, the defect type, defect location information and severity of the metal component are marked, and a defect detection report and abnormal signal are generated and output to the result output module.

[0048] Finally, the result output module receives the defect detection results of metal components, visualizes the results and issues alarms to promptly remind staff to address the defects.

[0049] This implementation method can automate the entire process of data acquisition, preprocessing, feature extraction, defect identification, and result output without requiring extensive manual intervention. It can reduce the labor intensity of staff, improve detection efficiency, meet the high-efficiency detection needs of metal components in large-scale power equipment, and provide a reliable guarantee for the safe and stable operation of the power system.

[0050] For example, the result output module includes a visual display and an alarm device. The visual display is used to show the defect detection results, corresponding defect location annotations, defect detection reports, and historical detection data query results in real time. The alarm device is used to automatically trigger an alarm when the defect results of the metal component indicate the presence of a defect of a preset severity, so as to remind the staff to handle it in a timely manner. The alarm device can be an audible and visual alarm, automatically triggering an alarm when a moderate or higher-level defect is detected, reminding the staff to handle it promptly. The alarm device can also be used to send alarm information to the staff's mobile terminal via a wireless communication module.

[0051] This implementation method, by combining visual display, audible and visual alarms, and remote mobile terminal reminder mechanisms, can ensure that staff can obtain defect information in a timely manner and quickly handle defects, thus preventing defects from escalating and causing equipment failure.

[0052] The defect detection system for metal components of power equipment provided by the present invention will be further described below through Examples 1 and 2.

[0053] Example 1: Specific configuration and operation of each module of the power equipment metal component defect detection system.

[0054] (1) Data acquisition module: Three types of equipment are used to collect data in a coordinated manner: ① a high-definition industrial camera, which is mounted on a fixed bracket around the transformer to collect image data of the metal shell surface; ② an ultrasonic detector, which is attached to the shell surface with a magnetic probe to collect ultrasonic reflection signal data; and ③ an infrared thermal imager, which is coaxially mounted with the high-definition industrial camera to collect infrared thermal imaging data of the shell surface.

[0055] Historical inspection data collection: Collect inspection data of 10 transformers over the past 6 years, including 8,000 sets of image data, 8,000 sets of ultrasonic inspection data and 8,000 sets of infrared thermal imaging data, covering normal status and various defect status data under different seasons and operating conditions.

[0056] Data collection for testing: Real-time data collection is conducted on 10 transformers once a month, with each collection consisting of 30 sets of image data, 30 sets of ultrasonic test data, and 30 sets of infrared thermal imaging data for each transformer.

[0057] (2) Data preprocessing module: Data cleaning: Outlier screening was performed on 8,000 sets of historical test data using the 3σ principle, removing 320 sets of data with abnormal ultrasound amplitude and infrared temperature; salt-and-pepper noise in image data and infrared thermal imaging data was removed using a 5×5 template median filtering algorithm; and 180 missing values ​​in ultrasound test data were filled using linear interpolation.

[0058] Data standardization: According to the normalization formula, the pixel values ​​of the cleaned image data and infrared thermal imaging data are normalized to the [0,1] interval, and the amplitude of the cleaned ultrasound detection data is normalized to the [0,1] interval to eliminate the difference in data dimensions.

[0059] Data augmentation: The normalized image data was randomly rotated (0-90 degrees), horizontally flipped, vertically flipped, and its brightness was adjusted by ±15%, with random cropping to 200×200 pixels. Random Gaussian noise with a mean of 0 and a variance of 0.01 was added to the normalized ultrasound detection data, and signal shifting was performed at ±5 sampling points. The normalized infrared thermal imaging data was adjusted for contrast by ±20% and stretched to a temperature threshold of 10℃-60℃. After data augmentation, the total amount of historical detection data was expanded to 24,000 sets.

[0060] Data partitioning: The 24,000 sets of historical detection data obtained after data augmentation were divided into training set, validation set and test set in a ratio of 7:2:1. The training set contains 16,800 sets of historical detection data, the validation set contains 4,800 sets of historical detection data, and the test set contains 2,400 sets of historical detection data.

