An image recognition-based power equipment abnormal state intelligent identification system

Abstract

This invention discloses an intelligent identification system for abnormal states of power equipment based on image recognition, comprising the following modules: an acquisition and suppression module for acquiring inspection image sequences and generating a set of component hints and a reflection suppression weight map; a component segmentation module for inputting an improved SAM2 model to generate component masks; a mask self-healing module for cross-view propagation and mask self-healing to obtain a set of component masks; an interface micro-segmentation module for generating connection interface bands and obtaining connection interface micromasks; an evidence extraction module for extracting texture, morphology, and connection interface evidence; a screening and updating module for generating an abnormal candidate list based on stability, evidence strength, and interface consistency; and a structured recognition module for inputting abnormal candidates into Qwen2-VL and outputting structured recognition results. This invention improves the robustness and interpretability of abnormal identification in reflective and multi-view scenarios.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection and image recognition technology for power equipment, and in particular to an intelligent identification system for abnormal states of power equipment based on image recognition. Background Technology

[0002] With the expansion of power grids and the normalization of intelligent inspections, image recognition and intelligent discrimination technologies for identifying external defects, abnormal connection interfaces, and potential operational hazards in power equipment have received widespread attention. Existing technologies typically employ manual, experience-based inspections, defect detection based on traditional image features, or single-frame identification of equipment components using general object detection / semantic segmentation models, followed by simple thresholding or classifiers to output anomaly conclusions. However, these technologies generally suffer from the following problems in practical applications: The acquired inspection images are susceptible to reflections, haze, overexposure, and occlusion, leading to blurred component boundaries, mask drift, and missegmentation. Furthermore, the lack of suppression and error correction mechanisms for reflective areas makes it difficult to reliably obtain component masks for subsequent analysis. In multi-view inspection scenarios, existing methods primarily process single frames independently, lacking cross-view propagation and stability quantification, making it difficult to identify deformation drift caused by viewpoint changes and local defects caused by occlusion. Interface anomalies typically occur at the junctions of adjacent components, and general segmentation often fails to maintain fine consistency within narrow-band interfaces, resulting in interface cracks, loosening, and ablation being interfered with by background textures. Simultaneously, existing anomaly judgments often rely on single scores or coarse-grained labels, lacking structured encapsulation, threshold mapping, and consistency verification of texture, morphology, and interface evidence, making it difficult to form traceable evidence citations and handling suggestions, thus affecting the reliable output of inspection results and the closure of work orders.

[0003] Therefore, how to provide an intelligent identification system for abnormal states of power equipment based on image recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose an intelligent identification system for abnormal states of power equipment based on image recognition. This invention utilizes inspection image sequences and quality and operating condition parameters to generate a component prompt set and a reflection suppression weight map, and employs an improved SAM2 to achieve component segmentation. The drift type is determined by cross-view propagation and stability indicators, triggering point filling, bounding box filling, and reliable frame rollback to complete mask self-healing, generating a micromask of the connection interface and extracting texture, morphology, and interface evidence, filtering and updating to obtain abnormal candidates, and finally outputting structured results by Qwen2-VL. It has the advantages of anti-reflection, cross-view stability, accurate recognition, and interpretability.

[0005] An intelligent identification system for abnormal states of power equipment based on image recognition according to an embodiment of the present invention includes the following modules: The acquisition and suppression module is used to acquire inspection image sequences and record the inspection image quality and working condition characteristic parameters, and generate component prompt set and reflection suppression weight map; The component segmentation module is used to input the inspection image sequence, component cue set and reflection suppression weight map into the improved SAM2 model to generate component masks; The mask self-healing module is used to perform cross-view propagation on the component mask, generate mask stability index and determine the drift type, trigger point correction, box correction or trusted frame rollback to achieve mask self-healing, and obtain the component mask set. The interface micro-segmentation module is used to generate the connection interface band, and together with the inspection image sequence and the reflection suppression weight map, it is input into the improved SAM2 model to perform local segmentation and obtain the connection interface micro-mask. The evidence extraction module is used to extract texture evidence, morphological evidence and connection interface evidence from the component mask set and the micro-mask of the connection interface to form a candidate anomaly evidence set. The filtering and updating module is used to filter and update the confidence level of candidate anomaly evidence sets based on mask stability index, evidence strength of candidate anomaly evidence sets and consistency of connection interface evidence, so as to obtain an anomaly candidate list. The structured recognition module is used to input the component views, mask locations, and evidence values ​​corresponding to the anomaly candidate list into the Qwen2-VL model and output the structured recognition results.

[0006] An intelligent identification method for abnormal states of power equipment based on image recognition according to an embodiment of the present invention includes the following steps: Step 1: Collect inspection image sequences of the same power equipment and record the inspection image quality and operating condition characteristic parameters. Generate a component prompt set based on the power equipment type and component structure prior, and generate a reflection suppression weight map based on the inspection image quality and operating condition characteristic parameters. Step 2: Input the inspection image sequence, component cue set, and reflection suppression weight map into the improved SAM2 model to generate a component mask. The improved SAM2 model includes an image encoding module, a cue encoding module, a memory construction and propagation module, and a mask decoding module. Step 3: Perform cross-view propagation on the component mask, generate mask stability index based on area drift, centroid drift, boundary jitter and cross-view overlap, and determine the drift type. Trigger point correction, box correction or trusted frame rollback to achieve mask self-healing and obtain the component mask set. Step 4: Generate a connection interface band based on the component mask set, and input the inspection image sequence, the connection interface band and the reflection suppression weight map into the improved SAM2 model to perform local segmentation to obtain the connection interface micromask; Step 5: Extract texture evidence, morphological evidence, and connection interface evidence from the component mask set and the micromask at the connection interface to form a candidate anomaly evidence set; Step 6: Based on the mask stability index, the evidence strength of the candidate anomaly evidence set, and the consistency of the connection interface evidence, the candidate anomaly evidence set is screened and the confidence level is updated to obtain the anomaly candidate list; Step 7: Input the component views, mask locations, and evidence values ​​corresponding to the anomaly candidate list into the Qwen2-VL model, and output a structured identification result containing the anomaly component, anomaly phenomenon, risk level, evidence citation, and handling suggestions.

[0007] Optionally, step one specifically includes: Collect a sequence of visible light inspection images of the same power equipment within the same inspection task cycle from multiple perspectives, and record the focal length, exposure and shooting distance parameters for each perspective inspection image; The image quality and working condition characteristics parameters of the inspection image are calculated by combining the sharpness parameter, blur parameter, brightness overexposure ratio parameter, haze index parameter, reflectivity ratio parameter and occlusion ratio parameter of the inspection image. The type of power equipment is determined by the equipment ledger or inspection task information. The prior knowledge of the component structure is provided by a pre-set equipment component structure knowledge base, which includes the power equipment type and component list, as well as the adjacency and connection relationships between components. Based on the prior knowledge of power equipment type and component structure, generate a component hint set for each component according to the component list; Based on the reflectivity ratio parameter, the bright areas are weighted and a reflectivity suppression weight map is generated.

