Power equipment fault identification method based on improved YOLO model

Abstract

The invention provides a power equipment fault identification method based on an improved YOLO model. The method comprises the following steps: acquiring and preprocessing image data of power equipment; performing multiple times of feature extraction on the image data to obtain first image features under different scales; the first image features comprise edge features, color features and texture features; setting a target function, and performing combined analysis on the edge features, the color features and the texture features by taking the target function as a guide to obtain M combined features; probability prediction is carried out on the M combined features according to each anchor point of each scale, and N effective combined features are obtained; wherein the effective combination feature comprises an equipment type, a fault type and a probability value; and classifying power equipment faults according to the effective combination features to realize fault identification of the power equipment. According to the electrical equipment fault identification method based on the improved YOLO model, the high recall rate is guaranteed, and meanwhile the accuracy is kept to be 95% or above.

Description

Technical Field

[0001] This invention relates to the field of fault identification, and in particular to a method for identifying power equipment faults based on an improved YOLO model. Background Technology

[0002] Power equipment faults are often minor defects (such as insulator cracks, loose bolts, and cable corrosion). Traditional YOLO models suffer from a loss of shallow details due to cross-layer connections, resulting in a high false negative rate for small targets smaller than 32px. In power scenarios, the ratio of normal equipment to faulty samples is vastly different; directly applying cross-entropy loss biases the model towards the majority class (normal state). Fixed-size default anchor frames struggle to match the size differences of equipment at different voltage levels (e.g., the volume difference between 110kV and 500kV transformers is as much as 8 times). Substation environments contain numerous similar textures (conductors, fittings) and lighting variations (shadows, reflections), making traditional single-modal features prone to generating false alarms. UAV inspections require processing >24 frames per second, but while the classic two-stage detector (Faster R-CNN) offers high accuracy, it cannot meet the speed requirements. The physical characteristics of power equipment (such as the temperature gradient distribution of overheating faults) and topological relationships (electrical connection constraints between equipment) are not utilized. Detection results lack interpretability, cannot provide maintenance suggestions, and require inefficient manual secondary verification. Summary of the Invention

[0003] The purpose of this invention is to provide a power equipment fault identification method based on an improved YOLO model. Through four core technological innovations—a dual-stream feature collaboration mechanism, an adaptive anchor frame system, a dynamic sample allocation strategy, and knowledge-guided feature combination—this method systematically solves key challenges in power equipment fault detection, such as missed detections of small targets, false alarms in complex backgrounds, variable equipment morphologies, and sample imbalance. While maintaining real-time performance, it achieves significant improvements in detection accuracy, robustness, and adaptability, making it particularly suitable for intelligent inspection needs in complex scenarios such as transmission lines and substations.

[0004] To achieve the above objectives, this invention provides a power equipment fault identification method based on an improved YOLO model, comprising the following steps: Acquire and preprocess image data of power equipment; Multiple feature extractions are performed on the image data to obtain first image features at different scales; the first image features include edge features, color features, and texture features; Set an objective function, and use the objective function as a guide to perform combined analysis on edge features, color features and texture features to obtain M combined features; M combined features are probabilistically predicted at each anchor point of each scale to obtain N effective combined features; among them, the effective combined features include equipment type, fault type and probability value; Power equipment faults are classified according to effective combination characteristics to achieve fault identification of power equipment.

[0005] Preferably, the improved backbone network performs multiple feature extractions on the image data to obtain first image features at different scales, including the following steps: The image data is processed by the first branch, and then obtained by splicing double pooling, double-layer MLP compression and expansion, and adaptive weight adjustment to obtain second image features with higher correlation to the fault point. The image data is processed by the second branch, and after being processed by the standard convolutional block, it is processed by dimensionality reduction, spatial feature fusion and spatial fine-tuning to obtain the third image features with multi-scale receptive fields. The second image features optimized by the channel dimension and the third image features optimized by the spatial dimension are combined through a cascade operation to obtain the enhanced backbone network feature output.

