A power system defect identification method

Abstract

The application discloses a kind of power system defect identification methods, specific steps are as follows: S1, the real-time image data and prompt image data collected from the key parts of electric power equipment are respectively input into backbone network, and real-time image feature and normal image feature are extracted;S2, real-time image feature is input into image feature fusion module, and fusion feature is obtained;S3, prompt image feature is input into visual prompt feature extraction module, and target feature is obtained;S4, fusion feature and target feature are input into prompt feature fusion module and are deeply fused, and depth fusion feature is obtained;S5, depth fusion feature is input into detection head, and the position and category information of defect target are obtained.Once abnormal or potential defect is detected, the system will automatically trigger an alarm and notify the maintenance personnel to check and handle.The present scheme is suitable for power system or other industrial control scenarios.

Description

Technical Field

[0001] This invention relates to the field of large-scale models, and in particular to a method for identifying defects in power systems. Background Technology

[0002] As power systems become increasingly large and complex, traditional defect detection methods are struggling to meet the demands of modern power equipment maintenance. Traditional methods typically rely on manual inspection or simple sensors, which are not only time-consuming and labor-intensive but also susceptible to human error, leading to inaccurate results. With the rise of AI, more and more deep learning-based methods are being applied to power system defect detection tasks. However, current applications in power system defect detection often involve small, specialized models targeting specific defects. Each defect requires a separate model to be trained, which is time-consuming, labor-intensive, and necessitates the collection of dedicated data for training each time.

[0003] Currently, some researchers have begun to use multimodal large language models to identify power system defects, which to some extent achieves the goal of solving multiple problems with one model. However, some defects in power systems (such as missing bolts, damaged insulation, etc.) are hidden and complex, making them difficult to accurately describe and identify through language. In such cases, using language modal prompts cannot accurately describe and identify the defects. Summary of the Invention

[0004] This invention mainly addresses the technical problems of low versatility and poor recognition rate of existing defect identification models, and provides a power system defect identification method that can accurately complete the identification without precise verbal description.

[0005] The present invention addresses the aforementioned technical problems primarily through the following technical solution: a power system defect identification method, comprising the following steps: S1. Real-time image data collected from key parts of the power equipment is input into the backbone network to extract real-time image features; prompt image data is input into the backbone network to obtain prompt image features; the prompt image is an image containing defects, preferably with the defective part extracted, other parts deleted or covered, and the location information of the defective part is marked; real-time image features and prompt image features are collectively referred to as image features; the backbone network can be a commonly used network such as convnext or vit, and the extracted image features are the features output by the last three layers of the backbone network; S2. Input the real-time image features into the image feature fusion module to obtain the fused features; the image feature fusion module fuses different image features, and can be a commonly used fusion module such as FPN or PAN. S3. Input the features of the prompt image into the visual prompt feature extraction module to obtain the target features; S4. Input the fused features and target features into the prompt feature fusion module for deep fusion to obtain deep fused features; S5. Input the deep fusion features into the detection head to obtain the location and category information of the defect target.

[0006] Once an anomaly or potential defect is detected, the system will automatically trigger an alarm to notify maintenance personnel to inspect and handle the issue.

[0007] Preferably, in step S3, the specific process of obtaining the target features through the visual cue feature extraction module is as follows: S301. The image features include three layers of depth features. Each of the three depth features is adjusted by passing it through a CBR block. Each CBR block contains a 1*1 convolutional network, a batch normalization layer, and a ReLU layer connected in sequence. The depth features are adjusted by the CBR blocks to further enhance the expressive power of the features. S302. The adjusted second and third layer depth features are amplified through an upsampling layer to match the scale of the adjusted first layer depth features. S303. The three layers of consistent depth features are concatenated according to the channel direction using a concat layer. S304. The features obtained by concatenation in step S303 are passed through a CBR block to obtain feature_D. The CBR block here is also composed of a 1*1 convolutional network, a batch normalization layer and a ReLU layer connected in sequence, but the number of channels in each CBR block is different. S305. The obtained feature_D is a feature of B*C*H*W dimensions, where B (batch) represents the batch size, C (channel) represents the number of channels, and H and W are the height and width of the feature map, respectively. Feature_D enters two branches: the first branch reshapes feature_D to obtain a vector of B*(HW)*C dimensions, and then passes it through three one-layer transformer encoding layers to obtain a vector of B*(HW)*C dimensions, feature_E1; at the same time, the second branch passes feature_D through a global pooling layer to obtain a vector of B*1*C dimensions, feature_E2. S306. Add feature_E1 and feature_E2 along dimension C to obtain a vector of B*(HW)*C, and then reshape the vector to obtain a vector of dimension B*C*H*W. S307. Pass the vector feature_E through a CBR block to obtain the vector feature_F; the CBR block here has the same structure as the previous CBR block, only the number of channels is different; S308. Using a region alignment (roialign) layer, the target features to be detected are extracted from the vector feature_F. The target feature dimensions are B*C*h*w. The function of roialign is to extract a small feature map corresponding to the target's position in the global feature map based on the target's position information (label coordinates) in the image. During training, the position of the defect target in the image is known, where h and w are the height and width of the defect target, respectively. If the image only contains defects, the extracted target features are the same as the original vector feature_F.