[0008] Optionally, the improved SAM2 model includes an image encoding module, a cue encoding module, a memory construction and propagation module, and a mask decoding module: The image encoding module performs concatenated encoding of convolutional layers and layer normalization layers on the inspection images from each viewpoint in the inspection image sequence, and performs feature transformation through the GELU activation function to form multi-scale image features. A reflection suppression guidance mechanism is introduced, and the reflection suppression weight map is aligned by bilinear interpolation according to the scale and then multiplied element-wise with the feature map of the corresponding scale to complete the weight reduction of the reflection area features. The weighted feature map is then subjected to channel dimension weighted summation to obtain the multi-scale image features after reflection suppression. The prompt encoding module performs embedding encoding on the component prompt set to obtain the component prompt embedding vector. It introduces a joint segmentation mechanism of skeleton prompt and boundary prompt. The point prompt embedding vector in the component prompt embedding vector is defined as the skeleton prompt, and the box prompt embedding vector in the component prompt embedding vector and the component mask boundary band sampling points output by the mask decoding module in the previous round are defined as the boundary prompt. The prompt encoding module uses a multi-head attention layer to interactively fuse the skeleton prompt and the boundary prompt. The multi-head attention layer projects the skeleton prompt through the first linear mapping layer to obtain the query vector, projects the boundary prompt through the second linear mapping layer to obtain the key vector, and projects it through the third linear mapping layer to obtain the value vector. It calculates the attention weight based on the query vector and the key vector and normalizes it through the Softmax function. It uses the attention weight to perform a weighted summation on the value vector to obtain the fused vector, and outputs the joint prompt embedding vector through the feedforward network layer and the residual connection. The mask decoding module uses a cross-attention layer to calculate the attention mapping of the joint cue embedding vector to the multi-scale image features, and performs element-wise addition of the attention mapping result with the multi-scale image features to form fused features. The mask decoding module also outputs the component mask quality score. The memory construction and propagation module performs element-wise multiplication between the component mask and the corresponding multi-scale feature map to obtain the features within the mask, and encodes them into memory features through convolutional layers and layer normalization layers and writes them into the memory queue. A component-specific memory gating mechanism is introduced. For each component, the inspection image quality, working condition feature parameters and component mask quality score are read separately. According to the preset strong update flag, weak update flag and no update flag, the corresponding memory writing weight is selected, and after performing element-wise multiplication with the memory features, it is written into the memory queue. The module outputs a reliable frame index and a memory rollback candidate set. Output the component mask, component mask quality score, memory features, and trusted frame index and memory rollback candidate set corresponding to the inspection image sequence.

[0009] Optionally, step three specifically includes: The system receives component masks, component mask quality scores, memory features, trusted frame indexes, and memory rollback candidate sets. It performs cross-view propagation on each component mask based on memory features, using adjacent view inspection images as units. It also determines the propagation update method and whether to trigger mask self-healing based on the component mask quality score. For the same component, the number of mask pixels is counted for each component mask in two adjacent viewpoints to obtain the area drift. The mean value of the pixel coordinates in the mask is calculated to obtain the centroid coordinates and the centroid drift is calculated. The set of pixels at the mask boundary is extracted and the boundary jitter is calculated. The component masks in two adjacent viewpoints are clipped and aligned according to the component's bounding rectangle and the intersection-union ratio is calculated to obtain the cross-viewpoint overlap. The mask stability index is obtained by linearly weighting area drift, centroid drift, boundary jitter and cross-viewpoint overlap according to preset weights, and the drift region is determined based on the mask stability index and the symmetric difference pixel set. The set of pixels with a pixel weight greater than a preset weight threshold in the reflection suppression weight map is defined as a high weight region. When the overlap ratio between the drift region and the high weight region in the reflection suppression weight map is greater than the preset reflection overlap threshold, it is determined to be a reflection adsorption drift. When the overlap ratio between the drift region and the occlusion region is greater than the preset occlusion overlap threshold, it is determined to be an occlusion drift. After excluding reflection adsorption drift and occlusion drift, the viewpoint deformation drift is determined based on the change in the aspect ratio of the bounding rectangle and the change in the principal axis direction. Mask self-healing is triggered based on the drift type, and the mask self-healing includes point correction, bounding box correction, and trusted frame rollback; Output the set of cross-view consistent component masks after mask self-healing and the corresponding mask stability index.

[0010] Optionally, step four specifically includes: Based on the consistent set of component masks across different viewpoints, component pairs with interconnected relationships are identified, and the first and second component masks of the component pairs are extracted in each viewpoint. Morphological dilation and erosion are performed on the first component mask and the second component mask respectively, and the difference is taken to obtain the first boundary band and the second boundary band. The candidate pixel set of the connection interface is obtained by filtering based on the nearest distance between the first boundary band and the second boundary band. The connection interface band is generated by connected component filtering and morphological closing operation on the candidate pixel set of the connection interface. The inspection images corresponding to the connecting interface bands in the inspection image sequence, the connecting interface bands and the reflection suppression weight map are input into the improved SAM2 model to perform local segmentation, and the connecting interface micromasks corresponding to each inspection image are output.

[0011] Optionally, step five specifically includes: Evidence is extracted for each component mask and its corresponding connection interface micromask using the inspection image as a unit. After removing pixels with weights exceeding a preset weight threshold in the reflection suppression weight map within the pixel area defined by the component mask, texture evidence is calculated. Morphological evidence is calculated within the pixel area defined by the component mask, and connection interface evidence is calculated within the pixel area defined by the connection interface micromask. Texture evidence, morphological evidence, and connection interface evidence are structurally encapsulated according to "component identifier, viewpoint number, mask identifier, evidence item name, and evidence value" to form a candidate anomaly evidence set.

[0012] Optionally, step six specifically includes: The candidate abnormal evidence set is grouped according to component identification and viewpoint number, and the evidence strength is obtained by interval mapping of the evidence value of each evidence item according to the preset evidence threshold table. When the mask stability index is less than the preset stability threshold or the component mask quality score is less than the preset quality threshold, the evidence strength under the current view is downweighted and marked as a verification state. A consistency check is performed on the connection interface evidence. The consistency check is to compare the strength of the connection interface evidence with the changes in the number of connected components and the maximum area ratio of the connected component in the connection interface micromask in adjacent viewpoints to see if they meet a preset consistency threshold. Based on the reduced-weighted evidence strength and consistency verification results, the texture evidence, morphological evidence and connection interface evidence of the same component are weighted and fused to generate anomaly confidence and form an anomaly candidate list.

[0013] Optionally, step seven specifically includes: And based on the component identifier and viewpoint number in the anomaly candidate list, the corresponding component view is extracted from the inspection image sequence; The component view is overlaid with the component mask and the connection interface micro-mask to generate a mask positioning map, and the evidence item names and evidence values ​​in the anomaly candidate list are structured and organized to form evidence text. The component view, mask positioning map and evidence text are input into the Qwen2-VL model to generate structured recognition results, which include abnormal components, abnormal phenomena, risk levels, evidence citations and handling suggestions. The structured recognition results are associated with and stored with the list of anomaly candidates, and then output.