[0006] Preferably, the objective function is: ; In the formula, The bounding box regression loss is used to guide the model to learn the precise location and size of electrical equipment, and the generalized intersection-union loss function is employed. The classification loss is used to guide the model to learn the correct fault type classification, and the weighted cross-entropy loss function is employed. This represents the confidence loss, used to guide the model in determining whether there are electrical devices within the prediction box. It employs the binary cross-entropy loss function. These are the weighting coefficients for each loss term, used to balance the contribution weights of different loss terms.

[0007] Preferably, the feature fusion network performs weighted fusion of multi-scale features to obtain M combined features, including the following steps: Regions are divided according to edge features, and feature labels are set for edge features, color features, and texture features based on the region division results; Overlap analysis is performed on edge features, color features, and texture features based on feature labels. Edge features, color features, and texture features at different scales are superimposed and combined to obtain M combined features.

[0008] Preferably, the M combined features are probabilistically predicted according to each anchor point of each scale to obtain N effective combined features, including the following steps: Adaptive anchor boxes are generated by clustering based on the device size distribution. The detection head predicts the bounding box offset, confidence, and class probability for each anchor point on the M combined features, generating M anchor boxes. By using confidence thresholding and non-maximum suppression, most overlapping and unreliable anchor boxes are filtered out, resulting in N effective combined features.

[0009] Preferably, generating adaptive anchor boxes by clustering based on device size distribution includes: Extract the width and height of all real annotation boxes, normalize and filter the width and height of the real annotation boxes, and then use K-class cluster analysis to obtain N sets of basic anchor box dimensions. Define a learnable scaling factor α∈[0.8,1.2] to act on the base anchor frame; Positive and negative samples are dynamically divided based on the Intersection over Union (IoU). Each iteration calculates the average IoU between the current anchor box and the ground plane; If the average IoU is lower than the target value, then expand the range of α; otherwise, narrow the range. Use a smooth function to transition threshold: f=f 0 +β (1) -exp ( -t / β 0)); In the formula, f Indicates the threshold. f 0 represents the base threshold. β This indicates the attenuation level, controlling the range of increase in the threshold. t Indicates the current training round. β 0 represents the decay rate, which controls the speed at which the threshold increases; When a new device type is detected, initiate rapid re-clustering; The anchor frame is automatically scaled based on the camera resolution.

[0010] A power equipment fault identification system based on an improved YOLO model, comprising: The feature extraction and preliminary enhancement module is used to suppress background noise, and the features of power equipment and its potential fault areas are significantly enhanced. The multi-task attention module enables the model to flexibly switch its focus across multiple object detection tasks by perceiving different tasks. Multi-scale feature fusion is used to fuse features from different levels, so that each layer of features has both strong semantics and high resolution. The detection head generates a large number of overlapping prediction boxes, and finally retains only the most reliable prediction box to output the recognition result.

[0011] Therefore, the present invention employs the above-mentioned power equipment fault identification method based on an improved YOLO model, and the technical effects are as follows: A dual-stream feature collaboration mechanism is proposed, employing a parallel design scheme of channel attention and spatial attention. The first branch utilizes an MLP-based gating mechanism to adaptively weight the channel dimensions of the feature map, highlighting feature channels relevant to power equipment faults while suppressing background information channels, thus achieving channel-level feature selection. The second branch establishes the correlation between different spatial locations in the feature map through multi-scale receptive field convolutional kernels and a spatial attention module, strengthening the spatial proximity and geometric constraints between power equipment components (such as insulators and bushings). The feature outputs of the two branches are cascaded and fused to achieve complementary enhancement of the channel and spatial dimensions.

[0012] An adaptive anchor frame evolution system: Addressing the adaptation problem of diverse power equipment forms (such as insulator strings, transformer bushings, circuit breakers, etc.), a dynamically learnable anchor frame mechanism is proposed. Based on the basic anchor frame size obtained through K-means offline clustering, a learnable scaling factor is introduced, enabling the anchor frames to adaptively adjust their size during training. Simultaneously, a curriculum learning strategy is employed to dynamically adjust the IoU matching threshold for positive samples: a lower IoU threshold is set in the early stages of training, making it easier for the model to obtain positive sample supervision signals; as training progresses, the IoU threshold is gradually increased, only labeling high-quality anchor frame-GT matches as positive samples, forcing the network to learn more accurate target localization. This avoids convergence difficulties caused by insufficient samples in the early stages of training, while improving localization accuracy in the later stages, preventing the model from overfitting to low-quality matches. Multimodal feature combination mathematical modeling encodes edge gradient histograms, HSV color statistics, and Gabor texture responses into a unified feature vector, and automatically learns the optimal combination weights through an attention mechanism. Attached Figure Description