[0061] (3) Multi-level feature extraction module: The first-level feature extraction unit consists of three convolutional layers with a kernel size of 3×3, a stride of 1, and a padding mode of "same". Each convolutional layer in the shallow convolutional layer is followed by a batch normalization layer and a ReLU activation function. After the shallow convolutional layer, a max pooling layer with a kernel size of 2×2 and a stride of 2 is connected. The first-level feature extraction unit is used to extract first-level features from the normalized preprocessed data, i.e., the normalized detection data, including normalized historical detection data and the data to be detected. These features include the contour features of the transformer metal shell, surface texture features, basic grayscale features, and preliminary amplitude features of the ultrasonic signal.

[0062] The second-level feature extraction unit employs a deep convolutional layer with 4 groups, a kernel size of 5×5, a stride of 1, and a padding method of "same". After the deep convolutional layer, a batch normalization layer and a ReLU activation function are connected. Subsequently, two dilated convolutional layers with kernel sizes of 3×3 and dilatations of 2 and 3 are connected. Based on the first-level features, the second-level feature extraction unit extracts second-level features, including features of fine cracks on the shell contour edge, gray-level differences caused by local corrosion, ultrasonic signal frequency features, and infrared thermal imaging temperature gradient features.

[0063] The third-level feature extraction unit performs a compression operation through global average pooling to compress the channel dimension of the second-level features. Then, it performs an activation operation through a fully connected layer of dimension 256 + ReLU activation + a fully connected layer of dimension 64 + sigmoid activation to obtain the channel weights. After multiplying each channel weight with the second-level features channel by channel, it inputs the result into the global average pooling layer and outputs a 256-dimensional fixed-dimensional deep feature.

[0064] (4) Model training module: Model Structure: An improved convolutional neural network and Transformer fusion model is constructed as the defect detection model, consisting of an input layer, a feature fusion layer, a fully connected layer, a dropout layer, and an output layer. The input layer receives 256-dimensional deep features; the feature fusion layer uses a multi-head self-attention mechanism with four attention heads and a feedforward neural network with a hidden layer dimension of 1024; the fully connected layer consists of two layers with output dimensions of 512 and 128, respectively, and each layer is followed by a batch normalization layer and a ReLU activation function; the dropout layer has a dropout probability of 0.5; the output layer uses a softmax activation function and has an output dimension of 5, corresponding to the five defect categories: crack, corrosion, porosity, deformation, and no defect.

[0065] Training process: The Xavier normal model was used to initialize the model parameters; the initial learning rate was set to 0.001, and a cosine annealing strategy was used to decay the learning rate with a weight decay coefficient of 0.0001. The batch size was set to 32, and the number of iterations was set to 100 rounds. The defect detection model was validated with a validation set every 10 rounds. In the 85th round, the loss did not decrease for 15 consecutive rounds, triggering an early stopping strategy to stop training. The defect detection model was tested using a test set. The performance test results were: detection accuracy of 98.7%, recall of 98.1%, F1 score of 98.4%, and false negative rate of 0.8%. These performance test results met the preset thresholds in the preset performance requirements, and the trained defect detection model was obtained.

[0066] (5) Defect detection module: Preprocessing of the data to be tested: For the 30 sets of data to be tested for each transformer acquired in real time, including image data, ultrasonic test data and infrared thermal imaging data, 3σ outlier removal, median filtering for noise reduction, linear interpolation to fill missing values ​​are performed, and then normalized to the [0,1] interval. No data augmentation or data partitioning is performed.

[0067] Feature extraction and detection: The preprocessed data to be detected is input into the multi-level feature extraction module to obtain the depth features to be detected. The depth features to be detected are input into the trained defect detection model to obtain the probability values ​​of various defects output by the model. The defect category with the highest probability value is determined as the metal component defect type. The defect location is located based on the pixel coordinates of the image data. The severity is comprehensively evaluated by combining the ultrasonic signal amplitude, infrared temperature gradient and defect area. Specifically, if the ultrasound signal amplitude is less than 0.3, it is rated as mild; if the ultrasound signal amplitude is greater than or equal to 0.3 and less than 0.6, it is rated as moderate; if the ultrasound signal amplitude is greater than or equal to 0.6, it is rated as severe. If the infrared temperature gradient is less than 5℃, it is rated as mild; if the infrared temperature gradient is greater than or equal to 5℃ and less than 10℃, it is rated as moderate; if the infrared temperature gradient is greater than or equal to 10℃, it is rated as severe. If the defect area is less than 5cm², it is rated as mild; if the defect area is greater than or equal to 5cm² and less than 15cm², it is rated as moderate; if the defect area is greater than or equal to 15cm², it is rated as severe.