[0014] The beneficial effects of this invention are: Under the real-world conditions of "high interference, strong on-site presence, and traceability" in power equipment inspection, the anomaly identification process has been upgraded from an unstable process of "single-frame segmentation + single-score judgment" to an engineering system of "stable segmentation - cross-perspective self-healing - interface fine inspection - evidence closed loop - structured output," thereby significantly improving the reliability, interpretability, and feasibility of the results. Its importance lies in the fact that defects in power equipment often manifest in their early stages as subtle texture changes, local boundary damage, or abnormal connection interfaces, which are easily masked by reflections, overexposure, haze, and occlusion. This invention, through full-process constraints of the reflection suppression weight map, the combined prompts and memory gating of the improved SAM2, and drift judgment and error correction rollback driven by stability indicators, ensures that the component mask remains consistent and self-recovering under multi-view conditions, reducing the risks of false positives, missed positives, and drift accumulation. Simultaneously, through local segmentation of connection interfaces with defined boundaries and structured encapsulation of texture / morphology / interface evidence, threshold mapping, and consistency verification, the "anomaly judgment" is transformed from a black-box score into a referable chain of evidence. Finally, Qwen2-VL outputs a structured result containing risk levels and handling suggestions, facilitating direct integration into inspection work orders and maintenance decisions. Overall, this invention not only improves the accuracy of single-shot identification but also enhances the stability and auditability of long-term inspections, possessing key engineering value for ensuring the safe operation of power grid equipment, reducing manual review costs, and improving defect closure efficiency. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart is a process for an intelligent identification system for abnormal states of power equipment based on image recognition proposed in this invention. Figure 2 This is a schematic diagram of an intelligent identification method for abnormal states of power equipment based on image recognition proposed in this invention; Figure 3 This is a framework diagram of the improved SAM2 model in the intelligent identification method for abnormal states of power equipment based on image recognition proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figure 1 An intelligent identification system for abnormal states of power equipment based on image recognition includes the following modules: The acquisition and suppression module is used to acquire inspection image sequences and record the inspection image quality and working condition characteristic parameters, and generate component prompt set and reflection suppression weight map; The component segmentation module is used to input the inspection image sequence, component cue set and reflection suppression weight map into the improved SAM2 model to generate component masks; The mask self-healing module is used to perform cross-view propagation on the component mask, generate mask stability index and determine the drift type, trigger point correction, box correction or trusted frame rollback to achieve mask self-healing, and obtain the component mask set. The interface micro-segmentation module is used to generate the connection interface band, and together with the inspection image sequence and the reflection suppression weight map, it is input into the improved SAM2 model to perform local segmentation and obtain the connection interface micro-mask. The evidence extraction module is used to extract texture evidence, morphological evidence and connection interface evidence from the component mask set and the micro-mask of the connection interface to form a candidate anomaly evidence set. The filtering and updating module is used to filter and update the confidence level of candidate anomaly evidence sets based on mask stability index, evidence strength of candidate anomaly evidence sets and consistency of connection interface evidence, so as to obtain an anomaly candidate list. The structured recognition module is used to input the component views, mask locations, and evidence values ​​corresponding to the anomaly candidate list into the Qwen2-VL model and output the structured recognition results.

[0018] refer to Figure 2-3 A method for intelligent identification of abnormal states of power equipment based on image recognition includes the following steps: Step 1: Collect inspection image sequences of the same power equipment and record the inspection image quality and operating condition characteristic parameters. Generate a component prompt set based on the power equipment type and component structure prior, and generate a reflection suppression weight map based on the inspection image quality and operating condition characteristic parameters. Step 2: Input the inspection image sequence, component cue set, and reflection suppression weight map into the improved SAM2 model to generate a component mask. The improved SAM2 model includes an image encoding module, a cue encoding module, a memory construction and propagation module, and a mask decoding module. Step 3: Perform cross-view propagation on the component mask, generate mask stability index based on area drift, centroid drift, boundary jitter and cross-view overlap, and determine the drift type. Trigger point correction, box correction or trusted frame rollback to achieve mask self-healing and obtain the component mask set. Step 4: Generate a connection interface band based on the component mask set, and input the inspection image sequence, the connection interface band and the reflection suppression weight map into the improved SAM2 model to perform local segmentation to obtain the connection interface micromask; Step 5: Extract texture evidence, morphological evidence, and connection interface evidence from the component mask set and the micromask at the connection interface to form a candidate anomaly evidence set; Step 6: Based on the mask stability index, the evidence strength of the candidate anomaly evidence set, and the consistency of the connection interface evidence, the candidate anomaly evidence set is screened and the confidence level is updated to obtain the anomaly candidate list; Step 7: Input the component views, mask locations, and evidence values ​​corresponding to the anomaly candidate list into the Qwen2-VL model, and output a structured identification result containing the anomaly component, anomaly phenomenon, risk level, evidence citation, and handling suggestions.

[0019] In this embodiment, step one specifically includes: Collect a sequence of visible light inspection images of the same power equipment within the same inspection task cycle from multiple perspectives, and record the focal length, exposure and shooting distance parameters for each perspective inspection image; The sharpness parameter is calculated based on the grayscale gradient change of the viewpoint inspection image. The sharpness parameter is the gradient magnitude statistics of the viewpoint inspection image after grayscale conversion within the entire image range. The blur parameter is calculated based on the edge expansion width of the viewpoint inspection image. The blur parameter is the edge expansion width statistics of the viewpoint inspection image in multiple directions. The brightness overexposure ratio parameter is calculated based on the proportion of pixels in the viewpoint inspection image whose pixel brightness reaches the saturation threshold. The haze index parameter is calculated based on the dark channel intensity and contrast attenuation of the viewpoint inspection image. The reflection ratio parameter is calculated based on the connected region area ratio and edge density of the bright region in the viewpoint inspection image. The bright region is the set of pixels whose pixel brightness exceeds a preset brightness threshold and whose saturation is lower than a preset saturation threshold. The occlusion ratio parameter is calculated based on the structural differences and edge gradients between adjacent viewpoints of the viewpoint inspection image. The occlusion ratio parameter is the proportion of connected region area of ​​pixels whose edge gradient is lower than a preset gradient threshold and whose structural differences occur between adjacent viewpoints. The parameters of sharpness, blur, brightness overexposure ratio, haze index, reflectivity ratio, and occlusion ratio are combined to form the inspection image quality and working condition characteristic parameters. The type of power equipment is determined by the equipment ledger or inspection task information. The prior knowledge of the component structure is provided by a pre-set equipment component structure knowledge base, which includes the power equipment type and component list, as well as the adjacency and connection relationships between components. Based on the prior knowledge of the power equipment type and component structure, a component hint set is generated for each component according to the component list. The component hint set includes component center point hints and component circumscribed rectangle hints. Based on the reflection ratio parameter, a weight mapping is performed on the bright area to generate a reflection suppression weight map. The weight mapping sets the pixel weight of the connected domain of the bright area to be less than the pixel weight of the non-bright area.

