Power monitoring system and method integrating image recognition and data analysis

Abstract

The invention relates to the field of electric power monitoring, and discloses an electric power monitoring system and method fusing image recognition and data analysis, and the method comprises the steps: carrying out the visual angle coverage modeling of a target equipment group through a multi-type visual collection unit disposed at a transformer substation and a power distribution terminal; performing cross-frame fine-grained texture differential analysis on the equipment state image sequence, and constructing an image event time window in combination with synchronous disturbance characteristics of multi-source monitoring parameters; based on the high-vigilance candidate frame set, fusing the image structure variability index and the operation data multi-dimensional deviation vector by using a feature encoder, and constructing a multi-modal state coupling feature tensor; map mapping is carried out on the potential fault evolution trend, and semantic association is established between structural nodes with abnormal attributes in the image and frequently fluctuating parameter indexes in the monitoring data; and combining a node interference path in the local fault association subgraph with fault precursor distribution induced in a historical accident sample. The method has the advantage that the operation safety is improved.

Description

Technical Field

[0001] The present invention relates to the field of power monitoring, and in particular to a power monitoring system and method integrating image recognition and data analysis. Background Art

[0002] With the development of intelligent power systems, key facilities such as substations and distribution rooms are increasingly deploying image monitoring equipment to visualize equipment status. However, existing image monitoring systems rely primarily on manual inspections or simple video stream recording, failing to automatically identify and analyze equipment anomalies. While data monitoring systems can provide real-time parameters, they lack the ability to perceive the equipment's external status. These two systems operate independently and lack effective collaboration, leading to information silos and decision delays during equipment anomaly warnings and fault tracing. For example, existing systems for monitoring the operation of low-voltage distribution cabinets often use thermoelectric sensors to detect temperature rise within the cabinet, but are unable to identify safety hazards caused by visual issues such as impact, doors not properly closed, and aging cables. Even when video cameras are deployed on-site, they lack the ability to structure image information, making it difficult to intelligently capture anomalies and provide real-time feedback. Therefore, it is imperative to design a power monitoring system and method that integrates image recognition and data analysis to enhance operational safety. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention provides a power monitoring system and method that integrates image recognition and data analysis, which has the advantage of improving operational safety and solves the problems in the above-mentioned background technology.

[0004] To achieve the above-mentioned purpose of improving operational safety, the present invention provides the following technical solution: a power monitoring method integrating image recognition and data analysis, comprising the following steps: By deploying multiple types of visual acquisition units at substations and distribution terminals, we model the view coverage of the target equipment group, extract high-risk image areas such as disconnector contacts, busbar hot spots, and cable connections, and generate a sequence of equipment status images. Perform cross-frame fine-grained texture differential analysis on the device status image sequence, combine the synchronous disturbance characteristics of multi-source monitoring parameters, construct image event time windows, and mark image frames with occlusion as a set of high-alert candidate frames; Based on a set of high-alert candidate frames, a feature encoder is used to fuse image structure variability indicators with multi-dimensional deviation vectors of operating data to construct a multimodal state coupling feature tensor. Attention weights are applied to the mutation boundary areas in the tensor to derive potential fault evolution trends. Perform graph mapping on the evolution trend of potential faults, establish semantic associations between structural nodes with abnormal attributes in the image and parameter indicators with frequent fluctuations in the monitoring data, and form a local fault correlation subgraph centered on the abnormal source; Combining the node interference paths in the local fault correlation subgraph with the fault precursor distribution summarized in historical accident samples, a set of early warning response strategies for the target area is generated.

[0005] Preferably, the process of generating a device status image sequence is: Deploy multiple types of visual acquisition units within the target power facilities, including high-resolution industrial cameras, thermal infrared imaging modules, and low-light environment-adapted lenses, to capture equipment structural appearance, temperature rise distribution, and dark zone operating status. By integrating a time synchronization module to perform timestamp calibration and frame alignment for each visual data source, combined with an adaptive illumination correction algorithm and image dynamic enhancement mechanism, image distortion or blurred areas caused by environmental disturbances are eliminated. Based on the equipment inventory and layout topology, regional template matching and target recognition are performed on key parts in the image, such as disconnectors, circuit breakers, busbar connectors, and cable outlets. The processed image sequences are grouped and reconstructed according to device type and time sequence to construct a device status image sequence with device hierarchical structure and time continuity.

