Power operation risk supervision video identification method and system based on AI technology

Abstract

The invention belongs to the technical field of electric power safety, and discloses an electric power operation risk supervision video identification method and system based on an AI technology, and the method comprises the steps: carrying out the parallel analysis through a double-flow pyramid network architecture, extracting dynamic features and static features in a multi-source video stream, carrying out the feature fusion, and generating different feature axes; identifying dynamic and static risk factors in the feature axis by constructing a hybrid encoder, predicting a cascade evolution path of the risk in combination with a preset risk diffusion simulator, and generating a risk diffusion probability topological graph; and evaluating the generated risk diffusion probability topological graph by using an entropy decision algorithm, calculating the risk level of the power operation scene, triggering a preset hierarchical response mechanism based on the risk level, and performing visual display of the risk level in the video. According to the method, the double-flow pyramid network architecture is adopted to perform parallel extraction and fusion analysis on the dynamic and static features, and the real-time performance and accuracy of risk monitoring can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power safety, and particularly relates to an AI technology-based power operation risk supervision video identification method and system. BACKGROUND

[0002] With the rapid development of the power industry, the complexity of power operation sites continues to increase, and potential safety risks in the operation process are increasingly prominent. Traditional power operation safety management methods mainly rely on manual inspection, experience judgment and post-analysis, which have problems such as response lag, strong subjectivity and insufficient coverage, and have been difficult to meet the requirements of modern power systems for efficient, real-time and intelligent safety supervision.

[0003] In recent years, artificial intelligence (AI) technology, especially computer vision, deep learning and image processing technology, has developed rapidly, providing a new idea for power operation risk supervision. The AI-based automatic monitoring system can efficiently extract key features from massive video data, analyze operation behavior and environmental state in real time, realize intelligent monitoring and risk identification of power operation sites, and greatly improve the real-time, accuracy and automation level of risk supervision.

[0004] However, the existing power operation risk identification method still has limitations, usually only focusing on a single feature (such as personnel behavior or equipment state), and is difficult to fully explore the correlation between multiple features, resulting in incomplete risk identification. At the same time, most AI systems fail to effectively integrate dynamic features (such as personnel actions and behavior changes) and static features (such as environmental and equipment states), making it difficult to accurately depict the complex risk causes and evolution process, and limiting the overall supervision effectiveness of the system.

[0005] Therefore, how to provide an AI technology-based power operation risk supervision video identification method and system is a problem to be solved at present. SUMMARY

[0006] The embodiments of the present application provide an AI technology-based power operation risk supervision video identification method and system to solve the above technical problems in the prior art.

[0007] To have a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This part is not a general review, nor is it intended to determine the key / important components or delineate the protection scope of these embodiments. Its only purpose is to present some concepts in a simple form as a prelude to the detailed description that follows.

[0008] According to a first aspect of the embodiments of the present application, an AI technology-based power operation risk supervision video identification method is provided.

[0009] In one embodiment, the power operation risk supervision video recognition method based on AI technology comprises: Collecting multi-source video streams of the power operation site, performing parallel analysis using a dual-stream pyramid network architecture, extracting dynamic and static features in the multi-source video streams, and performing feature fusion to generate multi-dimensional feature axes; By constructing a hybrid encoder, identifying dynamic and static risk factors, and combining a pre-set risk diffusion simulator to predict the cascade evolution path, a risk diffusion probability topology graph is generated; The generated risk diffusion probability topology graph is evaluated using an entropy value decision algorithm to calculate the risk level of the power operation scene, and a pre-set hierarchical response mechanism is triggered based on the risk level, and the risk level is visually displayed in the video.

[0010] In one embodiment, collecting multi-source video streams of the power operation site, performing parallel analysis using a dual-stream pyramid network architecture, extracting dynamic and static features in the multi-source video streams, and performing feature fusion to generate multi-dimensional feature axes comprises: Through the interface of the video platform, real-time collection of multi-source video streams of the power operation site is performed, and the multi-source video streams are transcoded to generate standardized target video streams; Performing geometric radiometric correction and image enhancement processing on the target video stream based on scene analysis, and using a deep optical flow network to perform motion estimation, and combining a pre-constructed video background model to preliminarily separate dynamic targets and static backgrounds; Combining a frame interrelation algorithm based on Transformer, the preliminary separation results are time-optimized and occlusion-inferred to generate dynamic target sequences and static background sequences; Constructing a dual-stream pyramid network architecture, respectively extracting dynamic and static features from the dynamic target sequences and static background sequences, and fusing the dynamic and static features through a scene-guided attention mechanism; According to the fused static and dynamic features, the feature axes including personnel behavior, equipment state and environmental state are extracted in parallel.

[0011] In one embodiment, performing geometric radiometric correction and image enhancement processing on the target video stream based on scene analysis, and using a deep optical flow network to perform motion estimation, and combining a pre-constructed video background model to preliminarily separate dynamic targets and static backgrounds comprises: Using a semantic segmentation network to perform scene analysis on the video stream of the power operation site, identifying structured regions in the scene, and based on the identification result, using a random sample consensus algorithm to perform geometric correction on each frame of image in the target video stream; The geometric correction result is subjected to radiation correction processing by using a radiation correction algorithm based on a retinal cortex theory, and a quality score of each frame of image is calculated, and a corresponding image enhancement strategy is selected from a preset image enhancement strategy library according to the quality score to perform image enhancement processing; The motion estimation is performed on the continuous video frames by using the deep optical flow network, the optical flow field is generated, and the potential motion area is identified by analyzing the vector change of the optical flow field; According to the identified potential motion area and the pre-constructed video background model, the pixel-level background fusion algorithm is used to aggregate multi-frame background data to establish a static background baseline, and the current frame is compared with the background baseline to preliminarily separate the dynamic target and the static background.