[0068] Test results: In the test of the third transformer, the probability value of the "corrosion" defect category output by the defect detection model was 0.96, the location coordinates were (320,450)-(480,620), the ultrasonic signal amplitude was 0.42, the infrared temperature gradient was 7.3℃, and the defect area was 8.6cm², which was rated as moderate corrosion; the test results of the other 9 transformers were no defects.

[0069] (6) Result Output Module: Output results: The visualization display shows the corrosion defect type of the third transformer in real time, the location map of the defect area marked with a red rectangle, and the defect detection report including defect parameters and evaluation criteria. At the same time, the audible and visual alarm device is activated, and the alarm information and defect report are sent to the mobile terminal of the substation operation and maintenance personnel through the wireless communication module.

[0070] (7) Data storage module: Data storage: The data storage module stores real-time detection data, preprocessed data, feature data, detection result reports, and historical training data, which facilitates subsequent tracking of defect development trends and model optimization.

[0071] Example 2: Specific configuration and operation of each module of the power equipment metal component defect detection system.

[0072] (1) Data acquisition module: A combination of mobile and fixed data acquisition methods was adopted: ① A drone equipped with a high-definition zoom camera flew along the power transmission line to collect image data of the connecting bolts; ② A portable ultrasonic testing instrument was used by maintenance personnel on-site to collect ultrasonic testing data by placing the probe of the portable ultrasonic testing instrument against the surface of the bolts; ③ A handheld infrared thermal imager was used by maintenance personnel to simultaneously collect infrared thermal imaging data of the bolts and connections.

[0073] Historical testing data collection: Collect testing data from 50 disconnect switches over the past 8 years, including 5,000 sets of image data, 5,000 sets of ultrasonic testing data, and 5,000 sets of infrared thermal imaging data, covering various scenarios such as normal bolt condition, cracks, deformation, and corrosion.

[0074] Data collection for testing: A full-line test is conducted once per quarter. The drone flies at a speed of 5 m / s. 20 sets of image data are collected for each set of disconnect switches. At the same time, maintenance personnel collect 20 sets of ultrasonic test data and 20 sets of infrared thermal imaging data.

[0075] (2) Data preprocessing module: Data cleaning: 410 outliers in historical test data were removed using the 3σ principle; noise in image data and infrared thermal imaging data was removed using 3×3 template median filtering; and 130 missing values ​​in ultrasound test data were filled using linear interpolation.

[0076] Data standardization: The pixel values ​​of the cleaned image data, infrared thermal imaging data, and the amplitude of the ultrasonic detection data are uniformly normalized to the [0,1] interval to eliminate the dimensional differences of data collected by different devices.

[0077] Data augmentation: The normalized image data was randomly rotated from 0 to 180 degrees, horizontally flipped, vertically flipped, and its brightness was adjusted by ±20%, with random cropping to 150×150 pixels. Random Gaussian noise with a mean of 0 and a variance of 0.008 and signal shifting of ±3 sampling points were added to the normalized ultrasound detection data. The contrast of the normalized infrared thermal imaging data was adjusted by ±15%, and the temperature threshold was stretched from 5℃ to 40℃. After data augmentation, the total amount of historical detection data was expanded to 15,000 sets.

[0078] Data partitioning: The 15,000 sets of historical detection data obtained after data augmentation were divided into training set, validation set and test set in a ratio of 7:2:1. The training set contains 10,500 sets of historical detection data, the validation set contains 3,000 sets of historical detection data, and the test set contains 1,500 sets of historical detection data.

