Power grid safety supervision method and system

Abstract

The invention provides a power grid safety supervision method and system, and relates to the technical field of information, and the method comprises the steps: carrying out the grade classification and hidden danger rating of hidden dangers according to a preliminary hidden danger analysis result, a construction site business rule, and the severity of the hidden dangers, and obtaining a hidden danger grade classification result; hidden danger grades are judged according to the hidden danger grade classification result, the hidden danger grades are divided into grade I (major hidden danger), grade II (larger hidden danger), grade III (serious hidden danger) and grade IV (general hidden danger), and a corresponding response instruction sequence is issued to a construction equipment controller based on the hidden danger grades; generating a summary report of power grid supervision according to the fused multi-modal feature vectors; and the generated standardized construction report is stored in a database and pushed to related management personnel, so that the management personnel can know the construction site and the power grid state in time. The risk of safety accidents is reduced, and powerful support is provided for construction safety management.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a power grid security supervision method and system. Background Art

[0002] In a complex and ever-changing construction site, how to identify various construction targets in real time and accurately, and analyze potential safety hazards in a timely manner, is a technical problem that needs to be solved urgently. The traditional manual inspection method is inefficient, difficult to fully cover the entire construction site, and is easily affected by subjective factors, resulting in inaccurate and untimely identification of hidden dangers. Simply using machine vision technology for target detection and identification is difficult to adapt to the dynamic changes and complex environment of the construction site, and is prone to problems such as missed detection and misjudgment. In addition, how to quickly and accurately analyze key information from massive construction site data and take corresponding preventive and emergency measures according to the level of hidden dangers is also a complex technical challenge. Relying solely on manual experience and judgment, it is difficult to fully consider various factors, and the response speed is slow, making it difficult to meet real-time requirements. Simply using methods such as rule engines or expert systems is difficult to cope with the variability and uncertainty of the construction site, and is prone to problems such as improper decision-making and untimely response. In short, how to achieve all-round, real-time and accurate identification and response of safety hazards in a complex and ever-changing construction site is a key technical bottleneck that needs to be broken through. Summary of the invention

[0003] In order to solve the above technical problems, the present invention provides a power grid security supervision method, which mainly includes:

[0004] According to the construction operation scenario, the construction business data is obtained, the camera equipment is called to collect the image data of the construction site, and the collected image data is subjected to real-time target detection and recognition to obtain the real-time target detection and recognition results;

[0005] Based on the real-time target detection and recognition results, a knowledge graph of the construction scene is formed, and the knowledge graph is used as the basis for reasoning to generate preliminary hidden danger analysis results;

[0006] Based on the preliminary hidden danger analysis results and in combination with the construction site business rules, the hidden dangers are classified and rated according to their severity to obtain the hidden danger level classification results;

[0007] The hidden danger level is determined according to the hidden danger level classification results. The hidden danger level is divided into level I, level II,

[0008] Level III and Level IV: issuing corresponding response instruction sequences to the construction equipment controller based on the hazard level;

[0009] Obtain a variety of on-site safety and supervision results data, including law enforcement device data, camera equipment data and sensor data, and combine real-time processing records to perform feature extraction and fusion, integrate data from different sources, and generate a fused multi-modal feature vector;

[0010] Generate a summary report of power grid supervision based on the fused multi-modal feature vector. The summary report includes the problems existing at the construction site, the level of hidden dangers, and emergency response measures.

[0011] According to the summary report of the power grid supervision, a text generation method based on natural language processing is used to generate a standardized construction report that meets the construction site management requirements. The construction report includes basic information of the construction site, existing problems, hidden danger levels and emergency response measures;

[0012] The generated standardized construction reports are stored in the database and pushed to relevant managers so that they can understand the status of the construction site and power grid in a timely manner.

[0013] The present invention provides a power grid security monitoring system, which mainly includes:

[0014] The data acquisition and real-time target detection module is used to obtain construction business data according to the construction operation scene, call the camera equipment to collect the image data of the construction site, perform real-time target detection and recognition on the collected image data, and obtain the real-time target detection and recognition results;

[0015] The preliminary hidden danger analysis module is used to form a knowledge graph of the construction scene based on the real-time target detection and recognition results, and use the knowledge graph as the basis for reasoning to generate preliminary hidden danger analysis results;

[0016] The hidden danger level classification module is used to classify and rate hidden dangers according to the severity of hidden dangers based on the preliminary hidden danger analysis results and the construction site business rules, and obtain hidden danger level classification results;

[0017] A response instruction generation and execution module is used to determine the hidden danger level according to the hidden danger level classification result. The hidden danger levels are divided into level I, level II, level III, and level IV, and send a corresponding response instruction sequence to the construction equipment controller based on the hidden danger level;

[0018] The multimodal data fusion module is used to obtain a variety of on-site safety and supervision results data, including law enforcement instrument data, camera equipment data and sensor data, and combine real-time processing records to perform feature extraction and fusion, integrate data from different sources, and generate a fused multimodal feature vector;

[0019] The supervision summary report generation module is used to generate a summary report of power grid supervision based on the fused multi-modal feature vector. The summary report includes the problems existing at the construction site, the level of hidden dangers, and the emergency response measures.

[0020] The standardized construction report generation module is used to generate a standardized construction report that meets the construction site management requirements based on the summary report of the power grid supervision and the text generation method based on natural language processing. The construction report content includes basic information of the construction site, existing problems, hidden danger levels and emergency response measures;

[0021] The report storage and push module is used to store the generated standardized construction reports in the database and push them to relevant managers so that they can understand the construction site and power grid status in a timely manner.

[0022] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0023] The present invention discloses a power grid safety supervision method and system. The method collects multi-source data from the construction site, uses target detection technology to identify safety hazards in real time, and combines knowledge graphs and agent models for intelligent analysis and reasoning. The present invention constructs a construction scene knowledge graph, which contains business rules, historical hidden danger data and countermeasures, and provides a knowledge basis for intelligent analysis. Through multimodal data fusion and feature extraction, combined with knowledge graph reasoning, automatic identification, classification and response of hidden dangers are achieved. For hidden dangers of different levels, the present invention can automatically trigger corresponding treatment measures, such as notifying a shutdown or prompting correction. In addition, the present invention can also automatically generate standardized construction reports and push them to management personnel. This intelligent construction safety supervision method significantly improves the accuracy and timeliness of hidden danger identification, realizes intelligent management of construction sites, effectively reduces the risk of safety accidents, and provides strong support for construction safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 The figure is a flow chart of the power grid security supervision method of the present invention. DETAILED DESCRIPTION

[0025] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. The present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the relevant inventions, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0026] like Figure 1 The power grid security supervision method of this embodiment may specifically include:

[0027] S101. According to the construction operation scene, construction business data is obtained, a camera device is called to collect image data of the construction site, and the collected image data is subjected to real-time target detection and recognition to obtain a real-time target detection and recognition result.