[0012] In one embodiment, the calculation formula of the quality score is: ; In the formula, Q Q represents the quality score; α W represents a weight coefficient; LoG edge E represents an edge sharpness index; β W represents a weight coefficient; SSIM ref S represents a structural similarity index; gamma W represents a weight coefficient; sigma noise V represents a noise variance; delta W represents a weight coefficient; mu grad G represents an average gradient intensity.

[0013] In one embodiment, a dual-stream pyramid network architecture is constructed to extract dynamic features and static features from a dynamic target sequence and a static background sequence respectively, and the dynamic features and the static features are fused by a scene-guided attention mechanism, including: Based on the generated dynamic target sequence and the static background, a dual-stream pyramid network architecture is constructed, and the dual-stream pyramid network architecture includes a spatio-temporal feature stream and a spatial feature stream; In combination with the dual-stream pyramid network architecture and the hierarchical pyramid structure, the spatio-temporal feature stream is used to capture the time sequence change of the dynamic target, and the spatial feature stream is used to extract multi-scale spatial semantic features of the static background, so as to obtain the dynamic features and the static features respectively; According to each level of the pyramid structure, the scene-guided attention mechanism is used to fuse the obtained dynamic features and static features.

[0014] In one embodiment, according to each level of the pyramid structure, the scene-guided attention mechanism is used to fuse the obtained dynamic features and static features, including: Based on each level of the pyramid, the static features extracted at the current level are used as guides to extract the motion information most relevant to the current scene from the dynamic features through an attention mechanism, so as to execute static features to guide dynamic features. The dynamic features, guided by the static features, are used in reverse to optimize the static features. The guided dynamic features are fused with the optimized static features to generate the fused features of the current level, and the fused features are then passed to the next level of the pyramid. The process is passed down layer by layer within the pyramid structure, integrating static and dynamic features at multiple levels until it converges at the top of the pyramid, resulting in the final integrated feature.

[0015] In one embodiment, by constructing a hybrid encoder to identify dynamic and static risk factors along the feature axis, and combining this with a pre-defined risk diffusion simulator to predict the cascading evolution path of risk, a risk diffusion probability topology map is generated, including: Based on the extracted feature axes and the pre-built risk knowledge graph, a hybrid encoder based on graph neural networks and convolutional networks is constructed, and the hybrid encoder is used to update the node states in the risk knowledge graph. Based on the updated risk knowledge graph, the dynamic and static risk factors related to personnel, equipment and environment are identified using the causal attention mechanism. By analyzing the causal relationships between nodes, the corresponding risk entities and their associated paths in the risk knowledge graph are output. Based on the type definitions in the risk knowledge graph, the identified risk entities and associated paths are structured and encoded, and the encoding results are mapped to a unified spatiotemporal semantic coordinate system through affine transformation to generate a set of risk entities. Based on the generated set of risk entities, a risk hypergraph is constructed, and combined with a preset risk diffusion simulator, the cascading evolution path of risks is predicted to generate a risk diffusion probability topology graph.

[0016] In one embodiment, a risk hypergraph is constructed based on the generated set of risk entities, and combined with a preset risk diffusion simulator, the cascading evolution path of risks is predicted to generate a risk diffusion probability topology graph, including: Using the knowledge hypergraph theory, risk entities in the risk entity set are constructed into a risk hypergraph based on the hierarchical structure. By combining a pre-set risk diffusion simulator, the dynamic diffusion process of risk in the risk hypergraph is simulated, and the cascading evolution path is predicted by analyzing the nodes in the risk hypergraph. Based on the predicted cascaded evolution path, a risk coupling tensor is constructed, and the diffusion probability algorithm is used to transform the constructed risk coupling tensor into a probabilistic topological graph.

[0017] In one embodiment, in combination with a preset risk diffusion simulator, a dynamic diffusion process of the risk in the risk hypergraph is simulated, and a cascade evolution path is predicted by analyzing nodes in the risk hypergraph, including: Based on the nodes in the risk hypergraph, a dynamic routing aggregation algorithm is used to initialize the state of each node and assign weights to the hyperedges; According to the initialization result, the activation state of the node is updated by iteration to simulate the dynamic propagation process of the risk in the risk hypergraph, and the risk propagation path between different nodes is simulated according to a preset hierarchical diffusion rule; The risk propagation path and the node feature are fused by using a spatio-temporal graph convolution network, and the cascade evolution path of the risk is predicted by combining spatial convolution and temporal convolution technology.

[0018] According to a second aspect of an embodiment of the present application, an AI technology-based power operation risk supervision video recognition system is provided.