[0006] Preferably, the process of marking image frames with occlusion as a set of high-alert candidate frames is as follows: Multi-scale grayscale difference and texture consistency comparison are performed on adjacent frames in the image sequence, and the structural change rate and edge distortion amplitude of the key image area are extracted using an adaptive sliding window. Combining temperature and humidity, voltage fluctuations, and short-term current surges collected from environmental sensors and operational data, we construct the synchronization disturbance period for the corresponding frame, and use occlusion patches, blurred boundaries, and high-frequency noise in the image frame as uncertainty markers. Using the occlusion scoring mechanism, the image integrity weight and resolvability index of each frame are calculated, and frames with significant image distortion are marked as high-alert candidate frames based on the set threshold; Finally, the high-alert candidate frames within the coverage of multiple sensors are summarized to form a set of candidate image frames with spatial redundancy and temporal local aggregation characteristics.

[0007] Preferably, the process of constructing the multimodal state coupling characteristic tensor is: Perform feature point matching, edge contour extraction, and texture histogram construction on the key structure areas extracted from the high-alertness candidate frames to generate an image structure variability index vector; The operating data at the corresponding moment, including bus current, cable temperature rise, ground resistance, voltage fluctuation rate and other parameters, are normalized and mapped into a state deviation vector; A deep feature fusion module is used to perform multi-channel convolution coupling processing on the image structure variability index and the operating state deviation vector to form a multimodal state coupling feature tensor.

[0008] Preferably, the process of deriving the potential fault evolution trend is: Perform time series modeling analysis on the constructed multi-modal state coupling feature tensor, and use an improved time series clustering algorithm to identify the state migration pattern of the tensor in a continuous time period; Extract local time windows in the state mutation zone, track high-weight change paths, and locate high-risk candidate trend lines with continuous evolution characteristics; Construct a state evolution path diagram, mark the starting point, transition point and abnormal convergence point of the fault trend line, and compare and analyze it with the standard operating state trajectory set; If there are multiple change channels in the trend line that are greater than the characteristic deviation rate threshold, it is determined to be a potential fault evolution trend; If there are multiple change channels in the trend line that are less than or equal to the characteristic deviation rate threshold, it is determined to be a non-potential fault evolution trend.

[0009] Preferably, the process of forming a local fault correlation subgraph centered on the abnormal source is: Extract key image areas and measurement point locations where structural mutations or frequent parameter fluctuations occur from the potential fault evolution trend and define them as abnormal nodes; Propagate upstream and downstream information about the device hierarchy where the abnormal node is located, and track its adjacent components, power lines, and grounding links in the topological network; Combining the electrical system connection relationship with the equipment operation and maintenance records, a multi-dimensional directed edge set between nodes is constructed to form a preliminary logical path diagram between devices; A semantic extension mechanism is introduced to physically bind the physical anomaly attributes reflected in the image with the feature offset indicators shown in the data, giving each node a multimodal fault label. Finally, a local fault correlation subgraph with the abnormal source as the center of the graph is generated.

[0010] Preferably, the process of generating an early warning response strategy set for a target area is: Perform structural feature extraction and propagation path analysis on the local fault-related subgraph to identify core nodes with critical connectivity and abnormal density points; Compare the subgraph structure with a pre-built library of typical power failure patterns to identify historical samples with the highest similarity to the current subgraph in terms of topology, label distribution, and propagation trends. Based on the verified effective emergency response records in the matching samples, a strategy sequence including isolation actions, load transfer, operating parameter reduction and manual re-inspection is extracted; A real-time scoring and equipment risk level evaluation mechanism is introduced into the strategy sequence to calculate the response lag risk and coverage adaptation rate of each strategy in the current scenario, and finally generate an early warning response strategy set.