[0028] A positioning calibration rod is used to obtain the reference coordinates of the construction site. The reference coordinates are used to establish a calibration coordinate system for the camera equipment. The collected image data is distorted and corrected through the calibration coordinate system to obtain standard image data. A residual neural network is used to extract construction target features based on the standard image data. The three-dimensional position data of the construction target space is calculated from the position coordinate data extracted from the target features and the position coordinate data of the preset calibration points. If the three-dimensional position data of the construction target space exceeds the pre-established construction area boundary coordinate range, an alarm is triggered and the motion trajectory data of the construction target is collected. The motion trajectory data of the construction target is smoothed by a Kalman filter, and the smoothed motion trajectory data is classified and identified by a support vector machine to obtain a construction target operation state determination result.

[0029] Exemplarily, the reference coordinates are collected based on the positioning calibration poles arranged at the construction site, and the cameras are arranged at multiple fixed positions at the construction site to collect images. The camera collects the images and establishes a calibration coordinate system according to the reference coordinates of the calibration poles. When the light intensity is less than 500 lux, the fill light is turned on, and the position and attitude parameters of the camera are calculated by positioning the calibration poles and the reference coordinates collected by the camera. The original image data is subjected to distortion correction and noise removal to obtain the standard image data of the construction site. Based on the standard image data of the construction site, the target features of the construction workers, mechanical equipment, and environmental obstructions in the image are extracted through the residual neural network, and the targets associated with the construction work are screened out from the target feature data. The position coordinate data is extracted from the target, and the position coordinate data is compared and calculated with the position coordinate data of the pre-set calibration points to obtain the three-dimensional position of the construction work target space. By comparing the three-dimensional position data of the construction target space with the pre-established construction area boundary coordinate data, the boundary range of the construction area is determined. If the construction target position coordinates are within the boundary range, the construction target position data is recorded; if the construction target position coordinates exceed the boundary range, an alarm of exceeding the boundary range is triggered and the construction target is tracked and photographed, and the motion trajectory data of the construction target is collected, and the target displacement, speed, and acceleration change parameters are extracted from the motion trajectory data. The motion trajectory data of the construction target is smoothed by a Kalman filter to remove the motion trajectory noise data, and the motion characteristics of the construction target are extracted from the smoothed motion trajectory data. The support vector machine is used to classify and identify the motion characteristic data of the construction target, and it is determined whether the construction target is in a normal construction state, and the construction target operation state determination result is stored. The deployment of calibration poles for positioning construction scenes is crucial to the establishment of camera equipment acquisition benchmarks. The height of the calibration poles ranges from 3 meters to 6 meters. The top and bottom of the calibration poles are engraved with standard scale marking points, with a scale marking point interval of 0.5 meters. The calibration poles are made of aluminum alloy and coated with anti-rust paint. Four calibration poles are deployed in areas with good vision in the construction scene to form a rectangular area. LED beacon lights are installed on the top of the calibration poles, with a flashing frequency of 1 Hz at night, which is convenient for image acquisition at night or in low light environments. The construction site camera equipment is deployed at a fixed position point with a deployment height of 5 meters to 8 meters. The camera equipment uses a high-definition network camera with a resolution of 1920x1080 pixels, a frame rate of 25 frames per second, a field of view of 120 degrees, and an automatic aperture adjustment range of F1.6 to F3.5. The camera equipment uses infrared fill light, and the fill light distance can reach 50 meters. The fill light is automatically turned on when the light intensity is lower than 500 lux.The standard image data processing of the construction site adopts distortion correction and noise removal. Distortion correction is mainly aimed at lens distortion, including radial distortion and tangential distortion. The radial distortion correction coefficients are k1=-0.32, k2=0.12, and the tangential distortion correction coefficients are p1=0.005, p2=0.003. Gaussian filtering is used for noise removal, with a filter kernel size of 5x5 pixels and a standard deviation of 1.5. The image clarity is improved by 30% after processing. The target feature extraction adopts residual neural network for identification, and the construction workers, mechanical equipment, and environmental obstructions are classified. The construction workers wear reflective vests and helmets. The mechanical equipment includes excavators, loaders, and transport vehicles. The environmental obstructions include temporary buildings, stacked materials, and construction fences. The position accuracy of the target detection frame is within plus or minus 5 pixels. The boundary of the construction area is delineated by pre-established coordinate data, and boundary points are set around the construction area. The boundary point interval is 10 meters. The coordinates of the boundary points use the geodetic coordinate system. The coordinates of the construction target position within the boundary range and the boundary points form a polygonal area to determine whether the construction target is within the polygonal area. The motion trajectory of the construction target is collected and recorded to record the change of the target position over time. The sampling interval is 0.5 seconds. The motion trajectory data includes the target displacement, velocity, and acceleration parameters. The displacement accuracy is at the centimeter level. The velocity calculation uses displacement difference, and the acceleration calculation uses velocity difference. The motion trajectory smoothing uses Kalman filter to remove tracking jitter caused by occlusion or illumination changes. The state of the construction target is determined by support vector machine classification and recognition. The training samples include normal construction operations and abnormal operation states. The features include motion trajectory features, target posture features, and operation area features. The classification accuracy reaches 90%, and the results of the construction target state determination are recorded and stored in real time.

[0030] S102. Based on the real-time target detection and recognition results, a knowledge graph of the construction scene is formed, and the knowledge graph is used as the basis for reasoning to generate preliminary hidden danger analysis results.

[0031] Acquire real-time target detection results at the construction site, wherein the target detection results include construction equipment location coordinate data and construction personnel clothing tag data; construct knowledge graph nodes in a triplet manner, wherein the knowledge graph nodes generate equipment nodes and personnel nodes according to the target detection results, and the equipment nodes and personnel nodes are associated according to the mapping relationship of the rule base; calculate the semantic association weights between nodes using a graph attention network according to the knowledge graph nodes, wherein the semantic association weights generate an action sequence graph structure based on the spatial distance between nodes and the degree of temporal association; trigger the Agent perception module to perform path reasoning through the action sequence graph structure, wherein the path reasoning is based on the comparison of node semantic association with the risk level threshold to obtain preliminary hidden danger analysis results.