[0019] In one embodiment, the AI technology-based power operation risk supervision video recognition system includes: A feature axis generation module is configured to collect multi-source video streams of a power operation site, perform parallel analysis using a dual-stream pyramid network architecture, extract dynamic features and static features in the multi-source video streams, and perform feature fusion to generate different feature axes; A risk topology graph generation module is configured to identify dynamic and static risk factors in the feature axes by constructing a hybrid encoder, predict a cascade evolution path of the risk in combination with a preset risk diffusion simulator, and generate a risk diffusion probability topology graph; A risk level display module is configured to evaluate the generated risk diffusion probability topology graph using an entropy value decision algorithm, calculate a risk level of the power operation scene, trigger a preset hierarchical response mechanism based on the risk level, and visually display the risk level in the video.

[0020] According to a third aspect of an embodiment of the present application, a computer device is provided.

[0021] In some embodiments, the computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0022] According to a fourth aspect of an embodiment of the present application, a computer readable storage medium is provided.

[0023] In one embodiment, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the above method.

[0024] The technical solutions provided by the embodiments of the present application can include the following beneficial effects: 1、The present application can comprehensively identify various risk factors in the power operation site by collecting multi-source video streams and using a dual-stream pyramid network architecture to extract and fuse dynamic and static features in parallel, thereby improving the real-time performance and accuracy of risk monitoring.

[0025] 2、The present application can predict the cascading evolution path of risks in the power operation scene by combining a preset risk diffusion simulator and generate a risk diffusion probability topology map; further, the present application can comprehensively evaluate multi-source risk information by using an entropy value decision algorithm, calculate the overall risk level of the scene, and trigger the corresponding hierarchical response mechanism accordingly, thereby improving the automation level and intelligent decision-making ability of risk management.

[0026] 3、The present application can visualize the risk level through video pictures, which helps the on-site personnel quickly grasp the risk situation, provides timely support for emergency decision-making, improves the response speed and disposal accuracy, thereby enhancing the operation safety and overall efficiency.

[0027] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application. BRIEF DESCRIPTION OF DRAWINGS

[0028] The accompanying drawings incorporated in the specification hereof and forming a part thereof illustrate embodiments consistent with the present application and together with the description are used to explain the principles of the present application.

[0029] Figure 1 is a flowchart of an AI technology-based power operation risk supervision video recognition method according to an exemplary embodiment; Figure 2 is a principle block diagram of an AI technology-based power operation risk supervision video recognition system according to an exemplary embodiment; Figure 3 is a structural schematic diagram of a computer device according to an exemplary embodiment. DETAILED DESCRIPTION

[0030] The following description and drawings are illustrative of specific embodiments thereof and are not intended to limit the scope of the embodiments. Parts and features of some embodiments can be included in, or alternative to, parts and features of other embodiments. The scope of the embodiments herein encompasses the whole scope of the claims and all available equivalents of the claims. In this document, the terms "first", "second", etc. are used merely to distinguish one element from another, and do not require or imply any actual relationship or order between the elements. In fact, the first element can also be referred to as the second element, and vice versa. Also, the terms "comprises", "comprising", or any other variations thereof are intended to cover a non-exclusive inclusion, such that a structure, device, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such structure, device, or apparatus. Without more limitations, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the structure, device, or apparatus that includes the element. Various embodiments are described in a progressive manner, each focusing on the differences from other embodiments, and the same or similar parts between various embodiments can be referred to each other.

[0031] In this document, the terms "longitudinal", "lateral", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", and the like indicate the orientation or positional relationship shown in the drawings, and are used only for the convenience of description and simplification of the description herein, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In the description herein, unless otherwise specified and limited, the terms "mount", "connect", "connection" should be interpreted broadly, for example, it can be a mechanical connection or an electrical connection, it can be a communication between two elements inside, it can be a direct connection or an indirect connection through an intermediate medium, and the specific meaning of the above terms can be understood by the person skilled in the art according to the specific circumstances.

[0032] In this document, the term "multiple" means two or more, unless otherwise specified.

[0033] In this document, the character " / " represents an "or" relationship between the objects before and after it. For example, A / B means: A or B.

[0034] In this document, the term "and / or" describes the relationship between the objects, which means that there can be three relationships. For example, A and / or B means: A or B, or, A and B, the three relationships.

[0035] It should be understood that although the steps in the flowchart are shown in a sequential order following the arrows, the steps are not necessarily executed in the order shown by the arrows. Unless otherwise explicitly stated herein, there is no strict order requirement for the execution of the steps, and the steps can be executed in other orders. Moreover, at least some of the steps in the figure can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution of the sub-steps or stages can not necessarily be sequential, but can be performed alternately or in rotation with at least part of other steps or sub-steps or stages of other steps.

[0036] Each module in the device or system of the present application can be implemented wholly or partially by software, hardware, and a combination thereof. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operations corresponding to the above modules.

[0037] In the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0038] Figure 1 An embodiment of the AI technology-based power operation risk supervision video recognition method of the present application is shown.

[0039] In this optional embodiment, the AI technology-based power operation risk supervision video recognition method comprises: Step S101, multiple-source video streams of a power operation site are collected, parallel analysis is performed by using a double-stream pyramid network architecture, dynamic features and static features in the multiple-source video streams are extracted, feature fusion is performed, and a multi-dimensional feature axis is generated; Step S102, a hybrid encoder is constructed, dynamic and static risk factors are identified, a preset risk diffusion simulator is combined to predict a cascading evolution path, and a risk diffusion probability topology graph is generated; Step S103, an entropy value decision algorithm is used to evaluate the generated risk diffusion probability topology graph, the risk level of the power operation scene is calculated, a preset hierarchical response mechanism is triggered based on the risk level, and the risk level is visually displayed in the video.