[0011] A power monitoring system integrating image recognition and data analysis, comprising: Perception modeling module: Deploys and coordinates multiple types of image acquisition equipment to extract high-risk image areas of key power components and construct a complete image sequence of equipment status; Image analysis module: performs cross-frame texture difference analysis and disturbance data fusion to identify occlusion and structural changes in the image, and screens high-alert image frames to form anomaly detection windows; Feature fusion module: This module jointly encodes image structural variability and multi-source operational deviation data to generate a unified state coupling tensor and mine potential fault evolution trends. Graph construction module: maps potential abnormal trends to the graph structure between structural nodes and monitoring indicators, and constructs a local fault correlation subgraph around the abnormal source; Strategy generation module: Based on the subgraph structure and historical sample knowledge base, it outputs a warning response strategy set containing fault intervention suggestions.

[0012] Compared with the existing technology, the present invention provides a power monitoring system and method that integrates image recognition and data analysis, which has the following beneficial effects: The present invention achieves comprehensive perception and accurate diagnosis of the status of power equipment by integrating image recognition technology with multi-source data analysis, significantly improving the intelligence level of the power monitoring system. By utilizing the perspective coverage modeling of multiple types of visual acquisition units, it can efficiently capture structural and temperature anomaly information of key equipment parts, ensuring the multi-dimensionality and high timeliness of monitoring data. Cross-frame fine-grained texture differential analysis combined with operating parameter perturbations enhances the detection capability of subtle anomalies and occlusions, effectively reducing missed reports and false alarms. The construction of multimodal state coupling feature tensors and their attention weight mechanism improves the accuracy of identifying potential fault evolution trends and strengthens the capture of early fault signals. The construction of local fault correlation subgraphs based on graph mapping achieves semantic fusion between equipment abnormal nodes and multi-dimensional monitoring parameters, providing an intuitive and reliable analysis basis for fault root cause location. Combined with the fault precursor distribution of historical accident samples, a targeted early warning response strategy set is intelligently generated, improving the pertinence and effectiveness of fault response, promoting the automation and refined management of power system risk prevention and control, greatly reducing equipment failure risks and operation and maintenance costs, and ensuring the safe, stable, and continuous reliability of power grid operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 Schematic diagram of the method of the present invention; Figure 2 Schematic diagram of the system of the present invention. DETAILED DESCRIPTION

[0014] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0015] Example 1: Please refer to Figure 1 As shown, an electric power monitoring method integrating image recognition and data analysis according to an embodiment of the present invention includes the following steps: S1: By deploying multiple types of visual acquisition units at substations and distribution terminals, we model the view coverage of the target equipment group, extract high-risk image areas such as disconnector contacts, busbar hot spots, and cable connections, and generate a sequence of equipment status images.

[0016] The process of generating the device status image sequence in S1 is as follows: Deploy multiple types of visual acquisition units within the target power facilities, including high-resolution industrial cameras, thermal infrared imaging modules, and low-light environment-adapted lenses, to capture equipment structural appearance, temperature rise distribution, and dark zone operating status. By integrating a time synchronization module to perform timestamp calibration and frame alignment for each visual data source, combined with an adaptive illumination correction algorithm and image dynamic enhancement mechanism, image distortion or blurred areas caused by environmental disturbances are eliminated. Based on the equipment inventory and layout topology, regional template matching and target recognition are performed on key parts in the image, such as disconnectors, circuit breakers, busbar connectors, and cable outlets. The processed image sequences are grouped and reconstructed according to device type and time sequence to construct a device status image sequence with device hierarchical structure and time continuity.