[0032] Exemplarily, based on the location coordinates of construction equipment and the clothing information of construction personnel in the real-time target detection and recognition results of the construction site, the knowledge graph nodes are established in combination with the basic business rule base of the construction site. The construction equipment nodes contain the attributes of location distance and placement direction, and the construction personnel nodes contain the attributes of safety helmet wearing mark and reflective clothing wearing mark. The mapping relationship between the rule nodes and the actual scene nodes is recorded in a triple storage method. The rule nodes contain a combination of preconditions and trigger actions, and the subject-object relationship is extracted from the real-time detection data to fill the node attribute data. From the historical hidden danger data records of the construction site, a convolutional neural network is used to extract the hidden danger feature association matrix between equipment, personnel, and operating behaviors. The equipment feature dimension in the hidden danger feature association matrix includes location coordinates, floor space, and mobile status, and the personnel feature dimension includes clothing completeness, operating area coordinates, and operating posture. The hidden danger scene knowledge graph nodes are established and the connection between nodes is constructed according to the spatial distance and the correlation between the human and equipment operating status. The response and disposal nodes are established for the historical hidden danger scene response and disposal methods and the hidden danger scene nodes are associated. The graph attention network is used to construct the semantic association between nodes of the knowledge graph of the construction site. The association weight is calculated according to the spatial distance and temporal association degree between nodes. The action sequence diagram is constructed for the nodes of construction equipment, construction personnel, and construction operation behavior. The rule constraint condition nodes and hidden danger scene response nodes are added to the graph structure. The construction scene data is obtained from the real-time target detection and recognition results to update the knowledge graph node status, and dynamic decision rules are established according to the node status changes. The Agent perception module is triggered by the construction scene node status in the knowledge graph. According to the node semantic association, path reasoning is performed under the rule constraint conditions to judge the distance between construction equipment and safety and the standardization of personnel dress. The construction scene risk status judgment result is obtained from the graph structure. The historical hidden danger scene response node is associated with the risk status. The risk level threshold is compared with the real-time detection data of the construction site to obtain the preliminary hidden danger analysis results. During the construction of the knowledge graph nodes at the construction site, the attributes of the construction equipment nodes include three dimensions: equipment type, spatial position, and operating status. The equipment types include excavators, loaders, and transport vehicles. The spatial position is represented by three-dimensional coordinates. The equipment operating status is divided into three types: static, mobile, and operating. The relationship between nodes is stored in the form of triples. The subject-predicate-object structure describes the association between nodes. For example, when the excavator is less than 3 meters away from the safety passage, the equipment placement is not standardized. The attributes of the construction personnel node include three dimensions: personnel identification, clothing status, and operating behavior. The clothing status includes three items: safety helmets, reflective clothing, and protective shoes. The operating behavior includes basic actions such as standing, walking, and bending. A spatial distance relationship is established between the personnel node and the equipment node. When the distance between the personnel node and the equipment node is less than 5 meters, the human-machine operation association judgment is triggered.The feature extraction of historical hidden danger data adopts convolutional neural network. The input data includes image features and position features. The image feature dimension is 256x256 pixels, including three RGB channels. The position feature dimension is 6, including spatial coordinates and motion parameters. The network structure includes 5 convolution layers, each with a convolution kernel size of 3x3. The pooling layer adopts maximum pooling, with a pooling kernel size of 2x2. The fully connected layer outputs a 1024-dimensional feature vector. The semantic association of knowledge graph nodes is constructed using graph attention network. The node feature vector dimension is 128, the number of attention heads is 8, and each attention head calculates the association weight between nodes. The weight calculation considers the spatial distance attenuation factor. The weight attenuation is 0 when the distance exceeds 10 meters. The time series association adopts exponential attenuation with an attenuation coefficient of 0.8. The risk status of the construction scene is determined based on the combination of node states. The risk determination rules for irregular equipment placement include the distance threshold from the safe channel and the boundary threshold of the mechanical motion range. The risk determination rules for personnel clothing violations include the confidence threshold of 0.9 for helmet wearing recognition and the confidence threshold of 0.85 for reflective clothing wearing recognition. The Agent perception module adopts a multi-layer perceptron structure, with an input layer dimension of 256, corresponding to the knowledge graph node state vector, a hidden layer dimension of 512, and an output layer dimension of 16, corresponding to different risk level categories. The risk level is divided into 4 levels, and each level is further divided into 4 subcategories. The risk level judgment threshold is determined according to the statistical distribution of historical hidden danger data. The risk warning judgment result includes three parts: risk level, risk description, and response measures. The risk description is generated using the corresponding hidden danger scenario node attributes in the knowledge graph. The response measures are selected from the historical response and disposal nodes according to the similarity sorting. The similarity calculation is based on the cosine distance of the node feature vector. The three response measure nodes with the highest similarity are selected to generate the warning response suggestion.

[0033] S103. Based on the preliminary hidden danger analysis results and in combination with the construction site business rules, the hidden dangers are classified and rated according to their severity to obtain hidden danger grade classification results.

[0034] According to the knowledge graph, construction rule nodes are retrieved to obtain quantitative indicator data, and a hidden danger determination matrix is ​​constructed through the quantitative indicator data. The hidden danger determination matrix calculates a rule violation quantification score; historical hidden danger record data is obtained, and a deep neural network is used to extract features of the historical hidden danger record data to obtain a hidden danger feature vector, and the hidden danger feature vector generates a rule violation quantification score through a feature extractor; the rule violation quantification score sequence is input into a long short-term memory network predictor, and the predictor calculates a state transition probability matrix according to a scene state migration rule to obtain a hidden danger prediction score; if the sum of the rule violation quantification score and the hidden danger prediction score is greater than a first threshold, it is determined to be a red risk and an emergency response rule is executed; if the sum of the scores is less than the first threshold and greater than the second threshold, it is determined to be an orange risk and an early warning response rule is executed.