[0040] It should be noted that the evaluation of the generated risk diffusion probability topology graph by using the entropy value decision algorithm, the calculation of the risk level of the power operation scene, the triggering of the preset hierarchical response mechanism based on the risk level, and the visual display of the risk level in the video comprise: First, based on the probability value of each node (representing different risk factors) in the risk diffusion probability topology graph, the information entropy (reflecting the degree of uncertainty) of each risk factor is calculated, and the entropy value calculation formula is: ; In the formula, H j denotes the entropy value; k denotes the standardization constant; p ij denotes the node probability distribution value.

[0041] Next, the difference coefficient of each factor is calculated: ; In the formula, d j denotes the difference coefficient; H j denotes the entropy value; the smaller the entropy value, the greater the difference coefficient, indicating that the factor has a more significant impact on the overall risk.

[0042] Subsequently, the weight is determined by normalizing the difference coefficient: ; In the formula, w j denotes the weight; d j denotes the difference coefficient; a weighted evaluation system is formed; the risk probabilities of all nodes in the topology graph are then weighted and aggregated to generate a comprehensive risk value; according to the preset threshold interval (such as low / medium / high risk), the risk level is divided, and the corresponding hierarchical response mechanism (such as early warning, shutdown or emergency dispatch) is automatically triggered.

[0043] Finally, the risk level and key risk node information are superimposed in real time on the video stream of the power operation site, and visualized through color coding (such as red / yellow / green) or labels, assisting decision-making.

[0044] In this optional embodiment, multi-source video streams of the power operation site are collected, parallel analysis is performed using a dual-stream pyramid network architecture, dynamic and static features in the multi-source video streams are extracted, and feature fusion is performed to generate multi-dimensional feature axes, including: Through the interface of the video platform, multi-source video streams of the power operation site are collected in real time, and the multi-source video streams are transcoded to generate standardized target video streams; The target video stream is subjected to geometric radiometric correction and image enhancement processing based on scene analysis, and motion estimation is performed using a deep light flow network, and a pre-constructed video background model is used to preliminarily separate dynamic targets and static backgrounds; In combination with the inter-frame association algorithm based on Transformer, the preliminary separation result is optimized in time sequence and the occlusion is inferred to generate dynamic target sequence and static background sequence. A dual-stream pyramid network architecture is constructed to extract dynamic features and static features from the dynamic target sequence and the static background sequence respectively, and the dynamic features and the static features are fused through the scene-guided attention mechanism. According to the fused static features and dynamic features, feature axes including personnel behavior, equipment state and environmental state are extracted in parallel.

[0045] It is necessary to supplement that the multi-source video stream of the power operation site is collected in real time through the video platform interface, and the standardized video stream is generated after codec conversion, which helps to ensure the consistency and compatibility of video data, and is convenient for subsequent processing. The geometric radiation correction and image enhancement processing are performed on the target video stream, which can effectively reduce the influence of environmental factors on image quality and improve the clarity and recognizability of the image.

[0046] Motion estimation is performed using a deep optical flow network, and a pre-constructed video background model is used to preliminarily separate dynamic targets and static backgrounds, which helps to accurately extract moving targets and eliminate static interference, improving the accuracy of target detection. Through the inter-frame association algorithm based on Transformer, the separation result is optimized in time sequence and the occlusion is inferred, which can effectively improve the coherence of moving targets, reduce misidentification caused by occlusion, and enhance the stability of target tracking.

[0047] After constructing a dual-stream pyramid network architecture and extracting dynamic and static features, the two are fused through the scene-guided attention mechanism, which can fully utilize the complementarity of dynamic and static information, improve the accuracy and robustness of feature extraction. Finally, based on the fused feature axes, the features of personnel behavior, equipment state and environmental state are extracted, which helps to analyze the safety risks and operating conditions of the power operation site comprehensively and provide more accurate decision support.

[0048] In addition, according to the fused static features and dynamic features, first, a multi-task feature decoder is designed for personnel, equipment and environment, respectively extracting personnel behavior feature axis, equipment state feature axis and environmental state feature axis, and through parallel feature branches, the fused features are decoded; the personnel behavior axis focuses on analyzing human key points, motion trajectories and action patterns, the equipment state axis focuses on detecting the running state, position information and abnormal features of the equipment, and the environmental state axis focuses on extracting global information such as light, weather and site changes in the background environment. Through sharing the bottom layer fusion features and combining their own high-level semantic features, each feature axis realizes parallel extraction and time sequence modeling of multi-dimensional features, providing fine-grained support for subsequent behavior recognition, state monitoring and environment perception.

[0049] In this optional embodiment, scene analysis-based geometric radiometric correction and image enhancement processing are performed on the target video stream, motion estimation is performed using a deep optical flow network, and a pre-constructed video background model is combined to preliminarily separate dynamic targets and static backgrounds, including: A semantic segmentation network is used to perform scene analysis on the video stream of the power operation site, identify structured regions in the scene, and based on the identification results, use a random sample consensus algorithm to perform geometric correction on each frame of image in the target video stream; A radiometric correction algorithm based on the retinal cortex theory is used to perform radiometric correction on the geometric correction results, and the quality score of each frame of image is calculated, and according to the quality score, the corresponding image enhancement strategy is selected from the pre-set image enhancement strategy library to perform image enhancement processing; A deep optical flow network is used to perform motion estimation on consecutive video frames to generate an optical flow field, and by analyzing the vector changes of the optical flow field, potential motion regions are identified; According to the identified potential motion regions and the pre-constructed video background model, a pixel-level background fusion algorithm is used to aggregate multi-frame background data to establish a static background baseline, and the current frame is compared with the background baseline to preliminarily separate dynamic targets and static backgrounds.