[0020] In this embodiment, the improved SAM2 model includes an image encoding module, a cue encoding module, a memory construction and propagation module, and a mask decoding module: The image encoding module performs concatenated encoding of convolutional layers and layer normalization layers on the inspection images from each viewpoint in the inspection image sequence and performs feature transformation through the GELU activation function to form multi-scale image features. The convolutional layer includes stride convolution for downsampling and pointwise convolution for feature fusion. A reflection suppression guidance mechanism is introduced. The reflection suppression weight map is aligned by bilinear interpolation according to the scale and then multiplied element-wise with the feature map of the corresponding scale to complete the weight reduction of the reflection area features. The weighted feature map is then subjected to channel-dimensional weighted summation to obtain the multi-scale image features after reflection suppression. The prompt encoding module performs embedding encoding on the component prompt set to obtain component prompt embedding vectors. The component center point prompt is mapped to a point prompt embedding vector through position encoding and linear layer mapping. The component's circumscribed rectangle prompt is concatenated with its two corner point position encodings and linear layer mappings to obtain a box prompt embedding vector. The point prompt embedding vectors and box prompt embedding vectors are then concatenated to form a component prompt embedding vector sequence. A joint segmentation mechanism of skeleton prompts and boundary prompts is introduced. The point prompt embedding vectors in the component prompt embedding vectors are defined as skeleton prompts, and the box prompt embedding vectors in the component prompt embedding vectors are sampled with the component mask boundary bands output by the mask decoding module in the previous round. Points are collectively defined as boundary cues, where the boundary band is the differential region obtained by morphological dilation and erosion of the component mask. The cue encoding module uses a multi-head attention layer to interactively fuse skeleton cues and boundary cues. The multi-head attention layer projects the skeleton cues through a first linear mapping layer to obtain a query vector, projects the boundary cues through a second linear mapping layer to obtain a key vector, and projects them through a third linear mapping layer to obtain a value vector. Attention weights are calculated based on the query vector and the key vector and normalized using the Softmax function. The value vector is then weighted and summed using the attention weights to obtain a fused vector, which is then output as a joint cue embedding vector through a feedforward network layer and a residual connection. The mask decoding module uses a cross-attention layer to calculate the attention mapping of the joint cue embedding vector to multi-scale image features, and performs element-wise addition of the attention mapping result with the multi-scale image features to form a fused feature. The fused feature is then upsampled to restore the spatial resolution step by step, and then outputs the probability map of the component mask through a convolutional layer and obtains the component mask through a sigmoid function. The mask decoding module also outputs the component mask quality score, which is obtained by linearly weighting and fusing the average confidence of the component mask probability map, the boundary gradient consistency statistic, and the cross-viewpoint overlap statistic. The memory construction and propagation module performs element-wise multiplication between the component mask and the corresponding multi-scale feature map to obtain the features within the mask. These features are then encoded into memory features by convolutional layers and layer normalization layers and written into the memory queue. A component-specific memory gating mechanism is introduced. For each component, the inspection image quality, working condition feature parameters, and component mask quality score are read separately. The corresponding memory writing weights are selected according to preset strong update flags, weak update flags, and no update flags. The memory writing weight corresponding to the strong update flag is 1, the memory writing weight corresponding to the weak update flag is less than 1 and greater than 0, and the memory writing weight corresponding to the no update flag is 0. After performing element-wise multiplication with the memory features, the memory features are written into the memory queue. A reliable frame index and a memory rollback candidate set are output. The reliable frame index is the set of view frame numbers whose component mask quality scores meet a preset threshold. The memory rollback candidate set is the set of memory features corresponding to the reliable frame index. Output the component mask, component mask quality score, memory features, and trusted frame index and memory rollback candidate set corresponding to the inspection image sequence.

[0021] This implementation uses SAM2 as the basic model for component segmentation. Its "cue-driven segmentation + memory propagation" structure is adapted to multi-component, multi-view inspection scenarios in power equipment. Compared to traditional single-frame segmentation, it maintains cross-view consistency more easily. Based on this, the four modules of SAM2—image encoding, cue encoding, mask decoding, and memory construction and propagation—are kept unchanged, but targeted improvements are made: In image encoding, a reflection suppression weight map is introduced. After bilinear alignment, element-wise multiplication is used to reduce weights, followed by channel-weighted summation to suppress missegmentation caused by reflection pseudo-textures. In cue encoding, skeleton cues and boundary cues are used for joint segmentation. Multi-head attention is used to fuse the skeleton as a query and the boundary as a key to output a joint cue embedding vector, enhancing boundary convergence and the separability of adjacent components. At the decoding end, component mask quality scores are output. At the memory end, a component-specific memory gating system is introduced, with strong / weak / no updates written and outputting a reliable frame index and rollback candidate set to avoid low-quality frames contaminating the memory and improve the stability and error correction of long sequences.

[0022] In this embodiment, step three specifically includes: The system receives component masks, component mask quality scores, memory features, trusted frame indexes, and memory rollback candidate sets. It performs cross-view propagation on each component mask based on memory features, using adjacent view inspection images as units. It also determines the propagation update method and whether to trigger mask self-healing based on the component mask quality score. For the same component, the number of mask pixels is counted for the component masks in two adjacent viewpoints to obtain the number of mask pixels in the first viewpoint and the second viewpoint. The absolute value of the difference between the two pixel counts is used as the area change. The normalization benchmark is the number of mask pixels in the first viewpoint or the average number of mask pixels in both viewpoints. After normalization, the area change is obtained as the area drift. The centroid coordinates are obtained by calculating the mean of the pixel coordinates within the mask. The pixel distance between the centroid coordinates of two adjacent viewpoints is calculated as the centroid change. The normalization benchmark is the diagonal pixel length of the component's bounding rectangle in the first viewpoint or the average of the diagonal pixel lengths of the bounding rectangles in both viewpoints. The centroid change after normalization is used as the centroid drift. The mask boundary pixel set is extracted. For each boundary pixel in the first view boundary pixel set, the pixel distance to each boundary pixel in the second view boundary pixel set is calculated and the minimum distance is taken to obtain the minimum corresponding distance of the boundary pixel. The unidirectional boundary offset is obtained by averaging all the minimum corresponding distances of the first view boundary pixel set. The above process is repeated for the second view boundary pixel set to obtain the reverse boundary offset. The unidirectional boundary offset and the reverse boundary offset are averaged as the boundary jitter. The cross-view overlap ratio is calculated after the masks of adjacent view components are clipped and aligned according to the bounding rectangle of the component. The mask stability index is obtained by linearly weighting area drift, centroid drift, boundary jitter, and cross-viewpoint overlap according to preset weights. The drift region is determined based on the mask stability index and the symmetric difference pixel set. The part masks of the same part in two adjacent views are respectively recorded as the first view part mask and the second view part mask. The pixel-level XOR operation is performed on the two to obtain the symmetric difference pixel set. The connected component analysis is performed on the symmetric difference pixel set to obtain multiple differential connected components. The number of pixels in each differential connected component is counted. Differential connected components with a number of pixels greater than the preset connected component area threshold are merged to obtain the drift region. When the mask stability index is less than the preset stability threshold, the drift region is enabled for drift type determination. When the mask stability index is greater than or equal to the stability threshold, no drift region is generated or the drift region is set to an empty set. The set of pixels with a pixel weight greater than a preset weight threshold in the reflection suppression weight map is defined as a high weight region. When the overlap ratio between the drift region and the high weight region in the reflection suppression weight map is greater than the preset reflection overlap threshold, it is determined to be a reflection adsorption drift. When the overlap ratio between the drift region and the occlusion region is greater than the preset occlusion overlap threshold, it is determined to be an occlusion drift. After excluding reflection adsorption drift and occlusion drift, the viewpoint deformation drift is determined based on the change in the aspect ratio of the bounding rectangle and the change in the principal axis direction. Mask self-healing is triggered based on the drift type. The mask self-healing includes point correction, bounding box correction, and trusted frame rollback. Point correction involves selecting the pixel with the largest boundary gradient magnitude in the drift region to generate a correction point prompt and passing it to the prompt encoding module of the improved SAM2 model to regenerate the joint prompt embedding vector. Bounding box correction involves generating a constraint box prompt using the bounding rectangle of the union of the masks of two adjacent viewpoint components and passing it to the prompt encoding module to regenerate the joint prompt embedding vector. Trusted frame rollback involves selecting the trusted frame number with the smallest distance from the current viewpoint number from the trusted frame index, reading the corresponding memory feature from the memory rollback candidate set to replace the current memory feature, and then performing cross-viewpoint propagation. Output the set of cross-view consistent component masks after mask self-healing and the corresponding mask stability index.