[0013] Figure 1 This is a flowchart of a power equipment fault identification method based on an improved YOLO model according to the present invention. Detailed Implementation

[0014] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0016] Example 1 like Figure 1 As shown, a fault identification method for power equipment based on an improved YOLO model includes the following steps: Acquire and preprocess image data of power equipment; A power grid company used a drone inspection system equipped with a high-definition camera to collect 100,000 visible light images of transformers, insulators, surge arresters, and other equipment at a 500kV substation. The images underwent the following preprocessing operations: Cropping: Removes background elements (such as sky and ground) that are irrelevant to the image edges. Normalization: Scales pixel values ​​to the [0,1] range; Noise reduction: Apply Gaussian filtering to eliminate sensor noise; Enhancement: Histogram equalization is performed on images in low-light areas.

[0017] Performing multiple feature extractions on image data to obtain first image features at different scales includes the following steps: The image data is processed by the first branch, and then obtained by stitching double pooling, double-layer MLP (perceptron) compression and expansion, and adaptive weight adjustment to obtain second image features with higher correlation to the fault point. The image data is processed by the second branch, and after being processed by the standard convolutional block, it is further processed by dimensionality reduction (1×1 convolution), spatial feature fusion (convolutional kernels with multiple receptive fields in parallel), and spatial fine-tuning (spatial attention mechanism) to obtain the third image features with multi-scale receptive fields. The second image features optimized by the channel dimension and the third image features optimized by the spatial dimension are combined through a cascade operation to obtain the enhanced backbone network features output as the first image features.

[0018] The first image features include edge features, color features, and texture features; Set the objective function as follows: ; The bounding box regression loss is used to guide the model to learn the precise location and size of electrical equipment, and the generalized intersection-union loss function is employed. The classification loss is used to guide the model to learn the correct fault type classification, and the weighted cross-entropy loss function is employed. This represents the confidence loss, used to guide the model in determining whether there are electrical devices within the prediction box. It employs the binary cross-entropy loss function. These are the weighting coefficients for each loss term, used to balance the contribution weights of different loss terms.

[0019] The objective function is minimized by the SGD optimizer, thereby achieving iterative optimization and convergence of the model parameters.

[0020] Guided by the objective function, a combined analysis of edge features, color features, and texture features is performed to obtain M combined features, including the following steps: Regions are divided according to edge features, and feature labels are set for edge features, color features, and texture features based on the region division results; Overlap analysis is performed on edge features, color features, and texture features based on feature labels. Edge features, color features, and texture features at different scales are superimposed and combined to obtain M combined features.

[0021] Probability prediction is performed on M combined features according to each anchor point of each scale to obtain N effective combined features, including the following steps: Adaptive anchor boxes are generated by clustering based on the device size distribution. The detection head predicts the bounding box offset, confidence, and class probability for each anchor point on the M combined features, generating M anchor boxes. Adaptive anchor boxes are generated based on device size distribution clustering, including: Extract the width and height of all real annotation boxes. After normalizing and filtering the width and height of the real annotation boxes, use K-class cluster analysis to obtain N sets of basic anchor box dimensions; Example: A learnable scaling factor α∈[0.8,1.2] is defined and applied to the base anchor box; initially α=1.0, during training it is found that small gaps are easy to miss detection, so α is automatically adjusted to 1.15 to expand the anchor box; in the first round of training, the dynamic threshold=0.4, and in the tenth round it is increased to 0.6, gradually tightening the positive sample standard.

[0022] Among the N=8 valid combination features, there are conductor cross-sectional features at different angles, covering 98% of the real fracture morphology.

[0023] Dynamically divide positive and negative samples based on the Intersection over Union (IoU): Foreground samples: IoU > dynamic threshold (increasing with training rounds); Background samples: IoU < minimum threshold (fixed value 0.3~0.4); Each iteration calculates the average IoU between the current anchor box and the ground plane; If the average IoU is lower than the target value (e.g., 0.6), then the range of α is increased; otherwise, the range is decreased.