[0017] Deploy multiple types of visual acquisition units in key areas of substations and distribution terminals, including high-resolution industrial cameras, thermal infrared imaging modules, and near-infrared lenses suitable for low-light environments, to capture the structural appearance, temperature rise distribution, and operating status images of equipment at night or in low-light conditions, ensuring multi-angle and multi-time visual coverage. Through the integrated time synchronization module, unified timestamp calibration and inter-frame alignment are performed on various types of image data sources. A synchronization mechanism based on NTP or GPS is combined with an image frame interpolation algorithm to achieve consistent image timing processing. At the same time, an adaptive illumination correction algorithm and an image dynamic enhancement mechanism are used to correct images affected by ambient light disturbances. Adjust the image brightness and restore edge details; after obtaining high-quality image data, build a regional template library containing targets such as disconnector contacts, circuit breaker contacts, busbar connectors and cable outlets based on the asset ledger and on-site layout topology information of the power equipment, and use the deep learning-based image recognition model to perform regional matching and target recognition on key parts of the image; structure the recognition results, group and reorder the image frames according to the equipment type, spatial position and acquisition time, eliminate redundant frames and blurred frames, and generate an equipment status image sequence with a clear equipment hierarchy and time continuity, laying a data foundation for subsequent texture analysis and anomaly detection.

[0018] S2: Perform cross-frame fine-grained texture differential analysis on the device status image sequence, combine the synchronous disturbance characteristics of multi-source monitoring parameters, construct image event time windows, and mark image frames with occlusion as a set of high-alert candidate frames.

[0019] The process of marking the image frames with occlusion as the high-alert candidate frame set in S2 is as follows: Multi-scale grayscale difference and texture consistency comparison are performed on adjacent frames in the image sequence, and the structural change rate and edge distortion amplitude of the key image area are extracted using an adaptive sliding window. Combining temperature and humidity, voltage fluctuations, and short-term current surges collected from environmental sensors and operational data, we construct the synchronization disturbance period for the corresponding frame, and use occlusion patches, blurred boundaries, and high-frequency noise in the image frame as uncertainty markers. Using the occlusion scoring mechanism, the image integrity weight and resolvability index of each frame are calculated, and frames with significant image distortion are marked as high-alert candidate frames based on the set threshold; Finally, the high-alert candidate frames within the coverage of multiple sensors are summarized to form a candidate frame set with spatial redundancy and temporal local aggregation characteristics.

[0020] Adjacent image frames in the equipment status image sequence are compared pair by pair, and a multi-scale grayscale difference algorithm is used to detect the image brightness change characteristics. Combined with the texture consistency measurement method based on LBP or Gabor filtering, the structural contour change rate and edge clarity in the key area are quantitatively evaluated. At the same time, an adaptive sliding window mechanism is introduced in the image frame to dynamically extract abnormal fragments that may be occluded, blurred or structurally dislocated. The above image analysis results are synchronously aligned with the data from the environmental sensors and the power operation system, and combined with disturbance factors such as drastic changes in temperature and humidity, short-term voltage fluctuations, and current shocks, A disturbance window at the image frame level is constructed, and local occlusion patches, edge blurred areas, and noise clusters in the image within the window are marked with uncertainty. An image occlusion scoring mechanism is used to assign image integrity weights and resolvability indicators to each frame, and an empirical threshold is set to classify the scoring results. Any image frame with a score below the threshold is marked as a high-alert candidate frame. The marked high-alert frames in the coverage area of ​​multiple acquisition terminals are aggregated by position and time, and duplicate records are eliminated to construct a set of high-alert candidate frames with spatial distribution redundancy and time series aggregation for subsequent feature fusion and fault trend extraction.

[0021] S3: Based on the high-alert candidate frame set, the feature encoder is used to fuse the image structure variability index and the multi-dimensional deviation vector of the operating data to construct a multimodal state coupling feature tensor. Attention weights are applied to the mutation boundary areas in the tensor to derive the potential fault evolution trend.

[0022] The process of constructing the multi-modal state coupling characteristic tensor in S3 is: Perform feature point matching, edge contour extraction, and texture histogram construction on the key structure areas extracted from the high-alertness candidate frames to generate an image structure variability index vector; The operating data at the corresponding moment, including bus current, cable temperature rise, ground resistance, voltage fluctuation rate and other parameters, are normalized and mapped into a state deviation vector; A deep feature fusion module is used to perform multi-channel convolution coupling processing on the image structure variability index and the operating state deviation vector to form a multimodal state coupling feature tensor.