[0008] Preferably, in step S4, the deep fusion process of the feature fusion module is as follows: S401. First, the fusion features with dimensions B*C*H*W extracted by the image feature fusion module and the target features with dimensions B*C*h*w extracted by the visual cue feature extraction module are concatenated in the H*W dimension to obtain a vector of (H*W+h*w)*B*C. S402. The concatenated features pass through a transformer layer to obtain fused features with dimensions (H*W+h*w)*B*C. Then, the dimensions are adjusted (reshaped) to obtain features with dimensions B*C*2*(H*W+h*w). S403. The features of dimension B*C*2*(H*W+h*w) are passed through a fully connected layer to reduce the dimension to B*C*H*W, which is the feature after deep fusion.

[0009] Preferably, the detection head is a YOLO series detection head.

[0010] Preferably, during model training, the real-time image data and the prompt image data input to the backbone network are the same data. The training data is labeled using manual or automated tools to identify and mark potential defect features for use in model training, optimizing model parameters to improve its recognition accuracy and robustness.

[0011] The substantial effect of this invention is that it allows the use of actual visual images of defects as cues to accurately guide large models to identify images similar to the cues. Visual cues can be used to simply define and describe power defects, making it applicable to various power defects without requiring separate training for each specific defect. The cue feature extraction and cue feature fusion modules are plug-and-play, enabling rapid transformation of traditional networks into network structures suitable for visual cues. Attached Figure Description

[0012] Figure 1 This is a flowchart of a circuit system defect identification method according to the present invention. Detailed Implementation

[0013] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0014] Example: This example illustrates a power system defect identification method, such as... Figure 1 As shown, it includes the following steps: S1. Real-time image data collected from key parts of the power equipment is input into the backbone network to extract real-time image features; prompt image data is input into the backbone network to obtain prompt image features; the prompt image is an image containing defects, preferably with the defective part extracted, other parts deleted or covered, and the location information of the defective part is marked; real-time image features and prompt image features are collectively referred to as image features; the backbone network can be a commonly used network such as convnext or vit, and the extracted image features are the features output by the last three layers of the backbone network; S2. Input the real-time image features into the image feature fusion module to obtain the fused features; the image feature fusion module fuses different image features, and can be a commonly used fusion module such as FPN or PAN. S3. Input the features of the prompt image into the visual prompt feature extraction module to obtain the target features; S4. Input the fused features and target features into the prompt feature fusion module for deep fusion to obtain deep fused features; S5. Input the deep fusion features into the detection head to obtain the location and category information of the defect target.

[0015] Once an anomaly or potential defect is detected, the system will automatically trigger an alarm to notify maintenance personnel to inspect and handle the issue.

[0016] In step S3, the specific process of obtaining the target features through the visual cue feature extraction module is as follows: S301. The image features include three layers of depth features. Each of the three depth features is adjusted by passing it through a CBR block. Each CBR block contains a 1*1 convolutional network, a batch normalization layer, and a ReLU layer connected in sequence. The depth features are adjusted by the CBR blocks to further enhance the expressive power of the features. S302. The adjusted second and third layer depth features are amplified through an upsampling layer to match the scale of the adjusted first layer depth features. S303. The three layers of consistent depth features are concatenated according to the channel direction using a concat layer. S304. The features obtained by concatenation in step S303 are passed through a CBR block to obtain feature_D. The CBR block here is also composed of a 1*1 convolutional network, a batch normalization layer and a ReLU layer connected in sequence, but the number of channels in each CBR block is different. S305. The obtained feature_D is a feature of B*C*H*W dimensions, where B (batch) represents the batch size, C (channel) represents the number of channels, and H and W are the height and width of the feature map, respectively. Feature_D enters two branches: the first branch reshapes feature_D to obtain a vector of B*(HW)*C dimensions, and then passes it through three one-layer transformer encoding layers to obtain a vector of B*(HW)*C dimensions, feature_E1; at the same time, the second branch passes feature_D through a global pooling layer to obtain a vector of B*1*C dimensions, feature_E2. S306. Add feature_E1 and feature_E2 along dimension C to obtain a vector of B*(HW)*C, and then reshape the vector to obtain a vector of dimension B*C*H*W. S307. Pass the vector feature_E through a CBR block to obtain the vector feature_F; the CBR block here has the same structure as the previous CBR block, only the number of channels is different; S308. Through the region alignment (roialign) layer, the target features to be detected are extracted from the vector feature_F. The target feature dimensions are B*C*h*w. The function of roialign is to extract the small feature map corresponding to the target's position in the global feature map based on the target's position information (label coordinates) in the image. During the training process, the position of the defective target in the image is known, and h and w are the height and width of the defective target, respectively.