[0023] This implementation uses a drift identification and self-healing strategy driven by stability indicators to achieve a closed-loop correction process of "drift detection - cause location - automatic error correction / rollback", reducing the cost of manual review and reshooting, and providing a more reliable mask basis for interface micro-segmentation and evidence judgment.

[0024] In this embodiment, step four specifically includes: Based on the consistent set of component masks across different viewpoints, component pairs with interconnected relationships are identified, and the first and second component masks of the component pairs are extracted in each viewpoint. Morphological dilation and erosion are performed on the first component mask and the second component mask respectively, and the difference is taken to obtain the first boundary band and the second boundary band. The candidate pixel set of the connection interface is obtained by filtering based on the nearest distance between the first boundary band and the second boundary band. The connection interface band is generated by connected component filtering and morphological closing operation on the candidate pixel set of the connection interface. The inspection images corresponding to the connecting interface band in the inspection image sequence, the connecting interface band and the reflection suppression weight map are input into the improved SAM2 model to perform local segmentation. The local segmentation is to output the connecting interface probability map only within the limited range of the connecting interface band and obtain the connecting interface micromask through the Sigmoid function, and set the connecting interface probability map outside the limited range of the connecting interface band to zero, and output the connecting interface micromask corresponding to each view inspection image.

[0025] This implementation method focuses the identification on the narrow band area at the junction of components by connecting the interface with constraints for local segmentation. This significantly improves the detection rate and positioning accuracy of minor interface anomalies such as cracks, loosening, and ablation, while reducing false alarms caused by background texture interference.

[0026] In this embodiment, step five specifically includes: Evidence extraction is performed on each component mask and its corresponding connecting interface micromask, using the viewpoint inspection image as a unit. Within the pixel region defined by the component mask, pixels with a weight exceeding a preset weight threshold in the reflection suppression weight map are removed. Texture evidence is then calculated, including: grayscale conversion of the pixel region and calculation of the gradient magnitude map; the percentage of pixels with gradient magnitudes above a preset threshold is used as the texture mutation density; local binary pattern encoding is performed on the grayscale result, and the histogram of the local binary pattern is calculated; the proportion of the main peak and the entropy value of the histogram are extracted as texture uniformity indicators; multi-directional Gabor filtering is performed on the grayscale result, and the mean and variance of the filtering response in each direction are calculated as directional texture intensity. Morphological evidence is then calculated within the pixel region defined by the component mask, including: extracting the outer contour of the component mask and calculating... The number of indentations and the depth of indentations on the outer contour are used as contour gap indicators. The ratio of the area of ​​the component mask to the area of ​​the circumscribed rectangle is used as shape compactness. The boundary gradient magnitude of pixels within the boundary band of the component mask is calculated, and the proportion of continuous boundary segments with gradient magnitudes lower than a preset boundary threshold is used as boundary damage indicators. Connection interface evidence is calculated within the pixel area defined by the connection interface micromask. The connection interface evidence includes: the number of connected components and the proportion of the largest connected component area of ​​the connection interface micromask are used as interface continuity indicators; the interface centerline is generated along the main axis of the connection interface micromask, and the mean and variance of the gray-scale gradient change intensity are calculated within a preset width range on both sides of the centerline as interface gradient change indicators; and the mean difference of color and the mean difference of brightness of pixels within the connection interface micromask are used as interface appearance change indicators. Texture evidence, morphological evidence, and connection interface evidence are structurally encapsulated according to "component identifier, viewpoint number, mask identifier, evidence item name, and evidence value" to form a candidate anomaly evidence set.

[0027] This implementation method uses texture, morphology, and interface multi-evidence joint extraction and structured encapsulation to upgrade anomaly judgment from a single score to a verifiable chain of evidence, improving the interpretability and traceability of the results, and facilitating auditing, review, and cross-team handover.

[0028] In this embodiment, step six specifically includes: The candidate abnormal evidence set is grouped according to component identification and viewpoint number. The evidence strength is obtained by interval mapping of the evidence value of each evidence item according to the preset evidence threshold table. The evidence threshold table is a correspondence table between the evidence item name and the normal interval and abnormal interval. When the mask stability index is less than the preset stability threshold or the component mask quality score is less than the preset quality threshold, the evidence strength under the current view is downweighted and marked as a verification state. A consistency check is performed on the connection interface evidence. The consistency check is to compare the strength of the connection interface evidence with the changes in the number of connected components and the maximum area ratio of the connected component in the connection interface micromask in adjacent viewpoints to see if they meet a preset consistency threshold. Based on the evidence strength and consistency verification results after weight reduction, the texture evidence, morphological evidence and connection interface evidence of the same component are weighted and fused to generate anomaly confidence, forming an anomaly candidate list that includes component identifier, viewpoint number, mask identifier, anomaly confidence and evidence item.

[0029] This implementation method achieves reliable calibration and sorting of anomaly confidence through evidence strength mapping, stability and quality score weighted review, and interface consistency verification. This reduces false alarms triggered by occasional noise and improves the rationality of alarm priority, which is conducive to the accurate allocation of operation and maintenance resources.

[0030] In this embodiment, step seven specifically includes: Based on the component identifier and view number in the abnormal candidate list, the corresponding component view is extracted from the inspection image sequence. The component view is an image block obtained by cropping the inspection image with the outer rectangle of the component mask. The component view is overlaid with the component mask and the connection interface micro-mask to generate a mask positioning map, and the evidence item names and evidence values ​​in the anomaly candidate list are structured and organized to form evidence text. The component view, mask positioning map, and evidence text are input into the Qwen2-VL model to generate structured recognition results. The structured recognition results include abnormal components, abnormal phenomena, risk levels, evidence citations, and handling suggestions. The evidence citations are output in the form of fields such as "component identifier, viewpoint number, mask identifier, evidence item name, and evidence value". The handling suggestions include review shooting instructions and maintenance handling instructions. The structured recognition results are associated with and stored with the list of anomaly candidates, and then output.