[0023] For image frames marked as high-alert candidates, a variety of structural feature information at the image level is extracted from the key structural areas identified (such as disconnector contacts, busbar connection terminals, cable joints, etc.), including feature point matching through SIFT or ORB algorithm, edge contour extraction based on Canny or Sobel operator, and texture histogram statistical features generated by gray-level co-occurrence matrix or local texture coding method, thereby forming an image structure variability index vector that characterizes the structural change trend; secondly, the equipment operation monitoring data corresponding to the above image frame timestamp is obtained, including high-frequency sampling parameters such as busbar current, cable temperature rise, grounding resistance, and voltage fluctuation rate, and z -score normalization or min-max standardization method are used to uniformly map them to the same numerical scale, and a structured operating state deviation vector is constructed according to the physical category to which the parameters belong, retaining their physical correlation and mutation characteristics; on this basis, a deep feature fusion module is introduced, and a multi-channel convolutional neural network is used to perform feature-level coupling processing on the image structure variability index and the operating state deviation vector. The response weight of the key feature dimension is improved through the channel attention mechanism, and the fusion results are spliced ​​at the convolution output end, finally forming a multimodal state coupling feature tensor containing image structure, operating parameters and their interactive relationship, providing a unified data foundation for subsequent potential fault trend modeling and atlas mapping.

[0024] The potential fault evolution trend process introduced in S3 is as follows: Perform time series modeling analysis on the constructed multi-modal state coupling feature tensor, and use an improved time series clustering algorithm to identify the state migration pattern of the tensor in a continuous time period; Extract local time windows in the state mutation zone, track high-weight change paths, and locate high-risk candidate trend lines with continuous evolution characteristics; Construct a state evolution path diagram, mark the starting point, transition point and abnormal convergence point of the fault trend line, and compare and analyze it with the standard operating state trajectory set; If there are multiple change channels in the trend line that are greater than the characteristic deviation rate threshold, it is determined to be a potential fault evolution trend; If there are multiple change channels in the trend line that are less than or equal to the characteristic deviation rate threshold, it is determined to be a non-potential fault evolution trend.

[0025] The constructed multimodal state coupling feature tensor is modeled and analyzed in continuous time periods, and an improved time series clustering algorithm (such as the K-Shape clustering method based on DTW distance optimization) is introduced to classify and aggregate the characteristic response curves of the tensor in multiple time segments, and identify the state migration trajectory with obvious dynamic change patterns; for the segments with sudden changes or morphological transitions in the clustering results, the corresponding local time windows are extracted, and the feature channel attention mechanism is used to enhance the feature dimensions with high change frequency and large change amplitude, and a change channel tracking matrix is ​​constructed. The paths with temporal continuity and significant change gradients are tracked frame by frame, and high-risk trend line candidates with significant cumulative effects and staged transition characteristics are located; on this basis, a state evolution path diagram corresponding to the trend line is constructed. The starting state point, transition turning point and abnormal convergence point of each trend line are marked, and the structural similarity of the trend lines is compared using typical samples in the standard operating state trajectory set to identify the degree of deviation in the spatial and temporal dimensions. If there are two or more continuous change channels in the trend line that are greater than the set threshold of the feature deviation rate (for example, the image texture change rate exceeds 30%, the temperature change rate is greater than 5°C / min, the current fluctuation frequency is abnormally enhanced, etc.), the system determines that the trend is a "potential fault evolution trend" and needs to enter the next stage of the graph mapping and early warning strategy generation process. On the contrary, if there are only short-term disturbances in the change channel and the overall deviation rate does not exceed the threshold, it is regarded as a non-fault trend and automatically archived by the system and excluded from the risk response process, thereby achieving accurate capture of early signs of faults and controlling the false alarm rate.

[0026] S4: Graph mapping is performed on the evolution trend of potential faults. Structural nodes with abnormal attributes in the image are semantically associated with parameter indicators that frequently fluctuate in the monitoring data to form a local fault association subgraph centered on the abnormal source.