[0017] In step S4, the deep fusion process of the feature fusion module is as follows: S401. First, the fusion features with dimensions B*C*H*W extracted by the image feature fusion module and the target features with dimensions B*C*h*w extracted by the visual cue feature extraction module are concatenated in the H*W dimension to obtain a vector of (H*W+h*w)*B*C. S402. The concatenated features pass through a transformer layer to obtain fused features with dimensions (H*W+h*w)*B*C. Then, the dimensions are adjusted (reshaped) to obtain features with dimensions B*C*2*(H*W+h*w). S403. The features of dimension B*C*2*(H*W+h*w) are passed through a fully connected layer to reduce the dimension to B*C*H*W, which is the feature after deep fusion.

[0018] The detection head is a YOLO series detection head.

[0019] During model training, the real-time image data and prompt image data input to the backbone network are the same. The training data is labeled using manual or automated tools to identify and mark potential defect features for use in model training, optimizing model parameters to improve its recognition accuracy and robustness.

[0020] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0021] Although this paper uses terms such as backbone network, visual cue feature extraction module, and cue feature fusion module frequently, the possibility of using other terms is not excluded. These terms are used merely for the convenience of describing and explaining the essence of this invention; interpreting them as any additional limitation would contradict the spirit of this invention.

Claims

1. A method for identifying defects in a power system, characterized in that, Includes the following steps: S1. Input the real-time image data collected from key parts of the power equipment into the backbone network and extract the real-time image features; The prompt image data is input into the backbone network to obtain the prompt image features. The prompt image is an image containing defects. S2. Input the real-time image features into the image feature fusion module to obtain the fused features; S3. Input the features of the prompt image into the visual prompt feature extraction module to obtain the target features; S4. Input the fused features and target features into the prompt feature fusion module for deep fusion to obtain deep fused features; S5. Input the deep fusion features into the detection head to obtain the location and category information of the defect target; In step S3, the specific process of obtaining the target features through the visual cue feature extraction module is as follows: S301. The image features include three layers of depth features. The three layers of depth features are adjusted by passing them through a CBR block. Each CBR block contains a 1*1 convolutional network, a batch normalization layer, and a ReLU layer connected in sequence. S302. The adjusted second and third layer depth features are amplified through an upsampling layer to match the scale of the adjusted first layer depth features. S303. The three layers of consistent depth features are stitched together according to the channel direction through a stitching layer; S304. The features obtained by splicing in step S303 are processed through a CBR block to obtain feature_D; S305. The obtained feature_D is a feature of dimension B*C*H*W, where B represents the batch size, C represents the number of channels, and H and W are the height and width of the feature map, respectively. Feature_D enters two branches: the first branch adjusts the dimensions of feature_D to obtain a vector of dimension B*(HW)*C, and then passes it through three one-layer transformer encoding layers to obtain a vector of dimension B*(HW)*C, feature_E1; at the same time, the second branch passes feature_D through a global pooling layer to obtain a vector of dimension B*1*C, feature_E2. S306. Add feature_E1 and feature_E2 along dimension C to obtain a vector of B*(HW)*C, and then adjust the dimensions to obtain a vector of B*C*H*W dimensions, feature_E. S307. Pass the vector feature_E through a CBR block to obtain the vector feature_F; S308. Through the region alignment layer, extract the target features to be detected from the vector feature_F. The dimensions of the target features are B*C*h*w, where h and w are the height and width of the defect, respectively.

2. The power system defect identification method according to claim 1, characterized in that, In step S4, the deep fusion process of the feature fusion module is as follows: S401. First, the fusion features with dimensions B*C*H*W extracted by the image feature fusion module and the target features with dimensions B*C*h*w extracted by the visual cue feature extraction module are concatenated in the H*W dimension to obtain a vector of (H*W+h*w)*B*C. S402. The concatenated features pass through the transformer layer to obtain the fused features with dimensions (H*W+h*w)*B*C. Then, the dimensions are adjusted to obtain features with dimensions B*C*2*(H*W+h*w). S403. The features of dimension B*C*2*(H*W+h*w) are passed through a fully connected layer to reduce the dimension to B*C*H*W, which is the feature after deep fusion.

3. The power system defect identification method according to claim 1, characterized in that, The detection head is a YOLO series detection head.

4. The power system defect identification method according to claim 1, characterized in that, During model training, the real-time image data and the prompt image data input to the backbone network are the same data.

Images