[0035] Exemplarily, based on the preliminary hidden danger analysis results, the construction rule node is retrieved through the knowledge graph agent to obtain the quantitative index data of the construction site, including the safety spacing value of construction equipment, the spacing value of safety protection facilities, and the numerical value of personnel clothing requirements, and a hidden danger judgment matrix is ​​constructed. The row vectors of the matrix represent the measured values ​​of each quantitative indicator, and the column vectors represent the standard values ​​of the corresponding indicators. The difference between the measured value and the standard value is calculated to generate the quantitative score of the first rule violation. The historical hidden danger record data is extracted from the knowledge graph, and the hidden danger features are extracted using a deep neural network. The hidden danger feature vector dimension is set to 256 dimensions, including three sub-feature blocks of equipment type, number of personnel, and scene environment. Each sub-feature block is input into the corresponding feature extractor to generate the quantitative score of the second rule violation. The hidden danger scoring standard is constructed in combination with the historical hidden danger level labels, and the standard score range is divided into four scoring ranges according to the historical statistical distribution. The quantified scores of the first rule violation and the quantified scores of the second rule violation are averaged to obtain the average of the quantified scores of the first rule violation and the second rule violation. A long short-term memory network is used to construct a construction scene state sequence predictor. The quantified scores of the first rule violation and the average of the quantified scores of the second rule violation are input. The number of hidden layer neurons is set to 512, and the output layer dimensions correspond to four scoring intervals, namely level I (major hidden dangers), level II (large hidden dangers), level III (serious hidden dangers), and level IV (general hidden dangers). The state transition probability matrix is ​​calculated through the scene state migration rules in the hidden danger scene knowledge graph, and the hidden danger evolution trend is predicted to generate a hidden danger prediction score. According to the rule violation quantitative score and hidden danger prediction score, the Agent classifies the hidden danger according to the scoring standard. If the hidden danger score is greater than 0.8, it is determined as a red risk, corresponding to level I (major hidden danger), and the emergency response rule is executed; if the hidden danger score is between 0.5 and 0.8, it is determined as an orange risk, corresponding to level II (large hidden danger), and the early warning response rule is executed; if the hidden danger score is between 0.3 and 0.5, it is determined as a yellow risk, corresponding to level III (serious hidden danger), and the rectification response rule is executed; if the hidden danger score is less than 0.3, it is determined as a blue risk, corresponding to level IV (general hidden danger), and the monitoring response rule is executed. The construction rule quantitative index is constructed in the form of a multidimensional matrix. Taking the safety spacing of construction equipment as an example, the standard value of the spacing between the excavator operation area and the safety passage is 5 meters. The actual measured value is obtained by image recognition. When the actual measured value is less than the standard value, the violation score is calculated according to the difference ratio. When the actual measured value is 3 meters, the violation score is 0.4, and when the actual measured value is 2 meters, the violation score is 0.6. The standard spacing standards for safety protection facilities include the distance between the construction area and the fence, the distance between safety warning signs, and the distance between safety protection railings. The standard spacing between the fence and the construction area is 3 meters, the standard spacing between safety warning signs is 10 meters, and the maximum spacing between protection railings is 1.5 meters. The actual measured values ​​are compared with the standard values ​​to generate a facility deployment violation score.The personnel dress requirements include the wearing of helmets, reflective clothing, and protective shoes. The integrity of personnel dress is judged by image recognition. The confidence threshold for helmet wearing is 0.9, the confidence threshold for reflective clothing is 0.85, and the confidence threshold for protective shoes is 0.8. If it is lower than the threshold, it is recorded as a violation and the violation score is calculated. In the deep neural network feature extraction, the dimension of the equipment type feature block is 64, encoding the type, quantity, and distribution location information of the equipment. The dimension of the personnel quantity feature block is 32, recording the number and distribution status of the operating personnel. The dimension of the scene environment feature block is 160, including environmental parameters such as light intensity, weather conditions, and ground conditions. The feature vector is mapped through 3 layers of fully connected layers to generate a violation quantification score. The long short-term memory network predicts the evolution trend of hidden dangers. The sequence of violation quantification scores in the last 24 hours is input. The network memory unit saves the scene state transition law. The state transition probability matrix is ​​used to predict the development trend of hidden dangers in the next 4 hours. The state transition probability is obtained based on historical data statistics, reflecting the transformation relationship between different hidden danger levels. Agent classifies hidden dangers based on the scoring range. Red risks correspond to severe levels, endangering the safety of life or causing major property losses. Orange risks correspond to general levels, with risks of personal injury or equipment damage. Yellow risks correspond to minor levels, violating operating specifications but not causing direct harm. Hidden danger response rules adopt differentiated disposal measures for different risk levels. For red risks, emergency plans are activated and related operating activities are stopped immediately. For orange risks, early warning notices are issued to strengthen on-site control. For yellow risks, rectification notices are issued, rectification is completed within a time limit, and regular inspections and tracking are carried out.

[0036] S104. Determine the hidden danger level according to the hidden danger level classification results. The hidden danger level is divided into Level I,

[0037] Level II, Level III, and Level IV, a corresponding response instruction sequence is issued to the construction equipment controller based on the hidden danger level.

[0038] Response rules are extracted according to the hidden danger level classification results, and the hidden danger level threshold judgment conditions are obtained from the response rule library. The hidden danger level is judged based on the hidden danger level threshold judgment conditions to obtain the hidden danger level judgment results; the hidden danger level judgment results are processed by a deep neural network to obtain a response instruction sequence, which includes an operation area control instruction, a personnel evacuation instruction, and an equipment shutdown instruction; the construction site monitoring data stream is processed by a rule reasoner to judge the hidden danger evolution trend. If the increase in the hidden danger value exceeds the amplitude threshold and continues to exceed the sampling period, the response instruction upgrade is triggered; the industrial control bus is used to send the response instruction sequence to the construction site control terminal. After the response instruction sequence is executed, the disposal result is verified through the construction site video monitoring data.