[0024] Use a smooth function to transition threshold: f=f 0 +β (1) -exp ( -t / β 0)); In the formula, f Indicates the threshold. f 0 represents the base threshold. β This indicates the attenuation level, controlling the range of increase in the threshold. t Indicates the current training round. β 0 represents the decay rate, controlling the speed at which the threshold increases; Example: Discharge detection of disconnector ceramic pillars: With an initial threshold f0 = 0.4, a decay coefficient β = 0.1, and 50 training epochs t = 50: f = 0.4 + 0.1×(1 - exp(-50 / 10)) ≈ 0.4 + 0.1×(1-0.0067) ≈ 0.493; As training progresses, the threshold gradually increases to reduce interference from low-quality candidate boxes.

[0025] Results: In the later stages, only high-confidence proposals with IoU > 0.6 were retained, resulting in a 40% reduction in false positive rate.

[0026] When a new device type is detected, initiate rapid re-clustering (only new samples need to participate in the iteration); when a new GIS device detection requirement is added: Traditional methods require re-clustering all historical data; This solution uses only 200 newly acquired GIS images of basin-type insulators to quickly update the anchor frame size distribution; Advantages: The adaptation time for new devices is reduced from 7 days to 2 hours.

[0027] The anchor frame is automatically scaled based on the camera resolution.

[0028] By using confidence thresholding and non-maximum suppression, most overlapping and unreliable anchor boxes are filtered out, resulting in N effective combined features. In capacitor leakage detection: Initially, 200 candidate boxes are generated, and after filtering with a confidence threshold of 0.7, 50 remain. The NMS algorithm uses IoU=0.3 to remove overlapping boxes, and finally retains 3 non-overlapping high-confidence boxes; The output includes the device type (capacitor), fault type (leakage), and probability value (0.92).

[0029] The system directly drives the alarm system, prompting maintenance personnel to prioritize handling high-risk devices.

[0030] Among them, effective combination features include equipment type, fault type, and probability value; Power equipment faults are classified according to effective combination characteristics to achieve fault identification of power equipment.

[0031] Phase 1: Feature Extraction and Preliminary Enhancement (Backbone Network and CBAM); Preliminary feature extraction (CSPDarknet); The input is a 640x640x3 RGB image, which is first passed through the YOLO11 backbone network (CSPDarknet). This network consists of multiple convolutional layers and CSP modules, progressively downsampling the image to extract features from low to high levels. Specifically, this includes: Shallow networks capture basic features such as the edges and colors of insulators, wire textures, and sky backgrounds.

[0032] Mid-level networks combine basic features to form more complex features, such as the string structure of insulators or the outline of transformers.

[0033] Deep networks extract high-level semantic features, such as the overall concept of an "insulator" and abnormal texture patterns in "damaged" areas.

[0034] CBAM attention calibration applies the CBAM module to the key feature layer in Backbone, forming a two-level attention calibration: This approach integrates shallow CBAM, convolutional downsampling, and the C3k2 module. It learns which feature channels are most important for power equipment detection (e.g., highlighting "ceramic white" and "metallic luster," while suppressing "blue sky") through channel attention. Example: When processing example images, the model might assign lower weights to the "blue sky" channel and higher weights to channels related to "ceramic white" and "dark cracks." It learns to ignore irrelevant background and focus on the equipment itself and signs of failure. It also learns where location information is most useful (e.g., the area where insulators are located) through spatial attention. Example: The model might produce a bright response in areas with insulator strings and bird nests on towers, while responding weakly in open sky or uniform power line areas.

[0035] Deep CBAM: Applied between the C3k2 and SPPF modules. Attention calibration is performed on higher-level semantic features to enhance the high-level semantic understanding of fault types.

[0036] After two layers of CBAM calibration, background noise is effectively suppressed, and the characteristics of power equipment and fault areas are significantly enhanced, forming high-quality backbone network output characteristics.

[0037] Phase Two: Task-Oriented Feature Decoupling (Multi-Task Attention Module); Its input comes from the backbone network and is enhanced by CBAM to form a feature map F_cbam with rich semantic information.