[0050] It should be noted that using a semantic segmentation network to perform scene analysis on the video stream of the power operation site and identifying structured regions in the scene helps to accurately understand the video content and facilitate subsequent processing and analysis. By using a random sample consensus algorithm based on the identification results to perform geometric correction on each frame of image, image distortion caused by camera angle or other factors can be effectively removed to ensure the accuracy and consistency of image data. Using a radiometric correction algorithm based on the retinal cortex theory for radiometric correction helps to eliminate the influence of environmental factors (such as light changes) on image quality and improve image quality and reliability.

[0051] At the same time, by calculating the quality score of each frame of image and selecting the most suitable enhancement strategy from the image enhancement strategy library according to the score, the optimal enhancement method can be adopted for images of different quality to improve the visual effect and recognizability of the images. Using a deep optical flow network to perform motion estimation on consecutive video frames and generating an optical flow field can accurately capture the motion trajectory of targets in the scene, providing necessary dynamic information for subsequent target recognition and tracking. By analyzing the vector changes of the optical flow field to identify potential motion regions, possible dynamic targets can be identified in advance to enhance the detection efficiency of dynamic targets.

[0052] Finally, according to the identified potential motion regions and the background model, a pixel-level background fusion algorithm is used to establish a static background baseline and perform differential comparison to effectively separate dynamic targets and static backgrounds, providing clear image data basis for target detection and subsequent analysis.

[0053] In addition, the deep optical flow network is a deep learning-based method for estimating pixel-level motion information (optical flow field) from consecutive video frames, which includes inputting two frames of images, extracting features through a convolutional neural network, and outputting optical flow vectors for each pixel point to describe its motion direction and speed; the network is trained through supervised learning, and a common loss function is used to measure the difference between the predicted optical flow and the real optical flow, which has the advantages of automatic learning, strong robustness and high estimation accuracy; common architectures include FlowNet and PWC-Net, and are widely used in tasks such as motion estimation, dynamic target detection and background modeling in videos.

[0054] In this optional embodiment, the calculation formula of the quality score is: ; In the formula, Q denotes the quality score; α denotes the weight coefficient; LoG edge denotes the edge sharpness index; β denotes the weight coefficient; SSIM ref denotes the structural similarity index; gamma denotes the weight coefficient; sigma noise denotes the noise variance; delta denotes the weight coefficient; mu grad denotes the average gradient intensity.

[0055] It should be noted that in the calculation of the quality score, the edge sharpness index measures the sharpness of the image edge through the Laplacian operator or the gradient amplitude, and the structural similarity index evaluates the perceptual similarity between the image and the reference image according to the brightness, contrast and structural information of the image; the noise variance is used to reflect the intensity of noise interference in the image, which is usually estimated by the variance of the local region; the average gradient intensity reflects the richness and clarity of details through the gradient information of the entire image; these indicators are based on edge detection, structural contrast, noise estimation and gradient analysis, etc. Standard image processing techniques are used to comprehensively evaluate the image quality and serve as the key basis for the calculation of the quality score.

[0056] In this optional embodiment, a dual-stream pyramid network architecture is constructed to extract dynamic features and static features from dynamic target sequences and static background sequences, respectively, and to fuse the dynamic features and static features through a scene-guided attention mechanism, including: Based on the generated dynamic target sequence and static background, a dual-stream pyramid network architecture is constructed, which includes a spatiotemporal feature stream and a spatial feature stream. In combination with the dual-stream pyramid network architecture and the hierarchical pyramid structure, the temporal-spatial feature stream is used to capture the temporal changes of dynamic targets, and the spatial feature stream is used to extract the multi-scale spatial semantic features of static backgrounds, so as to obtain dynamic features and static features respectively. According to each level of the pyramid structure, the scene-guided attention mechanism is used to fuse the obtained dynamic features and static features.

[0057] It should be noted that the dual-stream pyramid network architecture is constructed based on the generated dynamic target sequence and static background, which helps to effectively handle the differences between dynamic and static information, thereby improving the accuracy of target recognition and background modeling; the dual-stream pyramid network architecture is composed of a temporal-spatial feature stream and a spatial feature stream, wherein the temporal-spatial feature stream is used to capture the temporal changes of dynamic targets, which can track the dynamic evolution of targets in the time dimension and enhance the understanding of target motion trajectories; and the spatial feature stream is used to extract the multi-scale spatial semantic features of static backgrounds, which helps to accurately identify static elements in the scene and reduce the interference of static backgrounds.

[0058] In combination with the dual-stream pyramid architecture and the hierarchical pyramid structure, different scale features can be better extracted, thereby providing multi-level information support for the separation and identification of dynamic targets and static backgrounds; the scene-guided attention mechanism is used to fuse the obtained dynamic features and static features, which can improve the fusion effect, so that the model can more flexibly selectively focus on the features of dynamic targets and backgrounds according to the key areas and context information in the scene, and finally realize more accurate target detection and background modeling.