[0039] Exemplarily, according to the hidden danger level classification results, response rules are extracted from the knowledge graph warning response rule base, and the hidden danger level threshold judgment conditions are obtained from the response rule base. Hidden danger scores greater than 0.8 points are judged as major hidden dangers, hidden danger scores between 0.5 and 0.8 are judged as major hidden dangers, hidden danger scores less than 0.5 are judged as serious hidden dangers, and hidden danger scores less than 0.3 are judged as general hidden dangers. A mapping table of hidden danger levels and response instructions is constructed, and the response instructions include operation area control instructions, personnel evacuation instructions, and equipment shutdown instructions. A deep neural network is used to generate a response instruction sequence from the response instruction rule base. For major hidden dangers, a comprehensive control instruction for the operation area is issued, the on-site alarm is activated, the personnel evacuation image or video is captured through the camera in the operation area, and an emergency shutdown instruction is issued to the construction equipment controller; for major hidden dangers and serious hidden dangers, a local control instruction for the operation area is issued, and an illegal evidence collection instruction is issued to the camera; for general hidden dangers, a hidden danger recording instruction is issued, and the hidden danger data is stored in the hidden danger information database according to the timestamp. The rule reasoner is used to dynamically reason the hidden danger data of the construction site, and the hidden danger evolution trend curve is established based on the construction site monitoring data stream. When the hidden danger value increases by more than 0.2 within 5 minutes, it is judged as an upward trend. The upward trend continues for more than 3 sampling cycles to trigger the warning upgrade and execute a higher-level response instruction. When the hidden danger value fluctuates less than 0.1 within 10 minutes, it is judged as a stable trend and the original response level is maintained. The industrial control bus is used to send response instructions to the construction site control terminal in real time. The response instructions are encapsulated using the industrial control protocol. After receiving the instructions, the control terminal executes according to the instruction priority. The hidden danger level, hidden danger location, and disposal measure execution status are displayed through the display terminal of the power grid monitoring center. The disposal measure execution results are verified from the construction site video monitoring data. The disposal process data records are saved in the emergency response database. The record content includes the hidden danger occurrence time, hidden danger level, response measures, and disposal effect. The response rule base management is subdivided into multiple business levels. Taking the equipment hoisting operation scenario as an example, when the distance between the hoisted heavy object and the ground exceeds 5 meters, the distance between the hoisting equipment and the safe passage is less than 10 meters, which is judged as a major hidden danger and triggers a red warning. The safe passage distance is between 10 meters and 15 meters, which is judged as a serious hidden danger and triggers an orange warning. The safe passage distance is greater than 15 meters and is judged as a general hidden danger and triggers a yellow warning. The response instruction generation is constructed using a deep neural network. The network input layer dimension is 128, corresponding to the characteristics of the construction site monitoring data. The number of hidden layer neurons is set to 256, and the output layer corresponds to different levels of response instructions. Taking the hoisting operation as an example, the major hidden danger response instructions include three instructions: crane emergency braking, sling locking, and emergency evacuation of operators. The major and serious hidden danger response instructions include hoisting load reduction, operation area warning, and personnel transfer avoidance instructions. The general hidden danger response instructions include on-site recording and video surveillance tracking instructions.The rule reasoner dynamically monitors the hidden danger data of the construction site. Taking the monitoring data of the hoisting operation as an example, the distance between the hoisted heavy object and the ground decreased from 8 meters to 6 meters within 5 minutes, and the distance of the safe passage decreased from 12 meters to 9 meters, triggering the early warning upgrade mechanism, the response level was upgraded from orange warning to red warning, and the emergency braking command of the hoisting was executed. The industrial control bus adopts a real-time data transmission protocol. The length of the control instruction packet header is 32 bytes, the length of the instruction content is 64 bytes, and the communication cycle is 100 milliseconds. Taking the control of the hoisting equipment as an example, after the emergency braking command is issued, the hoisting equipment controller responds and executes the brake within 50 milliseconds. The braking process lasts for 2 seconds, and the braking status data is fed back to the monitoring center in real time. The recording resolution of the on-site video monitoring data is 1920x1080 pixels, and the frame rate is 25 frames per second. Taking the monitoring of the hoisting operation as an example, when the early warning is triggered, the hoisting equipment, the operator, and the surrounding environment are shot from multiple cameras. The video data is saved for 30 days. The hidden danger disposal process data includes the disposal start time, disposal completion time, disposal measure execution status, and disposal effect verification results. The execution of response instructions on the construction site adopts a multi-level linkage mechanism. Taking the prevention of hoisting accidents as an example, when the safety distance is insufficient during the hoisting process, the hoisting equipment is first triggered to slow down, and the alarm in the operation area is activated to remind the surrounding personnel to avoid it. If the safety distance continues to decrease, the hoisting emergency brake is executed, and the on-site personnel are notified to evacuate urgently. The monitoring center tracks the disposal process in real time until the hidden danger is eliminated. The priority division of response instructions reflects the risk control requirements during the construction operation. The response priority of hidden dangers that pose a great threat to personnel safety is the highest, followed by equipment safety, and the standardization of construction technology ranks third. Resource allocation and progress management have the lowest priority. Hidden dangers are hierarchically controlled through reasonable priority settings.

[0040] S105. Obtain various on-site safety and supervision result data, including law enforcement instrument data, camera equipment data and sensor data, and combine them with real-time processing records to perform feature extraction and fusion. Integrate data from different sources to generate a fused multi-modal feature vector.

[0041] Video frames, image data and sensor values ​​are collected respectively by law enforcement instruments, cameras and sensors, and the video frames, image data and sensor values ​​are stored in a data cache table according to timestamp indexes; a convolutional neural network is used to process the image data in the data cache table to obtain an image feature vector, and a sensor feature vector is obtained by calculating statistics on the sensor values. The image feature vector and the sensor feature vector are aligned and spliced ​​in the time dimension to form a spliced ​​feature vector; the spliced ​​feature vector is processed by a feature fusion device, and the feature fusion device uses a mapping matrix to transform the features and then weights them according to weight coefficients to obtain fused features, thereby obtaining a fused multimodal feature vector; the fused features are decomposed into principal components to obtain principal component features, and the principal component features are normalized to establish a corresponding relationship with the construction process stage and stored in a feature database.