[0079] (3) Multi-level feature extraction module: The first-level feature extraction unit consists of three convolutional layers with a kernel size of 3×3, a stride of 1, and the same padding method, forming a shallow convolutional layer. Each convolutional layer in the shallow convolutional layer is followed by a batch normalization layer and a ReLU activation function. The shallow convolutional layer is followed by a max pooling layer with a kernel size of 2×2 and a stride of 2. The first-level feature extraction unit is used to extract first-level features, including the bolt contour features, surface thread texture features, basic grayscale features, and preliminary amplitude features of ultrasonic signals.

[0080] The second-level feature extraction unit employs a deep convolutional layer with 4 groups, a kernel size of 5×5, a stride of 1, and a padding method of "same". After the deep convolutional layer, a batch normalization layer and a ReLU activation function are connected. Subsequently, two dilated convolutional layers with kernel sizes of 3×3 and dilatations of 2 and 3 are connected. The second-level feature extraction unit extracts second-level features including subtle changes in bolt head cracks, grayscale differences caused by thread corrosion, ultrasonic signal frequency features, and temperature gradient features between the bolt and the connector.

[0081] The third-level feature extraction involves a squeezing operation using global average pooling, followed by an activation operation using a 200-dimensional fully connected layer with ReLU activation, a 50-dimensional fully connected layer with sigmoid activation, and channel weights. These weights are then multiplied channel-by-channel with the second-level features. The multiplication result is then input into the global average pooling layer to obtain the 200-dimensional deep features output by the global average pooling layer.

[0082] (4) Model training module: Model Structure: An improved convolutional neural network and Transformer fusion model is constructed as the defect detection model, consisting of an input layer, a feature fusion layer, a fully connected layer, a dropout layer, and an output layer. The input layer receives 200-dimensional deep features; the feature fusion layer adopts a multi-head self-attention mechanism with 4 attention heads and a feedforward neural network with a hidden layer dimension of 800; the fully connected layer contains two layers with output dimensions of 512 and 128 respectively, each followed by a batch normalization layer and a ReLU activation function; the dropout layer sets the dropout probability to 0.5; the output layer outputs 5 dimensions, corresponding to the 5 defect categories of cracks, corrosion, porosity, deformation, and no defects.

[0083] Training process: The Xavier normal initialization method was used to initialize the model parameters; the initial learning rate was set to 0.001, and a cosine annealing strategy was used to decay the learning rate with a weight decay coefficient of 0.0001. The batch size was set to 32, and the number of iterations was set to 100 rounds. In the 78th round, the early stopping strategy was triggered after 15 consecutive rounds of verification that the loss had not decreased. The defect detection model was tested using a test set. The performance test results were as follows: detection accuracy was 98.3%, recall was 97.6%, F1 score was 97.9%, and false negative rate was 0.9%. These performance test results met the preset thresholds in the preset performance requirements, and the trained defect detection model was obtained.

[0084] (5) Defect detection module: Preprocessing of the data to be tested: 3σ outlier removal, median filtering, linear interpolation filling, and normalization to the [0,1] interval are performed on the 20 sets of data to be tested for each disconnector switch.

[0085] Feature Extraction and Detection: After preprocessing, the data to be detected undergoes multi-level feature extraction and is input into the defect detection model. The model outputs the defect category and its probability value. The defect location is determined based on image coordinates, and the severity is comprehensively assessed by combining ultrasonic signal amplitude, infrared temperature gradient, and defect area. Specifically, if the ultrasonic signal amplitude is less than 0.25, it is rated as mild; if the ultrasonic signal amplitude is greater than or equal to 0.25 and less than 0.5, it is rated as moderate; if the ultrasonic signal amplitude is greater than or equal to 0.5, it is rated as severe. If the infrared temperature gradient is less than 4℃, it is rated as mild; if the infrared temperature gradient is greater than or equal to 4℃ and less than 8℃, it is rated as moderate; if the infrared temperature gradient is greater than or equal to 8℃, it is rated as severe. If the defect area is less than 3cm², it is rated as mild; if the defect area is greater than or equal to 3cm² and less than 10cm², it is rated as moderate; if the defect area is greater than or equal to 10cm², it is rated as severe.