[0027] The process of forming a local fault correlation subgraph centered on the abnormal source in S4 is as follows: Extract key image areas and measurement point locations where structural mutations or frequent parameter fluctuations occur from the potential fault evolution trend and define them as abnormal nodes; Propagate upstream and downstream information about the device hierarchy where the abnormal node is located, and track its adjacent components, power lines, and grounding links in the topological network; Combining the electrical system connection relationship with the equipment operation and maintenance records, a multi-dimensional directed edge set between nodes is constructed to form a preliminary logical path diagram between devices; A semantic extension mechanism is introduced to physically bind the physical anomaly attributes reflected in the image with the feature offset indicators shown in the data, giving each node a multimodal fault label. Finally, a local fault correlation subgraph with the abnormal source as the center of the graph is generated.

[0028] Key image regions and corresponding monitoring point locations showing obvious structural mutations or frequent parameter fluctuations are screened from potential fault evolution trends, and these regions and monitoring points are identified as candidate abnormal nodes. Based on the hierarchical structure of power equipment, upstream and downstream information diffusion is performed on abnormal nodes. The system automatically tracks their neighboring components in the power grid topology, including connected power lines, busbars, circuit breakers, and grounding links, to form a preliminary scope of associated paths. Combining the electrical system's connectivity and equipment operation and maintenance history, a multidimensional set of directed edges between nodes is constructed, reflecting the electrical logic paths and possible fault propagation paths between devices, forming a preliminary device logic path diagram. A semantic extension mechanism is introduced to physically bind the physical anomaly attributes reflected in image recognition (such as local hot spots and signs of structural damage) with characteristic offset indicators (such as sudden temperature changes and abnormal current fluctuations) shown in the monitoring data. This allows each abnormal node to be assigned a multimodal fault label, improving the richness of node representation and diagnostic accuracy. Finally, this information is integrated to generate a local fault association subgraph centered on the anomaly source, covering key equipment nodes and their electrical connection paths, providing accurate structural support for fault location and early warning response.

[0029] S5: Combine the node interference paths in the local fault correlation subgraph with the fault precursor distribution summarized in the historical accident samples to generate a set of early warning response strategies for the target area.

[0030] The process of generating the early warning response strategy set for the target area in S5 is as follows: Perform structural feature extraction and propagation path analysis on the local fault-related subgraph to identify core nodes with critical connectivity and abnormal density points; Compare the subgraph structure with a pre-built library of typical power failure patterns to identify historical samples with the highest similarity to the current subgraph in terms of topology, label distribution, and propagation trends. Based on the verified effective emergency response records in the matching samples, a strategy sequence including isolation actions, load transfer, operating parameter reduction and manual re-inspection is extracted; A real-time scoring and equipment risk level evaluation mechanism is introduced into the strategy sequence to calculate the response lag risk and coverage adaptation rate of each strategy in the current scenario, and finally generate an early warning response strategy set.

[0031] Based on the local fault-related subgraph, a graph structure feature extraction algorithm is used to identify key connected nodes and abnormally dense areas. The importance of each node in the fault propagation path is evaluated through indicators such as node degree and betweenness centrality, and a fault propagation impact matrix is ​​constructed. The subgraph structure is matched with samples in a pre-established typical power fault pattern library for multi-dimensional similarity. Combined with the topological structure similarity, node label distribution consistency and fault propagation trend consistency, the most representative historical fault cases are selected as reference templates. Based on the selected historical samples, the emergency response strategy sequence contained therein is extracted, which covers the disconnector action, Multi-level measures such as load redistribution, operating parameter adjustment, and manual inspection and review ensure the comprehensiveness and pertinence of the strategy; in the strategy generation stage, the system further introduces a real-time scoring mechanism, dynamically assessing the response lag risk of each strategy based on the current equipment operating status and risk level, while considering the scope of application and coverage effect of the strategy, and quantifying the adaptability of each solution in the current environment; based on the above comprehensive evaluation results, the system automatically optimizes and adjusts the strategy priority and execution order, and generates a set of early warning response strategies that meet the needs of actual operating scenarios, have high timeliness and risk prevention and control capabilities, and provide a scientific basis and operational guidance for operation and maintenance decisions.