[0042] Exemplarily, according to the law enforcement instrument, camera equipment, and sensor collection timing marks deployed at the construction site, the clock synchronization signal is sent through the network time protocol server, the time accuracy of the marking signal is milliseconds, and the video frame, image data, and sensor value are extracted from the data stream obtained by the acquisition device. The data at the same time is organized into a data cache table according to the timing mark, and the data cache table stores the data of the last 10 seconds according to the timestamp index. The law enforcement instrument and camera equipment image data are read from the data cache table in time sequence, and the image features are extracted using a convolutional neural network. The network contains 5 convolutional layers and 3 fully connected layers, and finally outputs a 256-dimensional image space feature vector. At the same time, the sensor values ​​are read to generate a 32-dimensional sensor feature vector. The sensor feature vector contains 4 groups of values: temperature, vibration, displacement, and inclination. Each group of values ​​contains 4 statistics: mean, variance, peak, and frequency. The two types of feature vectors are aligned and spliced ​​in the time dimension. The feature fusion device is used to combine the spliced ​​feature vectors. The feature fusion device includes a feature mapping layer and a feature integration layer. The mapping layer performs linear transformation on the image features and sensor features respectively. The transformation matrix parameters are obtained through training of construction scene label data. The integration layer weights and sums the transformed features according to the weight coefficients. The weight coefficients are adaptively adjusted with the construction process stage. The fusion features are reduced in dimension by principal component decomposition. The principal components with cumulative variance contribution rate exceeding 95% are retained to construct multimodal feature vectors. The maximum and minimum values ​​of each dimension of the feature vector are normalized. The normalization interval is set to 0 to 1. The construction process stage mark is read from the real-time processing record, and the corresponding relationship between the feature vector and the process stage is established. The feature vector and the corresponding mark are written into the feature database. The feature data is organized according to the dual index of timestamp and process stage. The timing synchronization of multi-source data acquisition equipment at the construction site is crucial to data fusion. The network time protocol server broadcasts synchronization signals to the acquisition equipment every 100 milliseconds. The video frame rate of the law enforcement instrument is 30 frames per second, the image resolution of the fixed camera is 1920x1080 pixels, the temperature sensor sampling frequency is 10 times per second, the vibration sensor sampling frequency is 1000 times per second, the displacement sensor sampling frequency is 100 times per second, and the tilt sensor sampling frequency is 50 times per second. When the convolutional neural network extracts image features, the first convolution layer uses 64 3x3 convolution kernels, the second layer uses 128 3x3 convolution kernels, the third layer uses 256 3x3 convolution kernels, the fourth layer uses 512 3x3 convolution kernels, and the fifth layer uses 512 3x3 convolution kernels. Each convolution layer is followed by a maximum pooling layer with a pooling kernel size of 2x2. The number of neurons in the fully connected layer is 4096, 4096, and 256, respectively. The extracted feature vector reflects the spatial distribution and morphological characteristics of the target in the image.The sensor feature statistics show that the temperature value fluctuates between 15 and 35 degrees Celsius during the construction process, the vibration amplitude ranges from 0 to 200 mm per second, the displacement range is from 0 to 500 mm, and the inclination angle changes within plus or minus 15 degrees. The mean in the feature statistics reflects the overall level, the variance characterizes the degree of fluctuation, the peak indicates the limit state, and the frequency feature reflects the law of change. The feature fusion device assigns different weights to image features and sensor features. In the excavation process, the image feature weight is 0.7 and the sensor feature weight is 0.3. In the hoisting process, the image feature weight is 0.6 and the sensor feature weight is 0.4. In the welding process, the image feature weight is 0.5 and the sensor feature weight is 0.5. The weight coefficient is automatically adjusted with the process changes. When the principal component decomposition reduces the dimension of the fused feature, the cumulative variance contribution rate of the first 10 principal components can usually reach more than 95%. After the dimension reduction, the feature vector dimension is reduced from 288 to 32. The maximum and minimum value normalization ensures the uniformity of different feature dimensions, which is convenient for subsequent processing. The construction process stage marking includes earth excavation, foundation pouring, equipment hoisting, steel structure welding and other construction links. The corresponding relationship between the feature vector and the process stage is stored to facilitate data retrieval and analysis. The problem of missing data needs to be dealt with during the multi-source data fusion process. When the law enforcement device video data is missing, the fixed camera image data is used instead. When a sensor data is missing, the adjacent time data is interpolated to ensure the integrity of the feature vector. The feature fusion result reflects the overall state of the construction site, including visual information and physical parameters, providing a data basis for subsequent construction process monitoring.

[0043] S106. Generate a summary report of power grid supervision based on the fused multi-modal feature vector. The summary report includes information on problems existing at the construction site, hidden danger levels, and emergency response measures.

[0044] The construction rule graph nodes are retrieved according to the multimodal feature vector, and the cosine similarity between the feature vector and the rule node attribute value is calculated by using a deep neural network to obtain a rule violation degree data table; the rule violation degree data table is subjected to density clustering processing to obtain the corresponding hidden danger level; the response disposal node is obtained, and the graph convolutional neural network is used to perform similarity matching on the response disposal node to determine the disposal node corresponding to the hidden danger level interval; a summary report is generated according to the disposal node, and the summary report includes four chapters: project overview, construction site problems, hidden danger level identification, and emergency measures recommendations.

[0045] Exemplarily, based on the multimodal feature vector of the construction site, the Agent retrieves the construction rule graph node from the knowledge graph. The rule node attributes include the equipment spacing value, the standard value of personnel clothing, and the protection distance value of the power grid facility. The deep neural network is used to calculate the cosine similarity between the feature vector and the rule node attribute value. When the similarity is greater than 0.8, it is determined that the rule match is successful, and a rule violation degree data table is constructed, in which the rule number, violation degree value, and scene timestamp are recorded. The Agent clusters the rule violation degree data table, and uses the density clustering algorithm to divide the violation degree value into three hidden danger level intervals of red, orange, and yellow. The red interval corresponds to serious hidden dangers with a violation degree value greater than 0.8, the orange interval corresponds to general hidden dangers with a violation degree value between 0.5 and 0.8, and the yellow interval corresponds to minor hidden dangers with a violation degree value less than 0.5. A hidden danger problem description text is generated, and each description text contains three basic elements: the location of the hidden danger, the violation behavior, and the degree of harm. Response and disposal nodes are extracted from the Agent emergency plan map, and graph convolutional neural networks are used to match nodes for similarity. Node attributes include three dimensions: response level, disposal instructions, and execution time limit. The disposal node with the highest matching degree is selected for different hidden danger level intervals to generate emergency disposal text. The text contains three parts: response measures, disposal steps, and response requirements. The disposal text is arranged and organized according to the response time sequence relationship. Through the Agent report generator, the hidden danger problem description text and emergency disposal text are organized according to the supervision report format standard. The report format includes four fixed chapters: project overview, construction site problems, hidden danger level identification, and emergency measures recommendations. The construction area coordinates, work content, and supervision period information are filled in the project overview, the problem description text is filled in the construction site problem chapter, the hidden danger level determination basis is filled in the hidden danger level identification chapter, and the disposal text is filled in the emergency measures recommendation chapter to form a complete supervision report document. Agent knowledge graph rule retrieval takes hoisting operation as an example. The safety distance rule value of hoisting equipment in the graph node is 10 meters, the minimum safety distance between hoisted heavy objects and overhead lines is 5 meters, and the clothing requirements for hoisting workers include safety helmets, reflective clothing, and protective shoes. The deep neural network calculates the similarity between the multimodal feature vector and the rule attribute. When the hoisting equipment is 3 meters away from the power line, the rule violation degree value is 0.8, triggering a serious hidden danger judgment. When density clustering clusters the violation degree values, the safety distance between construction equipment and overhead lines is used as the benchmark. The violation degree value greater than 0.8 corresponds to the red interval, indicating that the equipment distance is less than 3 meters; the violation degree value between 0.5 and 0.8 corresponds to the orange interval, indicating that the equipment distance is between 3 and 5 meters; the violation degree value less than 0.5 corresponds to the yellow interval, indicating that the equipment distance is greater than 5 meters. The text format of the hidden danger description is "XX violation behavior was found at the XX coordinate position, violating Article XX of the XX regulations, and there is XX hazard".When the graph convolutional neural network matches the emergency plan, the handling nodes of the hoisting operation scene include three dimensions: emergency response, warning area, and personnel evacuation. When the red warning is triggered, the handling text is "immediately stop the hoisting operation, set a warning area with a radius of 30 meters, and evacuate all personnel in the area"; when the orange warning is triggered, the handling text is "lower the height of the boom, set up warning signs, and restrict personnel from entering"; when the yellow warning is triggered, the handling text is "strengthen on-site monitoring and prepare emergency equipment". The project overview in the supervision report format standard includes the coordinates of the construction site, the content of the construction process, and the supervision time range. The description of the construction site problem adopts the "location + object + behavior + degree" structure, such as "In the tower crane operation area in the northeast corner of the construction area, the closest distance between the boom and the 10 kV overhead line is 3 meters, which seriously violates Article 15 of the construction safety regulations and there is a risk of electric shock". The hazard level identification refers to relevant standard clauses, and the emergency measures are recommended to be listed according to the response level. The report data is indexed according to the three dimensions of construction area, supervision period, and hazard level to facilitate subsequent query statistics. The report review record includes reading time, feedback, and follow-up measures to form a closed-loop management. The rectification record of the construction unit includes the rectification time, rectification plan, and acceptance results, which serve as a reference for subsequent supervision. Statistical analysis of multiple report data reflects the trend of safety management on the construction site and provides data support for improving construction technology. The rule library is regularly updated and maintained, and rule nodes are supplemented in a timely manner in combination with new construction process requirements, equipment operating procedures, and changes in the working environment. The emergency plan library optimizes the disposal measures based on the actual disposal effect feedback to improve the accuracy of response disposal. The report template is dynamically adjusted in combination with regulatory requirements to ensure that the report content covers regulatory priorities.