[0059] In addition, the dual-stream pyramid network architecture combines the temporal-spatial feature stream and the spatial feature stream, which are respectively used to capture the temporal changes of dynamic targets and the multi-scale spatial semantic features of static backgrounds in the video; through the pyramid structure, the network extracts information of different scales from coarse to fine, and at each layer, the dynamic and static features are weighted and fused through the scene-guided attention mechanism, to ensure more accurate feature extraction and fusion, which can effectively distinguish dynamic targets from static backgrounds in video analysis, and provide support for subsequent target detection and background modeling.

[0060] In this optional embodiment, according to each level of the pyramid structure, the scene-guided attention mechanism is used to fuse the obtained dynamic features and static features, including: Based on each level of the pyramid, the static features extracted at the current level are used as a guide to extract the most relevant motion information from the dynamic features through the attention mechanism, so as to perform static feature-guided dynamic feature; The dynamic features guided by the static features are reversely applied to the static features, so as to perform dynamic feature-optimized static feature; The guided dynamic features are fused with the optimized static features to generate fusion features of the current level, and the fusion features are transmitted to the previous layer of the pyramid; The static and dynamic features are fused layer by layer in the pyramid structure, and the fusion is performed until the top layer of the pyramid to obtain the final fusion features.

[0061] It should be noted that in the pyramid structure, each layer uses the static features as a guide to extract the most relevant motion information from the dynamic features using the attention mechanism, which realizes the static feature guiding the dynamic feature and improves the scene relevance of the dynamic feature. Then, the guided dynamic features are used to optimize the static features, so that the static features pay more attention to the potential motion area and target boundary. Then, the fusion of the two generates the fusion features of the current level, which are transmitted layer by layer to realize the feature integration at multiple scales. Finally, a set of global fusion features with dynamic sensitivity and static semantics are aggregated at the top of the pyramid to improve the recognition ability of the target and background in complex scenes.

[0062] In this optional embodiment, by constructing a hybrid encoder, dynamic and static risk factors in the feature axis are identified, and a preset risk diffusion simulator is combined to predict the cascading evolution path of the risk to generate a risk diffusion probability topology map, including: According to the extracted feature axis and the pre-constructed risk knowledge graph, a hybrid encoder based on a graph neural network and a convolutional network is constructed, and the hybrid encoder is used to update the node state in the risk knowledge graph; Based on the updated risk knowledge graph, a causal attention mechanism is used to identify dynamic and static risk factors related to personnel, equipment, and environment, and by analyzing the causal relationship between nodes, the corresponding risk entities and associated paths in the risk knowledge graph are output; According to the type definition in the risk knowledge graph, the identified risk entities and associated paths are structured and coded, and the coding results are mapped to a unified spatio-temporal semantic coordinate system through affine transformation to generate a risk entity set; Based on the generated risk entity set, a risk hypergraph is constructed, and a preset risk diffusion simulator is combined to predict the cascading evolution path of the risk to generate a risk diffusion probability topology map.

[0063] It should be noted that by constructing a hybrid encoder based on a graph neural network and a convolutional network according to the extracted feature axis and the pre-constructed risk knowledge graph, the model's ability to represent complex risk knowledge can be effectively combined with graph structure and local features. This process allows the node state in the risk knowledge graph to be updated in real time, ensuring that the risk analysis is always consistent with the latest environment and situation.

[0064] The dynamic and static risk factors related to personnel, equipment and environment are identified by using the causal attention mechanism, so that the causal relationship between the risk factors can be accurately revealed, the risk propagation and evolution process can be better understood, and decision support can be optimized; through the analysis of the causal relationship between nodes, the key risk entities and their associated paths in the risk knowledge graph can be output, so that clear causal chains and important risk nodes are provided for decision-making.

[0065] According to the type definition in the knowledge graph, the risk entities and associated paths are structured and coded, and the results are mapped to a unified space-time semantic coordinate system through affine transformation, which helps to standardize and normalize the information of different risk entities, so that subsequent processing is more efficient; based on the generated risk entity set, a risk hypergraph is constructed, and a preset risk diffusion simulator is combined to predict the cascading evolution path of the risk, which helps to dynamically capture the risk propagation mode, and finally generates a risk diffusion probability topology graph to provide clear visual paths and propagation probabilities, supporting efficient risk management and emergency response.

[0066] In this optional embodiment, based on the generated risk entity set, a risk hypergraph is constructed, and a preset risk diffusion simulator is combined to predict the cascading evolution path of the risk, and a risk diffusion probability topology graph is generated, including: Using knowledge hypergraph theory, the risk entities in the risk entity set are constructed into a risk hypergraph according to the hierarchical structure; Combining a preset risk diffusion simulator, the dynamic diffusion process of the risk in the risk hypergraph is simulated, and the cascading evolution path is predicted by analyzing the nodes in the risk hypergraph; Based on the predicted cascading evolution path, a risk coupling tensor is constructed, and the constructed risk coupling tensor is converted into a probability topology graph using a diffusion probability algorithm.