[0086] Test results: In the 12th and 37th groups of disconnecting switches, "crack" defects (probability value 0.95, location coordinates (85,120)-(110,150), ultrasonic amplitude 0.58, temperature gradient 8.2℃, defect area 10.3cm², rated as severe) and "deformation" defects (probability value 0.93, location coordinates (210,300)-(240,330), ultrasonic amplitude 0.32, temperature gradient 5.1℃, defect area 6.7cm², rated as moderate) were detected, respectively; the remaining 48 groups showed no defects.

[0087] (6) Result Output Module: Output results: The visualization display shows the type, location markings, and inspection reports of the two defects, namely "crack" and "deformation". Severe crack defects trigger an audible and visual alarm, while moderate deformation defects trigger a light alarm. Alarm information and reports are sent to the mobile terminal of the line maintenance team via wireless communication.

[0088] (7) Data storage module: Data storage: Stores various types of data, including inspection data, feature data, and result reports, providing data support for subsequent bolt defect trend analysis and model optimization.

[0089] Those skilled in the art will understand that the above embodiments are specific implementations of the present invention, and in practical applications, various changes can be made in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A defect detection system for metal components of power equipment, characterized in that, include: The data acquisition module is used to collect historical test data and data to be tested of metal components of power equipment. Both historical test data and data to be tested include data of multiple data types. The data preprocessing module is used to preprocess historical detection data to obtain standardized preprocessed data, divide the standardized preprocessed data into training set, validation set and test set, and preprocess the data to be detected to obtain standardized data to be detected. The multi-level feature extraction module is used to extract the deep features corresponding to the standardized preprocessed data and the standardized data to be detected in a vertically progressive manner through a sequential feature extraction architecture. The model training module is used to construct an improved deep learning model as a defect detection model based on the deep features corresponding to the standardized preprocessed data, and to train the defect detection model based on the training set, validation set and test set, optimize the model parameters, and obtain the trained defect detection model. The defect detection module is used to input the deep features corresponding to the standardized data to be detected into the trained defect detection model to obtain the defect detection results of the metal components. The result output module is used to receive the defect detection results of metal components, visualize the defect detection results of metal components and provide abnormal alarms, so as to remind staff to deal with defects in a timely manner. The data storage module is used to store various types of data during the defect detection process of metal components in power equipment, so as to facilitate subsequent data traceability, model optimization and defect analysis.

2. The power equipment metal component defect detection system according to claim 1, characterized in that, Data of multiple types includes image data, ultrasound detection data, and infrared thermal imaging data.

3. The power equipment metal component defect detection system according to claim 2, characterized in that, The data preprocessing module is used to preprocess historical detection data to obtain standardized preprocessed data, which is then divided into training, validation, and test sets, including: The data preprocessing module is specifically used for: Data cleaning: Outliers in historical detection data are removed using the 3σ principle, noise in image data and infrared thermal imaging data in historical detection data is removed using the median filtering algorithm, and missing values ​​in ultrasonic detection data in historical detection data are filled using linear interpolation, resulting in cleaned historical detection data; Data standardization: Normalize the cleaned historical test data; Data augmentation: Perform data augmentation on the normalized historical detection data to improve the diversity of training samples; Data partitioning: The historical detection data after data augmentation is divided into training set, validation set and test set according to a preset ratio. The training set is used for model training, the validation set is used for model parameter adjustment, and the test set is used for model performance verification.