[0046] S107. Based on the summary report of the power grid supervision, a text generation method based on natural language processing is used to generate a standardized construction report that meets the construction site management requirements. The content of the construction report includes basic information of the construction site, existing problems, hidden danger levels and emergency response measures.

[0047] The text information of the summary report is obtained, and the structural features of the text information are extracted through the long short-term memory network to construct a structured feature matrix including the construction object dimension, the work content dimension and the supervision requirement dimension; the problem description text, the hidden danger determination text and the disposal measure text are extracted according to the structured feature matrix, and the text is segmented using a construction professional word segmentation dictionary; for the text after the word segmentation, the matching degree between the text semantics and the specification requirements is calculated through a bidirectional attention network, and if the matching degree is lower than the matching degree threshold, semantic replacement is performed according to the construction professional knowledge base; the semantically replaced text is filled into the corresponding chapter position according to the standard report format requirements, and the professional terminology standardization, sentence fluency and logical consistency of the text are verified through a grammar checker to generate a standardized construction report.

[0048] For example, based on the summary report of construction and power grid supervision, the text structure information is extracted through the long short-term memory network. The network input layer dimension is 512, corresponding to the text word vector, and the number of hidden layer neurons is 256. The construction site coordinates, process content, and supervision period information are extracted to construct a structured matrix. The matrix row vector represents the text paragraph, and the column vector represents the paragraph semantic feature. The feature dimension includes three basic dimensions: construction object, work content, and supervision requirements. Each basic dimension is subdivided into 4 sub-dimensions. The problem description, hidden danger determination, and disposal measures text are extracted from the structured matrix, and the text is segmented according to the construction site management specifications. The word segmentation dictionary contains four phrases: construction machinery list, work process terms, power grid equipment terms, and safety management terms. Each phrase contains professional terms and corresponding interpretations. According to the word segmentation results, the corresponding construction report standard terms are extracted from the language component library. The component library contains subject-predicate-object sentence templates, attributive adverbial templates, and conditional judgment templates. The semantics of the reorganized text is understood through a bidirectional attention network. The input includes text fragments and corresponding term definitions. The attention mechanism calculates the matching degree between the text semantics and the specification requirements, and replaces the semantic content below the matching threshold. The replacement rules are established based on the professional construction knowledge base, including action description rules, attribute description rules, and relationship description rules, to generate construction report paragraphs that meet management specifications. The report template is read from the construction report standard format library. The template contains five fixed chapters: project overview, construction content, quality control, safety management, and supervision suggestions. Each chapter sets a word limit and keyword requirements. The paragraph is filled in the corresponding chapter position. The grammar checker is used to verify the compliance of the text. The inspection content includes the standardization of professional terms, the fluency of sentences, and the logical consistency. The content that does not meet the requirements is marked and automatically corrected to generate a standardized construction report. When the long short-term memory network processes text structure information, taking the power grid construction report as an example, the network input layer receives the text word vector, the word vector dimension is 512, and contains basic information text such as "the construction site is located in the northeast corner of the substation, coordinates X = 120.5, Y = 30.8". Through the hidden layer, the structured features are extracted to obtain a standardized description including location, time, and work content to form a feature matrix. The word segmentation process uses professional dictionary matching, and the construction machinery phrase contains "

[0049] crawler crane, truck crane, lifting vehicle" and other equipment names, operation process phrases include "foundation pit excavation, steel bar binding, concrete pouring" and other process names, power grid equipment phrases include "transformer, circuit breaker, lightning arrester"

[0050] The names of devices such as "warning area, edge protection, grounding protection" and other terms are included in the safety management phrases. The sentence templates in the language component library are set for different content types. The problem description adopts the format of "XX equipment / personnel found to have XX violations at XX location", the hidden danger determination adopts the format of "violation of XX clauses of XX specification, belonging to XX level hidden danger", and the disposal measures adopt the format of "immediately take XX measures and take XX precautions". When the two-way attention network processes the text semantics, for the description of "the boom extension angle is too large during tower crane operation", the standard statement of "the boom extension angle during truck crane operation exceeds the rated working range" is obtained through standard comparison. The replacement basis comes from the specific requirements for lifting machinery operations in the construction specifications. The report chapter format has strict regulations, with a limit of 500 words for project overview, 1000 words for construction content, 800 words for quality control, 1000 words for safety management, and 500 words for supervision suggestions. Each chapter is required to contain specific keywords, such as the safety management chapter must contain words such as "warning signs, safety distance, emergency measures". When the grammar checker reviews the standardization of professional terms, "cable burial" is changed to "power cable laying" and "distribution cabinet installation" is changed to "distribution device installation" to ensure that the terms comply with power industry standards. Sentence coherence checks include subject-predicate collocation, use of conjunctions, and sentence integrity verification. Logical consistency checks ensure that there are no contradictions in the descriptions before and after. During report generation, the accuracy of the text description is verified in combination with on-site monitoring data. For example, when the report describes "the depth of the foundation pit is 4.5 meters", the measurement data is checked to confirm the accuracy of the description. On-site photo information is used to support the problem description to ensure that the text description is consistent with the actual situation on site. Suggested disposal measures are provided in combination with the historical case database to ensure that the suggested measures are operational.

[0051] S108. Store the generated standardized construction report in a database and push it to relevant management personnel so that the management personnel can understand the status of the construction site and power grid in a timely manner.