[0038] Feature input and branches: The feature map F_cbam is simultaneously fed into two parallel attention branches.

[0039] The classification attention branch is a key feature needed to learn "What is the fault?". This branch generates an attention map A_cls, where highlighted areas represent the parts important to the classification task. Performance in the example: For "insulator damage", A_cls will focus strongly on cracks, missing pieces, or abnormal textures on the insulator discs. These are key local features that distinguish "normal" from "damaged".

[0040] For "bird's nest", A_cls will focus on the special texture structure formed by intertwined branches and grass stalks.

[0041] F_cls = F_cbam * A_cls. This is a "filtered" feature map, in which features irrelevant to fault type identification (such as the overall shape and absolute position of the equipment) are weakened, while the detailed features of the fault itself are amplified.

[0042] The regression attention branch provides key features needed to learn "Where is the faulty target? What are its boundaries?". This branch generates an attention map A_reg, where highlighted areas represent parts important for precise localization. Performance in the example: A_reg focuses on the overall outer contour of the insulator string and the boundary at the connection between the bird's nest and the tower. It is concerned with the edges, corners, and overall shape of the target in order to generate a tightly enclosing bounding box.

[0043] F_reg = F_cbam * A_reg. This feature map emphasizes the spatial structure and boundary information of the object.

[0044] Feature fusion combines the feature maps F_cls and F_reg, which are specific to two tasks (e.g., concatenating channels followed by a 1x1 convolution). It combines the opinions of "classification experts" and "localization experts" to form a more comprehensive and less conflicting enhanced feature map, F_enhanced. F_enhanced includes both the subtle textures used to identify "damage" and the contour information used to outline the "insulator string."

[0045] Phase 3: Multi-scale feature fusion (Neck-PANet); Its input is the enhanced feature map F_enhanced, which has been decoupled and fused for the task.

[0046] Feature pyramid construction: Neck (such as PANet) receives F_enhanced and constructs a multi-scale feature pyramid. It contains feature maps of different resolutions (e.g., 80x80, 40x40, 20x20).

[0047] A fusion of top-down and bottom-up approaches: High-level features (20x20) have rich semantic information, which helps in detecting large targets (such as the entire transformer).

[0048] The low-level features (80x80) retain more details and location information, which helps detect small targets (such as tiny damage on a single insulator).

[0049] PANet fuses features from different layers through upsampling and downsampling operations, ensuring that each layer's features possess both strong semantics and high resolution. Its role in the example: The 80x80 detail helps to accurately locate small areas of insulator damage.

[0050] The 40x40 feature may be most effective for medium-sized bird nests.

[0051] The 20x20 semantic features ensure that the model does not misidentify the bird's nest as other objects.

[0052] Phase Four: Detection and Output (Detection Head); Its input consists of three fully fused and enhanced feature maps from Neck at three different scales.

[0053] Decoupling head prediction: YOLO11's decoupling head performs three predictions in parallel for each anchor at each scale: Classification branch: Based on the features highlighted by A_cls in F_enhanced, predict the probability that the object in the box belongs to each fault category (e.g., "insulator damage: 0.95", "bird's nest: 0.90").

[0054] Regression branch: Based on the features highlighted by A_reg in F_enhanced, predict the precise coordinates (x, y, w, h) of the bounding box.

[0055] Confidence branch: Predicts the probability that an object exists within the bounding box.

[0056] Post-processing (Non-maximum suppression - NMS): The model generates a large number of overlapping prediction boxes.

[0057] The NMS algorithm filters out redundant boxes based on confidence scores and IoU (Intersection over Union), keeping only the most reliable one.

[0058] Therefore, the present invention adopts the above-mentioned power equipment fault identification method based on the improved YOLO model, which achieves significant improvement in detection accuracy, robustness and adaptability while maintaining real-time performance, and is particularly suitable for intelligent inspection needs in complex scenarios such as transmission lines and substations.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for fault identification of power equipment based on an improved YOLO model, characterized in that, Includes the following steps: Acquire and preprocess image data of power equipment; Multiple feature extractions are performed on the image data to obtain first image features at different scales; the first image features include edge features, color features, and texture features; Set an objective function, and use the objective function as a guide to perform combined analysis on edge features, color features and texture features to obtain M combined features; M combined features are probabilistically predicted at each anchor point of each scale to obtain N effective combined features; among them, the effective combined features include equipment type, fault type and probability value; Power equipment faults are classified according to effective combination characteristics to achieve fault identification of power equipment.