[0032] Example 2: Figure 2 As shown, a power monitoring system integrating image recognition and data analysis includes: Perception modeling module: Deploys and coordinates multiple types of image acquisition equipment to extract high-risk image areas of key power components and construct a complete image sequence of equipment status; Image analysis module: performs cross-frame texture difference analysis and disturbance data fusion to identify occlusion and structural changes in the image, and screens high-alert image frames to form anomaly detection windows; Feature fusion module: This module jointly encodes image structural variability and multi-source operational deviation data to generate a unified state coupling tensor and mine potential fault evolution trends. Graph construction module: maps potential abnormal trends to the graph structure between structural nodes and monitoring indicators, and constructs a local fault correlation subgraph around the abnormal source; Strategy generation module: Based on the subgraph structure and historical sample knowledge base, it outputs a warning response strategy set containing fault intervention suggestions.

[0033] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0034] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A power monitoring method integrating image recognition and data analysis, characterized in that: The following steps are involved: By deploying multiple types of visual acquisition units at substations and distribution terminals, we model the view coverage of the target equipment group, extract high-risk image areas such as disconnector contacts, busbar hot spots, and cable connections, and generate a sequence of equipment status images. Perform cross-frame fine-grained texture differential analysis on the device status image sequence, combine the synchronous disturbance characteristics of multi-source monitoring parameters, construct image event time windows, and mark image frames with occlusion as a set of high-alert candidate frames; Based on a set of high-alert candidate frames, a feature encoder is used to fuse image structure variability indicators with multi-dimensional deviation vectors of operating data to construct a multimodal state coupling feature tensor. Attention weights are applied to the mutation boundary areas in the tensor to derive potential fault evolution trends. Perform graph mapping on the evolution trend of potential faults, establish semantic associations between structural nodes with abnormal attributes in the image and parameter indicators with frequent fluctuations in the monitoring data, and form a local fault correlation subgraph centered on the abnormal source; Combining the node interference paths in the local fault correlation subgraph with the fault precursor distribution summarized in historical accident samples, a set of early warning response strategies for the target area is generated.

2. The power monitoring method integrating image recognition and data analysis according to claim 1, characterized in that: The process of generating a device status image sequence is as follows: Deploy multiple types of visual acquisition units within the target power facilities, including high-resolution industrial cameras, thermal infrared imaging modules, and low-light environment-adapted lenses, to capture equipment structural appearance, temperature rise distribution, and dark zone operating status. By integrating a time synchronization module to perform timestamp calibration and frame alignment for each visual data source, combined with an adaptive illumination correction algorithm and image dynamic enhancement mechanism, image distortion or blurred areas caused by environmental disturbances are eliminated. Based on the equipment inventory and layout topology, regional template matching and target recognition are performed on key parts in the image, such as disconnectors, circuit breakers, busbar connectors, and cable outlets. The processed image sequences are grouped and reconstructed according to device type and time sequence to construct a device status image sequence with device hierarchical structure and time continuity.

3. The power monitoring method integrating image recognition and data analysis according to claim 2, characterized in that: The process of marking image frames with occlusion as a set of high-alert candidate frames is as follows: Multi-scale grayscale difference and texture consistency comparison are performed on adjacent frames in the image sequence, and the structural change rate and edge distortion amplitude of the key image area are extracted using an adaptive sliding window. Combining temperature and humidity, voltage fluctuations, and short-term current surges collected from environmental sensors and operational data, we construct the synchronization disturbance period for the corresponding frame, and use occlusion patches, blurred boundaries, and high-frequency noise in the image frame as uncertainty markers. Using the occlusion scoring mechanism, the image integrity weight and resolvability index of each frame are calculated, and frames with significant image distortion are marked as high-alert candidate frames based on the set threshold; Finally, the high-alert candidate frames within the coverage of multiple sensors are summarized to form a set of candidate image frames with spatial redundancy and temporal local aggregation characteristics.