4. The power equipment metal component defect detection system according to claim 2, characterized in that, The multi-level feature extraction module includes a first-level feature extraction unit, a second-level feature extraction unit, and a third-level feature extraction unit; The first-level feature extraction unit includes shallow convolutional layers and pooling layers of the convolutional neural network, which are used to perform shallow feature extraction on the standardized preprocessed data and the standardized data to be detected, respectively, to obtain the corresponding first-level features. The first-level features include the contour features, surface texture features, basic grayscale features and preliminary signal amplitude features of the metal components. The second-level feature extraction unit includes a deep convolutional layer and a dilated convolutional layer, which are used to extract mid-level features based on the first-level features to obtain the second-level features. The second-level features include subtle changes in the outline of the metal components, grayscale differences in local areas, frequency characteristics of ultrasonic signals, and temperature gradient characteristics of infrared thermal imaging. The second-level features are a further refinement and deepening of the first-level features. The third-level feature extraction unit includes an attention mechanism fusion layer and a global pooling layer, which are used to extract deep features based on the second-level features and obtain deep features. The attention mechanism fusion layer is used to assign weights to different channels of the second-level features using a channel attention mechanism, strengthening the weights of defect-related features and suppressing interference from irrelevant features. The obtained channel weights are multiplied with the second-level features channel by channel to obtain the attention-enhanced features. The global pooling layer is used to perform global average pooling on the attention-enhanced features to obtain fixed-dimensional deep features. The channel attention mechanism includes squeezing and activation operations. The squeezing operation is implemented through global average pooling, and the activation operation is implemented through a fully connected operation, ReLU activation, fully connected operation, and sigmoid activation in sequence.

5. The power equipment metal component defect detection system according to claim 1, characterized in that, The defect detection model is an improved convolutional neural network and Transformer fusion model, which includes an input layer, a feature fusion layer, a fully connected layer, a dropout layer, and an output layer; wherein, The input layer is used to receive the deep features output by the multi-level feature extraction module; The feature fusion layer adopts the encoder structure of Transformer to perform global feature association learning on deep features and output the fused global deep features. The encoder structure includes a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism is used to capture the global dependencies between deep features, and the feedforward neural network is used to perform nonlinear transformation on the features output by the multi-head self-attention mechanism. The fully connected layer consists of multiple fully connected layers with different output dimensions, and each fully connected layer is followed by a batch normalization layer and a ReLU activation function; The dropout layer is placed after the second fully connected layer to prevent the model from overfitting. The output layer uses the softmax activation function to output the probability values ​​of various defects in the metal components of power equipment.

6. The power equipment metal component defect detection system according to claim 4, characterized in that, The model training module is specifically used for: Initialize the parameters of the defect detection model using Xavier normal initialization. Input the deep features corresponding to the training set into the defect detection model to obtain the corresponding prediction output; The cross-entropy loss function is used to calculate the loss value between the predicted output and the corresponding label in the training set; The Adam optimizer is used to iteratively update the parameters of the defect detection model. The learning rate of the Adam optimizer is decayed using a cosine annealing strategy. In each iteration, a preset number of rounds are used to verify the performance of the defect detection model using a validation set. The validation accuracy and validation loss of the defect detection model are calculated. If the validation loss does not decrease in a specified number of consecutive rounds, an early stopping strategy is used to stop training in order to avoid model overfitting. After training stops, the defect detection model is tested using a test set to determine its performance. If the model performance meets the preset performance requirements, the trained defect detection model is obtained. If the model performance does not meet the preset performance requirements, the model parameters are adjusted and the defect detection model is retrained until the model performance meets the preset performance requirements.

7. The power equipment metal component defect detection system according to claim 1, characterized in that, The defect detection module is specifically used for: The deep features corresponding to the standardized data to be detected are input into the trained defect detection model to obtain the probability values ​​of various defects of the metal components of power equipment output by the model. The defect category with the highest probability value is determined as the defect type of the metal component. At the same time, the location information and severity of the defect are output as the defect detection result of the metal component. If the defect detection result of the metal component is that there are no defects, a normal inspection report is generated and output to the result output module. If the defect detection results of the metal component indicate the presence of a defect, the defect type, location information, and severity of the defect are marked, and a defect detection report and abnormal signal are generated and output to the result output module.

8. The power equipment metal component defect detection system according to claim 1, characterized in that, The result output module includes: A visual display is used to show the defect detection results, corresponding defect location annotation maps, defect detection reports, and historical detection data query results in real time. An alarm device is used to automatically trigger an alarm when the defect results of a metal component indicate the presence of a defect of a preset severity, so as to remind staff to handle the situation in a timely manner.

9. The power equipment metal component defect detection system according to claim 8, characterized in that, The alarm device uses both audible and visual alarms.

10. The power equipment metal component defect detection system according to claim 8, characterized in that, The alarm device is also used to send alarm information to staff's mobile terminals via a wireless communication module.

Images