[0031] This implementation method uses the Qwen2-VL model to perform multimodal joint understanding and generation of component views, mask positioning maps, and evidence text. It can bind numerical evidence with visual positioning and output a structured result of "abnormal component - abnormal phenomenon - risk level - evidence citation - handling suggestion". This avoids the problem of traditional classifiers only providing labels that are difficult to verify, and significantly improves the readability, verifiability, and work order executability of the conclusions.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to the daily intelligent inspection scenario of a 220kV substation in a coastal area. This substation has many exposed metal components, and the alternation of strong sunlight during the day and supplemental lighting at night easily causes specular reflections and localized overexposure at the GIS casing, disconnect switches, lead clamps, and busbar connections. Narrow passageways and obstructions caused by supports and fences result in significant boundary drift of the same component from different viewing angles. The inspection method employs a hybrid approach of handheld cameras and a track-based inspection robot. Within the same inspection cycle, 8–12 visible light images from different perspectives are collected for the same equipment. The image resolution is 4000×3000 pixels, the focal length is fixed at 8.0mm, the exposure time is adaptively adjusted within the range of 1 / 160s–1 / 400s, and the shooting distance is controlled between 1.6m and 2.4m. The equipment ledger provides the types of power equipment, and a pre-set equipment component structure knowledge base provides a list of components and their adjacency and connection relationships. Based on this, component center point prompts and component circumscribed rectangle prompts are generated as the component prompt set input.

[0033] After the image enters the system, the quality and condition characteristic parameters of the inspection image are calculated first. Clarity is calculated as the average grayscale gradient amplitude of the entire image; blurriness is calculated as the average edge expansion width in multiple directions; overexposure ratio is calculated as the percentage of pixels with brightness ≥250; haze index is obtained by combining dark channel intensity and contrast attenuation; reflectivity ratio is obtained by superimposing the area ratio of connected regions of bright pixels with brightness ≥240 and saturation ≤30 with bright edge density; occlusion ratio is obtained by the area ratio of connected regions with edge gradient ≤10 in areas of structural difference between adjacent viewpoints. The reflectivity suppression weight map is generated using weight mapping, with a pixel weight of 0.25 for bright areas and a pixel weight of 1.00 for non-bright areas, thus reducing the weight of reflective areas in the subsequent feature encoding stage. In the improved SAM2 component segmentation, during the image encoding stage, the reflection suppression weight map is bilinearly aligned by scale and then multiplied element-wise with multi-scale features, followed by channel-weighted summation to suppress reflection pseudo-texture. During the cue encoding stage, skeleton cues and boundary cues are used for joint segmentation. The skeleton is used as the query, and the boundary as the key, and multi-head attention fusion is performed to obtain the joint cue embedding vector. Mask decoding outputs the component mask and simultaneously outputs the component mask quality score. The memory construction and propagation module independently gates the writing process for each component. Mask quality scores ≥0.85 are strongly updated with a weight of 1.0, 0.70–0.85 are weakly updated with a weight of 0.5, and <0.70 are not updated with a weight of 0. Simultaneously, a reliable frame index and a memory rollback candidate set are formed, providing a rollback reference for subsequent self-healing.

[0034] Stability indices are calculated and drift types are determined during cross-view propagation and mask self-healing. Area drift is obtained by dividing the absolute value of the difference in the number of mask pixels between two adjacent views by the average number of pixels in both views. Centroid drift is obtained by dividing the distance between the centroid pixels of two adjacent views by the length of the diagonal of the bounding box. Boundary jitter is characterized by the average minimum distance between the two boundaries. Cross-view overlap is characterized by the cross-union ratio. The four indices are fused with equal weights to obtain the mask stability index, with a stability threshold of 0.78. Drift regions are generated and their types are determined when the value is below the threshold. If the overlap ratio between the drift region and the high-weight region of the reflection suppression weight map is ≥0.55, it is identified as reflection adsorption drift. If the overlap ratio with the occlusion region is ≥0.50, it is identified as occlusion drift. After excluding the first two types, if the aspect ratio of the bounding box changes by ≥0.18 and the principal axis changes by ≥12°, it is identified as viewpoint deformation drift. The self-healing strategy is triggered by type: reflective adsorption drift is corrected by prioritizing point filling, occlusion drift is corrected by prioritizing bounding box filling, and viewpoint deformation drift is corrected by prioritizing trusted frame rollback. Point filling error correction selects the pixel with the largest boundary gradient magnitude in the drift area to generate an error correction point prompt. Bounding box error correction uses the bounding box of the union of two adjacent viewpoint masks to generate a constraint box prompt. Trusted frame rollback selects the trusted frame closest to the current viewpoint from the trusted frame index and replaces the current memory feature with its memory feature before propagation.

[0035] The connection interface micro-segmentation generates connection interface bands for component pairs with interconnected relationships, such as "lead clamp - conductor" and "knife switch contact - terminal block". The connection interface band is generated by filtering candidate pixels from the closest distance between the boundary bands of two components (≤6 pixels), followed by connected component filtering (≥300 connected component pixels) and morphological closing operations. Local segmentation only outputs the connection interface probability map within the connection interface band and obtains a connection interface micromask via Sigmoid. The probability outside the interface band is set to zero, thus achieving higher resolution and consistent interface localization within a narrow band. Subsequently, texture, morphology, and interface evidence are extracted from the component mask and the connection interface micromask. An evidence threshold table maps evidence values ​​to evidence strength, and a weighted review is performed using mask stability indicators and mask quality scores. Finally, a consistency check is performed on the connection interface evidence (adjacent view intensity difference ≤0.15, connected component number change ≤2, and maximum area ratio change ≤0.10), updating the anomaly confidence and forming an anomaly candidate list. Finally, the component view, mask positioning map and evidence text are input into the Qwen2-VL model, which outputs abnormal components, abnormal phenomena, risk levels, evidence citations and handling suggestions. The handling suggestions are then mapped into review shooting instructions and maintenance suggestions, which facilitates direct entry into the inspection work order closed loop.

[0036] To verify the improvement effect of this invention compared to existing methods, 12 typical samples from the same week were selected for comparison, and the conclusions were cross-confirmed by two operations and maintenance personnel. Comparison method A uses general object detection + general semantic segmentation single-frame processing without reflection suppression, memory gating, or self-healing; comparison method B uses original SAM2 segmentation and propagation, but does not introduce reflection suppression guidance, skeleton-boundary joint prompts, component gating, or rollback. In Table 1, "Results" are represented by TP / FP / FN, and "Review Time" is the average time taken by operations and maintenance personnel to review and decide whether to dispatch a work order.