[0067] In this optional embodiment, combining a preset risk diffusion simulator, the dynamic diffusion process of the risk in the risk hypergraph is simulated, and the cascading evolution path is predicted by analyzing the nodes in the risk hypergraph, including: Based on the nodes in the risk hypergraph, the state of each node is initialized using a dynamic routing aggregation algorithm, and the weight of the hyperedge is assigned; According to the initialization result, the activation state of the node is updated by iteration, the dynamic propagation process of the risk in the risk hypergraph is simulated, and the risk propagation path between different nodes is simulated according to the preset hierarchical diffusion rule; The risk propagation path and the node feature are fused using a space-time graph convolution network, and the cascading evolution path of the risk is predicted by combining spatial convolution and temporal convolution technology.

[0068] Figure 2 An embodiment of the AI technology-based power operation risk supervision video recognition system of the application is shown.

[0069] In this optional embodiment, the power operation risk supervision video recognition system based on AI technology comprises: The feature axis generation module 201 is configured to collect multi-source video streams of a power operation site, perform parallel analysis by using a double-flow pyramid network architecture, extract dynamic features and static features in the multi-source video streams, perform feature fusion, and generate different feature axes. The risk topology graph generation module 202 is configured to identify dynamic and static risk factors in the feature axes by constructing a hybrid encoder, predict a cascading evolution path of the risk by combining a preset risk diffusion simulator, and generate a risk diffusion probability topology graph. The risk level display module 203 is configured to evaluate the generated risk diffusion probability topology graph by using an entropy value decision algorithm, calculate a risk level of the power operation scene, trigger a preset hierarchical response mechanism based on the risk level, and visually display the risk level in the video.

[0070] In one embodiment, a computer device, which can be a server, is provided, and an internal structure diagram of the computer device can be as shown in Figure 3 The computer device comprises a processor, a memory and a network interface connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store static information and dynamic information data. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement the steps in the above method embodiments.

[0071] Those skilled in the art can understand that Figure 3 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. A specific computer device can comprise more or fewer components than those shown in the figure, or some components can be combined, or have a different arrangement of components.

[0072] In addition, the present application also provides a computer device comprising a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0073] In addition, the present application also provides a computer readable storage medium having a computer program stored thereon. The computer program is executed by the processor to implement the steps in the above method embodiments.

[0074] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiment methods. Any reference to memory, storage, database or other medium used in each embodiment of the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0075] The present application is not limited to the structures described above and shown in the drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present application is only limited by the appended claims.