4. The power monitoring method integrating image recognition and data analysis according to claim 3, characterized in that: The process of constructing the multi-modal state coupling feature tensor is: Perform feature point matching, edge contour extraction, and texture histogram construction on the key structure areas extracted from the high-alertness candidate frames to generate an image structure variability index vector; The operating data at the corresponding moment, including bus current, cable temperature rise, ground resistance, voltage fluctuation rate and other parameters, are normalized and mapped into a state deviation vector; A deep feature fusion module is used to perform multi-channel convolution coupling processing on the image structure variability index and the operating state deviation vector to form a multimodal state coupling feature tensor.

5. The power monitoring method integrating image recognition and data analysis according to claim 4, characterized in that: The process of deriving the potential fault evolution trend is as follows: Perform time series modeling analysis on the constructed multi-modal state coupling feature tensor, and use an improved time series clustering algorithm to identify the state migration pattern of the tensor in a continuous time period; Extract local time windows in the state mutation zone, track high-weight change paths, and locate high-risk candidate trend lines with continuous evolution characteristics; Construct a state evolution path diagram, mark the starting point, transition point and abnormal convergence point of the fault trend line, and compare and analyze it with the standard operating state trajectory set; If there are multiple change channels in the trend line that are greater than the characteristic deviation rate threshold, it is determined to be a potential fault evolution trend; If there are multiple change channels in the trend line that are less than or equal to the characteristic deviation rate threshold, it is determined to be a non-potential fault evolution trend.

6. The power monitoring method integrating image recognition and data analysis according to claim 5, characterized in that: The process of forming a local fault correlation subgraph centered on the abnormal source is as follows: Extract key image areas and measurement point locations where structural mutations or frequent parameter fluctuations occur from the potential fault evolution trend and define them as abnormal nodes; Propagate upstream and downstream information about the device hierarchy where the abnormal node is located, and track its adjacent components, power lines, and grounding links in the topological network; Combining the electrical system connection relationship with the equipment operation and maintenance records, a multi-dimensional directed edge set between nodes is constructed to form a preliminary logical path diagram between devices; A semantic extension mechanism is introduced to physically bind the physical anomaly attributes reflected in the image with the feature offset indicators shown in the data, giving each node a multimodal fault label. Finally, a local fault correlation subgraph with the abnormal source as the center of the graph is generated.

7. The power monitoring method integrating image recognition and data analysis according to claim 6, characterized in that: The process of generating an early warning response strategy set for the target area is as follows: Perform structural feature extraction and propagation path analysis on the local fault-related subgraph to identify core nodes with critical connectivity and abnormal density points; Compare the subgraph structure with a pre-built library of typical power failure patterns to identify historical samples with the highest similarity to the current subgraph in terms of topology, label distribution, and propagation trends. Based on the verified effective emergency response records in the matching samples, a strategy sequence including isolation actions, load transfer, operating parameter reduction and manual re-inspection is extracted; A real-time scoring and equipment risk level evaluation mechanism is introduced into the strategy sequence to calculate the response lag risk and coverage adaptation rate of each strategy in the current scenario, and finally generate an early warning response strategy set.

8. A power monitoring system integrating image recognition and data analysis, applied to the method according to any one of claims 1 to 7, characterized in that: include: Perception modeling module: Deploys and coordinates multiple types of image acquisition equipment to extract high-risk image areas of key power components and construct a complete equipment status image sequence; Image analysis module: performs cross-frame texture difference analysis and disturbance data fusion to identify occlusion and structural changes in the image, and screens high-alert image frames to form anomaly detection windows; Feature fusion module: This module jointly encodes image structural variability and multi-source operational deviation data to generate a unified state coupling tensor and mine potential fault evolution trends. Graph construction module: maps potential abnormal trends to the graph structure between structural nodes and monitoring indicators, and constructs a local fault correlation subgraph around the abnormal source; Strategy generation module: Based on the subgraph structure and historical sample knowledge base, it outputs a warning response strategy set containing fault intervention suggestions.

Images