[0037] Table 1 Summary of Comparison Results of Typical Samples

[0038] As shown in Table 1, in samples with high reflectivity (S02, S04, S10), comparative method A is more likely to mistake reflective bright edges for abnormal boundaries, leading to component mask expansion or interface positioning drift, resulting in FN or FP. Although comparative method B improves the overall IoU, it still misjudges when strong reflective abrupt changes and local occlusion overlap, and the lack of a rollback mechanism makes it easy for "bad frame" effects to spread across viewing angles. In contrast, this invention maintains a higher IoU on these samples and corrects S04 from FP to TP. The key reason is that the reflectivity suppression weight map suppresses spurious responses in advance during the feature encoding stage, the skeleton-boundary joint cues improve edge-fitting ability, component memory gating avoids the spread of low-quality frame pollution, and the self-healing stage can perform targeted corrections for reflectivity adsorption drift, occlusion drift, and viewing angle deformation drift. In samples with a high occlusion ratio (S07, S08, S12), the probability of FN (Failure to Observe) increases when comparing methods A and B, especially in S08 where the occlusion ratio reaches 12.7%. In single frames or without self-healing propagation, component loss is more likely to occur. This invention improves continuity under occlusion conditions by triggering frame filling and rollback through stability indicators, thus avoiding "more and more errors as propagation progresses".

[0039] Regarding false alarm control, the control sample S06 exhibited false positives (FP) in both comparison methods A and B. On-site verification revealed that reflections and shadows near the terminals were easily mistaken for interface anomalies from a single viewpoint. This invention utilizes the evidence strength and interface consistency verification mapped by the evidence threshold table to downweight and mark occasional single-viewpoint anomaly evidence for verification, thereby obtaining a total non-discharge (TN) and reducing unnecessary work orders. In terms of verification efficiency, the average verification time of this invention decreased from 96 seconds in comparison method A and 79 seconds in comparison method B to 43 seconds. This is not only due to more stable segmentation but also because Qwen2-VL outputs structured results after jointly understanding "component view + mask positioning + evidence values," and clearly identifies the source of the anomaly using the evidence citation field. Maintenance personnel no longer need to repeatedly search for anomalies across multiple images; typically, verifying the location and evidence is sufficient for decision-making. Handling suggestions can also be directly converted into verification shooting requirements or maintenance actions, significantly improving the efficiency of work order closure.

[0040] Based on the data from this embodiment, under real inspection conditions with strong reflection, overexposure, haze, and obstruction, this invention improves segmentation robustness through reflection suppression guidance and joint prompts, enhances cross-view consistency through component memory gating and stability index-driven self-healing, improves the detection rate of interface-type anomalies through constrained local segmentation of connected interfaces, reduces false alarms through evidence strength mapping and consistency verification, and achieves structured output of "citationable evidence, gradable risk, and actionable recommendations" with the help of Qwen2-VL. Overall, this invention improves the reliability, interpretability, and engineering implementation value of inspection results.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent identification system for abnormal states of power equipment based on image recognition, characterized in that, Includes the following modules: The acquisition and suppression module is used to acquire inspection image sequences and record the inspection image quality and working condition characteristic parameters, and generate component prompt set and reflection suppression weight map; The component segmentation module is used to input the inspection image sequence, component cue set and reflection suppression weight map into the improved SAM2 model to generate component masks; The mask self-healing module is used to perform cross-view propagation on the component mask, generate mask stability index and determine the drift type, trigger point correction, box correction or trusted frame rollback to achieve mask self-healing, and obtain the component mask set. The interface micro-segmentation module is used to generate the connection interface band, and together with the inspection image sequence and the reflection suppression weight map, it is input into the improved SAM2 model to perform local segmentation and obtain the connection interface micro-mask. The evidence extraction module is used to extract texture evidence, morphological evidence and connection interface evidence from the component mask set and the micro-mask of the connection interface to form a candidate anomaly evidence set. The filtering and updating module is used to filter and update the confidence level of candidate anomaly evidence sets based on mask stability index, evidence strength of candidate anomaly evidence sets and consistency of connection interface evidence, so as to obtain an anomaly candidate list. The structured recognition module is used to input the component views, mask locations, and evidence values ​​corresponding to the anomaly candidate list into the Qwen2-VL model and output the structured recognition results.

2. The intelligent identification system for abnormal states of power equipment based on image recognition according to claim 1, characterized in that, The modules are connected in the following way: Step 1: Collect inspection image sequences of the same power equipment and record the inspection image quality and operating condition characteristic parameters. Generate a component prompt set based on the power equipment type and component structure prior, and generate a reflection suppression weight map based on the inspection image quality and operating condition characteristic parameters. Step 2: Input the inspection image sequence, component cue set, and reflection suppression weight map into the improved SAM2 model to generate a component mask. The improved SAM2 model includes an image encoding module, a cue encoding module, a memory construction and propagation module, and a mask decoding module. Step 3: Perform cross-view propagation on the component mask, generate mask stability index based on area drift, centroid drift, boundary jitter and cross-view overlap, and determine the drift type. Trigger point correction, box correction or trusted frame rollback to achieve mask self-healing and obtain the component mask set. Step 4: Generate a connection interface band based on the component mask set, and input the inspection image sequence, the connection interface band and the reflection suppression weight map into the improved SAM2 model to perform local segmentation to obtain the connection interface micromask; Step 5: Extract texture evidence, morphological evidence, and connection interface evidence from the component mask set and the micromask at the connection interface to form a candidate anomaly evidence set; Step 6: Based on the mask stability index, the evidence strength of the candidate anomaly evidence set, and the consistency of the connection interface evidence, the candidate anomaly evidence set is screened and the confidence level is updated to obtain the anomaly candidate list; Step 7: Input the component views, mask locations, and evidence values ​​corresponding to the anomaly candidate list into the Qwen2-VL model, and output a structured identification result containing the anomaly component, anomaly phenomenon, risk level, evidence citation, and handling suggestions.

3. The intelligent identification system for abnormal states of power equipment based on image recognition according to claim 2, characterized in that, Step one specifically includes: Collect a sequence of visible light inspection images of the same power equipment within the same inspection task cycle from multiple perspectives, and record the focal length, exposure and shooting distance parameters for each perspective inspection image; The image quality and working condition characteristics parameters of the inspection image are calculated by combining the sharpness parameter, blur parameter, brightness overexposure ratio parameter, haze index parameter, reflectivity ratio parameter and occlusion ratio parameter of the inspection image. The type of power equipment is determined by the equipment ledger or inspection task information. The prior knowledge of the component structure is provided by a pre-set equipment component structure knowledge base, which includes the power equipment type and component list, as well as the adjacency and connection relationships between components. Based on the prior knowledge of power equipment type and component structure, generate a component hint set for each component according to the component list; Based on the reflectivity ratio parameter, the bright areas are weighted and a reflectivity suppression weight map is generated.