Claims

1. An AI technology-based power operation risk supervision video recognition method, characterized in that, The method comprises: Collecting multi-source video streams of the power operation site, performing parallel analysis by using a double-stream pyramid network architecture, extracting dynamic features and static features in the multi-source video streams, and performing feature fusion to generate different feature axes; By constructing a hybrid encoder, the dynamic and static risk factors in the feature axis are identified, and the cascade evolution path of the risk is predicted by combining a preset risk diffusion simulator to generate a risk diffusion probability topology graph; An entropy value decision algorithm is used to evaluate the generated risk diffusion probability topology graph, calculate the risk level of the power operation scene, trigger a preset hierarchical response mechanism based on the risk level, and visually display the risk level in the video. 2.The AI technology-based power operation risk supervision video recognition method according to claim 1, characterized in that, The collection of multi-source video streams of the power operation site, the parallel analysis by using a double-stream pyramid network architecture, the extraction of dynamic features and static features in the multi-source video streams, and the feature fusion to generate different feature axes comprise: Through the interface of the video platform, the multi-source video streams of the power operation site are collected in real time, and the multi-source video streams are transcoded to generate standardized target video streams; Performing geometric radiometric correction and image enhancement processing based on scene analysis on the target video stream, and using a deep optical flow network to perform motion estimation, and combining a pre-constructed video background model to preliminarily separate dynamic targets and static backgrounds; Combining a frame interrelation algorithm based on Transformer, the preliminary separation result is time-optimized and occlusion-inferred to generate dynamic target sequences and static background sequences; A double-stream pyramid network architecture is constructed to extract dynamic features and static features from the dynamic target sequences and static background sequences respectively, and the dynamic features and static features are fused through a scene-guided attention mechanism; According to the fused static features and dynamic features, the feature axes including personnel behavior, equipment state and environment state are extracted in parallel. 3.The AI technology-based power operation risk supervision video recognition method according to claim 2, characterized in that, The performing of geometric radiometric correction and image enhancement processing based on scene analysis on the target video stream, and the using of a deep optical flow network to perform motion estimation, and the combining of a pre-constructed video background model to preliminarily separate dynamic targets and static backgrounds comprise: Using a semantic segmentation network to perform scene analysis on the video stream of the power operation site, identifying the structured areas in the scene, and based on the identification result, using a random sample consensus algorithm to perform geometric correction on each frame of image in the target video stream; Using a radiometric correction algorithm based on the theory of retinal cortex to perform radiometric correction processing on the geometric correction result, and calculating the quality score of each frame of image, and selecting the corresponding image enhancement strategy from a preset image enhancement strategy library according to the quality score to perform image enhancement processing; Using a deep optical flow network to perform motion estimation on consecutive video frames to generate an optical flow field, and identifying potential motion areas by analyzing the vector changes of the optical flow field; According to the identified potential motion areas and the pre-constructed video background model, pixel-level background fusion algorithm is used to aggregate multi-frame background data to establish a static background baseline, and the current frame is compared with the background baseline to preliminarily separate dynamic targets and static backgrounds. 4.The AI technology-based power operation risk supervision video recognition method according to claim 3, characterized in that, The calculation formula of the quality score is: ； wherein, Q denotes a quality score; α denotes a weight coefficient; LoG edge denotes an edge sharpness indicator; β denotes a weight coefficient; SSIM ref denotes a structural similarity index; γ denotes a weight coefficient; σ noise denotes a noise variance; δ denotes a weight coefficient; μ grad denotes an average gradient strength. 5.The AI technology-based power operation risk supervision video recognition method according to claim 4, characterized in that, The double-stream pyramid network architecture is constructed to extract dynamic features and static features from the dynamic target sequence and the static background sequence respectively, and the dynamic features and the static features are fused through the scene-guided attention mechanism, which includes: Based on the generated dynamic target sequence and static background, a double-stream pyramid network architecture is constructed, which includes a spatio-temporal feature stream and a spatial feature stream. By combining the double-stream pyramid network architecture and the hierarchical pyramid structure, the spatio-temporal feature stream is used to capture the temporal changes of the dynamic target, and the spatial feature stream is used to extract the multi-scale spatial semantic features of the static background, so as to obtain dynamic features and static features respectively. According to each level of the pyramid structure, the scene-guided attention mechanism is used to fuse the obtained dynamic features and static features. 6.The AI technology-based power operation risk supervision video recognition method according to claim 5, characterized in that, The scene-guided attention mechanism is used to fuse the obtained dynamic features and static features according to each level of the pyramid structure, which includes: Based on each level of the pyramid, the static features extracted at the current level are used as a guide to extract the most relevant motion information from the dynamic features through the attention mechanism to guide the dynamic features with the static features. The dynamic features guided by the static features are used to optimize the static features. The guided dynamic features and the optimized static features are fused to generate fusion features at the current level, and the fusion features are passed to the upper pyramid. The fusion of static and dynamic features at multiple levels is performed through layer-by-layer transmission in the pyramid structure until it is aggregated to the top of the pyramid to obtain the final fusion features. 7.The AI technology-based power operation risk monitoring video recognition method of claim 1, wherein, The hybrid encoder is constructed to identify dynamic and static risk factors in the feature axis, and the preset risk diffusion simulator is used to predict the cascading evolution path of the risk to generate a risk diffusion probability topology map, which includes: Based on the extracted feature axis and the pre-constructed risk knowledge graph, a hybrid encoder based on graph neural network and convolutional network is constructed, and the hybrid encoder is used to update the node state in the risk knowledge graph. Based on the updated risk knowledge graph, the causal attention mechanism is used to identify dynamic and static risk factors related to personnel, equipment and environment, and the causal relationship between nodes is analyzed to output the corresponding risk entities and associated paths in the risk knowledge graph. According to the type definition in the risk knowledge graph, the identified risk entities and associated paths are structured and coded, and the coding results are mapped to a unified spatio-temporal semantic coordinate system through affine transformation to generate a risk entity set. Based on the generated risk entity set, a risk hypergraph is constructed, and a risk diffusion probability topology map is generated by combining the preset risk diffusion simulator to predict the cascading evolution path of the risk. 8.The AI technology-based power operation risk supervision video recognition method according to claim 7, characterized in that, The risk hypergraph is constructed based on the generated risk entity set, and the risk diffusion probability topology map is generated by combining the preset risk diffusion simulator to predict the cascading evolution path of the risk, which includes: According to the hierarchical structure, the risk entities in the risk entity set are constructed into a risk hypergraph using the knowledge hypergraph theory. In combination with a preset risk diffusion simulator, a dynamic diffusion process of the risk in the risk hypergraph is simulated, and a cascading evolution path is predicted by analyzing nodes in the risk hypergraph; Based on the predicted cascading evolution path, a risk coupling tensor is constructed, and the constructed risk coupling tensor is converted into a probability topology graph by using a diffusion probability algorithm. 9.The AI technology-based power operation risk supervision video recognition method according to claim 8, characterized in that, The combination of the preset risk diffusion simulator simulates the dynamic diffusion process of the risk in the risk hypergraph, and predicts the cascading evolution path by analyzing the nodes in the risk hypergraph, which includes: Based on the nodes in the risk hypergraph, the state of each node is initialized by using a dynamic routing aggregation algorithm, and the weight of the hyperedge is assigned; According to the initialization result, the activation state of the node is updated by iteration, the dynamic propagation process of the risk in the risk hypergraph is simulated, and the risk propagation path between different nodes is simulated according to the preset hierarchical diffusion rule; The risk propagation path and the node feature are fused by using a spatio-temporal graph convolution network, and the cascading evolution path of the risk is predicted by combining spatial convolution and temporal convolution technology.

10. An AI technology-based power operation risk supervision video recognition system, characterized in that, The system includes: A feature axis generation module is configured to collect multi-source video streams of a power operation site, perform parallel analysis by using a dual-stream pyramid network architecture, extract dynamic features and static features in the multi-source video streams, perform feature fusion, and generate different feature axes; A risk topology graph generation module is configured to identify dynamic and static risk factors in the feature axes by constructing a hybrid encoder, predict a cascading evolution path of the risk in combination with a preset risk diffusion simulator, and generate a risk diffusion probability topology graph; A risk level display module is configured to evaluate the generated risk diffusion probability topology graph by using an entropy value decision algorithm, calculate a risk level of the power operation scene, trigger a preset hierarchical response mechanism based on the risk level, and visually display the risk level in a video.

Images