2. The power equipment fault identification method based on the improved YOLO model according to claim 1, characterized in that, The improved backbone network performs multiple feature extractions on image data to obtain first image features at different scales, including the following steps: The image data is processed by the first branch, and then obtained by splicing double pooling, double-layer MLP compression and expansion, and adaptive weight adjustment to obtain second image features with higher correlation to the fault point. The image data is processed by the second branch, and after being processed by the standard convolutional block, it is processed by dimensionality reduction, spatial feature fusion and spatial fine-tuning to obtain the third image features with multi-scale receptive fields. The second image features optimized by the channel dimension and the third image features optimized by the spatial dimension are combined through a cascade operation to obtain the enhanced backbone network feature output.

3. The power equipment fault identification method based on the improved YOLO model according to claim 1, characterized in that, The objective function is: ； In the formula, The bounding box regression loss is used to guide the model to learn the precise location and size of electrical equipment, and the generalized intersection-union loss function is employed. The classification loss is used to guide the model to learn the correct fault type classification, and the weighted cross-entropy loss function is employed. This represents the confidence loss, used to guide the model in determining whether there are electrical devices within the prediction box. It employs the binary cross-entropy loss function. These are the weighting coefficients for each loss term, used to balance the contribution weights of different loss terms.

4. The power equipment fault identification method based on the improved YOLO model according to claim 1, characterized in that, The feature fusion network performs weighted fusion of multi-scale features to obtain M combined features, including the following steps: Regions are divided according to edge features, and feature labels are set for edge features, color features, and texture features based on the region division results; Overlap analysis is performed on edge features, color features, and texture features based on feature labels. Edge features, color features, and texture features at different scales are superimposed and combined to obtain M combined features.

5. The power equipment fault identification method based on the improved YOLO model according to claim 1, characterized in that, Probability prediction is performed on M combined features according to each anchor point at each scale to obtain N effective combined features, including the following steps: Adaptive anchor boxes are generated by clustering based on the device size distribution. The detection head predicts the bounding box offset, confidence, and class probability for each anchor point on the M combined features, generating M anchor boxes. By using confidence thresholding and non-maximum suppression, most overlapping and unreliable anchor boxes are filtered out, resulting in N effective combined features.

6. The power equipment fault identification method based on the improved YOLO model according to claim 5, characterized in that, Adaptive anchor boxes are generated based on device size distribution clustering, including: Extract the width and height of all real annotation boxes, normalize and filter the width and height of the real annotation boxes, and then use K-class cluster analysis to obtain N sets of basic anchor box dimensions. Define a learnable scaling factor α∈[0.8,1.2] to act on the base anchor frame; Positive and negative samples are dynamically divided based on the Intersection over Union (IoU). In each iteration, the average IoU between the current anchor box and the ground plane is calculated; If the average IoU is lower than the target value, then expand the range of α; otherwise, narrow the range. Use a smooth function to transition threshold: f=f 0 +β (1) -exp ( -t / β 0)); In the formula, f Indicates the threshold. f 0 represents the base threshold. β This indicates the attenuation level, controlling the range of increase in the threshold. t Indicates the current training round. β 0 represents the decay rate, which controls the speed at which the threshold increases; When a new device type is detected, initiate rapid re-clustering; The anchor frame is automatically scaled based on the camera resolution.

7. A power equipment fault identification system based on an improved YOLO model according to claim 1, characterized in that, include: The feature extraction and preliminary enhancement module is used to suppress background noise, and the features of power equipment and its potential fault areas are significantly enhanced. The multi-task attention module enables the model to flexibly switch its focus across multiple object detection tasks by perceiving different tasks. Multi-scale feature fusion is used to fuse features from different levels, so that each layer of features has both strong semantics and high resolution. The detection head generates a large number of overlapping prediction boxes, and finally retains only the most reliable prediction box to output the recognition result.

Images