4. The intelligent identification system for abnormal states of power equipment based on image recognition according to claim 2, characterized in that, The improved SAM2 model includes an image encoding module, a cue encoding module, a memory construction and propagation module, and a mask decoding module: The image encoding module performs concatenated encoding of convolutional layers and layer normalization layers on the inspection images from each viewpoint in the inspection image sequence, and performs feature transformation through the GELU activation function to form multi-scale image features. A reflection suppression guidance mechanism is introduced, and the reflection suppression weight map is aligned by bilinear interpolation according to the scale and then multiplied element-wise with the feature map of the corresponding scale to complete the weight reduction of the reflection area features. The weighted feature map is then subjected to channel dimension weighted summation to obtain the multi-scale image features after reflection suppression. The prompt encoding module performs embedding encoding on the component prompt set to obtain the component prompt embedding vector. It introduces a joint segmentation mechanism of skeleton prompt and boundary prompt. The point prompt embedding vector in the component prompt embedding vector is defined as the skeleton prompt, and the box prompt embedding vector in the component prompt embedding vector and the component mask boundary band sampling points output by the mask decoding module in the previous round are defined as the boundary prompt. The prompt encoding module uses a multi-head attention layer to interactively fuse the skeleton prompt and the boundary prompt. The multi-head attention layer projects the skeleton prompt through the first linear mapping layer to obtain the query vector, projects the boundary prompt through the second linear mapping layer to obtain the key vector, and projects it through the third linear mapping layer to obtain the value vector. It calculates the attention weight based on the query vector and the key vector and normalizes it through the Softmax function. It uses the attention weight to perform a weighted summation on the value vector to obtain the fused vector, and outputs the joint prompt embedding vector through the feedforward network layer and the residual connection. The mask decoding module uses a cross-attention layer to calculate the attention mapping of the joint cue embedding vector to the multi-scale image features, and performs element-wise addition of the attention mapping result with the multi-scale image features to form fused features. The mask decoding module also outputs the component mask quality score. The memory construction and propagation module performs element-wise multiplication between the component mask and the corresponding multi-scale feature map to obtain the features within the mask, and encodes them into memory features through convolutional layers and layer normalization layers and writes them into the memory queue. A component-specific memory gating mechanism is introduced. For each component, the inspection image quality, working condition feature parameters and component mask quality score are read separately. According to the preset strong update flag, weak update flag and no update flag, the corresponding memory writing weight is selected, and after performing element-wise multiplication with the memory features, it is written into the memory queue. The module outputs a reliable frame index and a memory rollback candidate set. Output the component mask, component mask quality score, memory features, and trusted frame index and memory rollback candidate set corresponding to the inspection image sequence.

5. The intelligent identification system for abnormal states of power equipment based on image recognition according to claim 2, characterized in that, Step three specifically includes: The system receives component masks, component mask quality scores, memory features, trusted frame indexes, and memory rollback candidate sets. It performs cross-view propagation on each component mask based on memory features, using adjacent view inspection images as units. It also determines the propagation update method and whether to trigger mask self-healing based on the component mask quality score. For the same component, the number of mask pixels is counted for each component mask in two adjacent viewpoints to obtain the area drift. The mean value of the pixel coordinates in the mask is calculated to obtain the centroid coordinates and the centroid drift is calculated. The set of pixels at the mask boundary is extracted and the boundary jitter is calculated. The component masks in two adjacent viewpoints are clipped and aligned according to the component's bounding rectangle and the intersection-union ratio is calculated to obtain the cross-viewpoint overlap. The mask stability index is obtained by linearly weighting area drift, centroid drift, boundary jitter and cross-viewpoint overlap according to preset weights, and the drift region is determined based on the mask stability index and the symmetric difference pixel set. The set of pixels with a pixel weight greater than a preset weight threshold in the reflection suppression weight map is defined as a high weight region. When the overlap ratio between the drift region and the high weight region in the reflection suppression weight map is greater than the preset reflection overlap threshold, it is determined to be a reflection adsorption drift. When the overlap ratio between the drift region and the occlusion region is greater than the preset occlusion overlap threshold, it is determined to be an occlusion drift. After excluding reflection adsorption drift and occlusion drift, the viewpoint deformation drift is determined based on the change in the aspect ratio of the bounding rectangle and the change in the principal axis direction. Mask self-healing is triggered based on the drift type, and the mask self-healing includes point correction, bounding box correction, and trusted frame rollback; Output the set of cross-view consistent component masks after mask self-healing and the corresponding mask stability index.

6. The intelligent identification system for abnormal states of power equipment based on image recognition according to claim 2, characterized in that, Step four specifically includes: Based on the consistent set of component masks across different viewpoints, component pairs with interconnected relationships are identified, and the first and second component masks of the component pairs are extracted in each viewpoint. Morphological dilation and erosion are performed on the first component mask and the second component mask respectively, and the difference is taken to obtain the first boundary band and the second boundary band. The candidate pixel set of the connection interface is obtained by filtering based on the nearest distance between the first boundary band and the second boundary band. The connection interface band is generated by connected component filtering and morphological closing operation on the candidate pixel set of the connection interface. The inspection images corresponding to the connecting interface bands in the inspection image sequence, the connecting interface bands and the reflection suppression weight map are input into the improved SAM2 model to perform local segmentation, and the connecting interface micromasks corresponding to each inspection image are output.

7. The intelligent identification system for abnormal states of power equipment based on image recognition according to claim 2, characterized in that, Step five specifically includes: Evidence is extracted for each component mask and its corresponding connection interface micromask using the inspection image as a unit. After removing pixels with weights exceeding a preset weight threshold in the reflection suppression weight map within the pixel area defined by the component mask, texture evidence is calculated. Morphological evidence is calculated within the pixel area defined by the component mask, and connection interface evidence is calculated within the pixel area defined by the connection interface micromask. Texture evidence, morphological evidence, and connection interface evidence are structured and encapsulated according to "component identifier, viewpoint number, mask identifier, evidence item name, and evidence value" to form a candidate anomaly evidence set.

8. The intelligent identification system for abnormal states of power equipment based on image recognition according to claim 2, characterized in that, Step six specifically includes: The candidate abnormal evidence set is grouped according to component identification and viewpoint number, and the evidence strength is obtained by interval mapping of the evidence value of each evidence item according to the preset evidence threshold table. When the mask stability index is less than the preset stability threshold or the component mask quality score is less than the preset quality threshold, the evidence strength under the current view is downweighted and marked as a verification state. A consistency check is performed on the connection interface evidence. The consistency check is to compare the strength of the connection interface evidence with the changes in the number of connected components and the maximum area ratio of the connected component in the connection interface micromask in adjacent viewpoints to see if they meet a preset consistency threshold. Based on the reduced-weighted evidence strength and consistency verification results, the texture evidence, morphological evidence and connection interface evidence of the same component are weighted and fused to generate anomaly confidence and form an anomaly candidate list.

9. The intelligent identification system for abnormal states of power equipment based on image recognition according to claim 2, characterized in that, Step seven specifically includes: And based on the component identifier and viewpoint number in the anomaly candidate list, the corresponding component view is extracted from the inspection image sequence; The component view is overlaid with the component mask and the connection interface micro-mask to generate a mask positioning map, and the evidence item names and evidence values ​​in the anomaly candidate list are structured and organized to form evidence text. The component view, mask positioning map and evidence text are input into the Qwen2-VL model to generate structured recognition results, which include abnormal components, abnormal phenomena, risk levels, evidence citations and handling suggestions. The structured recognition results are associated with and stored with the list of anomaly candidates, and then output.

Images