[0052] The standardized construction report is stored in a distributed database, wherein the distributed database includes a main table for recording basic report information and a subtable for recording detailed content, and a reverse index table is used to store construction professional terms; according to the construction professional terms, a long short-term memory network is used to extract information from the content of the report to obtain a construction progress percentage label, a quality problem classification label, and a safety hazard level label; for the construction progress percentage label, the quality problem classification label, and the safety hazard level label, the standardized construction report is extracted in blocks according to the report content filtering rules to obtain the content of the construction management position, the content of the safety supervision position, and the content of the power grid operation and maintenance position; according to the content of the construction management position, the content of the safety supervision position, and the content of the power grid operation and maintenance position, the construction report information is sent to the management personnel terminal through the industrial data bus, and if the terminal returns a confirmation package, the message delivery time is recorded, and if the confirmation package is not received, the message retransmission is triggered.

[0053] Exemplarily, according to the standardized construction report structure, a distributed database is used to store the report content. The database main table records the basic information of the report, including the report number, generation time, construction area, and supervision period. The sub-table records the detailed content of the report, including construction progress, quality hazards, safety risks, and rectification measures. A reverse index table is established to store construction professional terms, and the mapping relationship between terms and report content is established through the inverted index. The report content is extracted through the long short-term memory network. The network input layer receives the report text vector, the hidden layer extracts text features, and the output layer generates the construction progress percentage, quality problem classification, safety hazard level, and rectification requirement classification labels. The management position responsibility list is obtained from the construction supervision management system, and the job responsibility keywords are mapped with the report content labels to generate job information filtering rules. The report content filtering rules are used to extract the construction report in blocks. The construction management position extracts the construction progress and process quality content, the safety supervision position extracts the hazard level and disposal measures content, and the power grid operation and maintenance position extracts the power facilities and safety distance content. Sensitive information is desensitized according to the management personnel's authority level. The senior authority displays all information, the intermediate authority blocks the rectification period, and the primary authority blocks the accountability. According to the division of labor of management personnel, the construction report information is pushed through the industrial data bus. The data bus adopts a publish-subscribe mechanism. The management terminal subscribes to the corresponding topic message. The message push adopts a confirmation retransmission mechanism. The terminal returns a confirmation package after receiving the message. Failure to receive the confirmation package triggers message retransmission, records the message delivery time, monitors the message reading status, marks the read messages, and stores the processing results in the management process database. The distributed database stores construction reports using a master-slave table structure. Taking the substation expansion project as an example, the main table records basic information such as project number GC2024001, supervision period from March 15 to April 15, 2024, and the north side of the substation in the construction area. The subtable records specific contents such as 35% completion of the construction progress, quality defects in foundation pit support, inadequate edge protection, and rectification before April 20. Professional entries include construction terms such as earth excavation, steel bar binding, and concrete pouring. When the long short-term memory network processes the report text, it inputs a descriptive text such as "The spacing of the support piles on the east side of the No. 3 foundation pit exceeds the design value by 50 mm". The network extracts numerical features and keywords, and outputs labels such as quality problem classification as Level 2, rectification deadline of 3 days, and construction process deviation. Among the management job responsibilities, the construction director focuses on the progress and quality of the project, the safety director focuses on illegal operations and hidden dangers, and the equipment director focuses on machinery and materials. The block extraction of report content follows the principle of differentiation. Construction management personnel obtain construction data such as "foundation pit excavation depth of 4.5 meters, support pile spacing deviation of 50 mm", safety supervision personnel obtain hidden danger information such as "edge guardrail missing 2 meters, warning signs missing 3 places", and power grid operation and maintenance personnel obtain "construction machinery is 4.5 meters away from the overhead line" and other data. Senior authority personnel see "

[0054] The rectification must be completed before April 20, and the construction unit will be held accountable." Intermediate-level personnel see "Rectification must be completed within a time limit", and junior-level personnel only see "Rectification in progress". The industrial data bus adopts a message queue mechanism. The topics include three channels: construction management, safety supervision, and equipment maintenance. The management terminal receives messages according to the subscribed topics. The message retransmission interval is 1 second, the maximum number of retransmissions is 3 times, and an alarm prompt is issued if no confirmation is received after the timeout. The message status includes three levels: sent, delivered, and read. The processing record includes reading time, processing opinions, and follow-up measures. The database query supports multi-dimensional retrieval. According to the time dimension, you can query a report for a certain supervision period. According to the regional dimension, you can query the report of the specified construction area, and according to the content dimension, you can search for reports related to quality problems and safety hazards. The reverse index table records the location of professional terms and supports full-text search. Entering "foundation pit support" can locate the relevant paragraph. The management process data record includes the entire process of problem discovery, rectification, implementation of measures, and acceptance and cancellation. Taking the foundation pit support problem as an example, it records the discovery of spacing deviation on March 20, the issuance of rectification notice on March 21, the addition of support piles on March 22, and the acceptance on March 23, forming a complete management chain. The problem handling record is used for subsequent work improvement, providing data support for the optimization of construction technology and the improvement of management mechanism.

[0055] The present invention provides a power grid security monitoring system, which mainly includes:

[0056] The data acquisition and real-time target detection module is used to obtain construction business data according to the construction operation scene, call the camera equipment to collect the image data of the construction site, perform real-time target detection and recognition on the collected image data, and obtain the real-time target detection and recognition results;

[0057] The preliminary hidden danger analysis module is used to form a knowledge graph of the construction scene based on the real-time target detection and recognition results, and use the knowledge graph as the basis for reasoning to generate preliminary hidden danger analysis results;

[0058] The hidden danger level classification module is used to classify and rate hidden dangers according to the severity of hidden dangers based on the preliminary hidden danger analysis results and the construction site business rules, and obtain hidden danger level classification results;

[0059] A response instruction generation and execution module is used to determine the hidden danger level according to the hidden danger level classification result. The hidden danger levels are divided into level I, level II, level III, and level IV, and send a corresponding response instruction sequence to the construction equipment controller based on the hidden danger level;

[0060] The multimodal data fusion module is used to obtain a variety of on-site safety and supervision results data, including law enforcement instrument data, camera equipment data and sensor data, and combine real-time processing records to perform feature extraction and fusion, integrate data from different sources, and generate a fused multimodal feature vector;

[0061] The supervision summary report generation module is used to generate a summary report of power grid supervision based on the fused multi-modal feature vector. The summary report includes the problems existing at the construction site, the level of hidden dangers, and the emergency response measures.

[0062] The standardized construction report generation module is used to generate a standardized construction report that meets the construction site management requirements based on the summary report of the power grid supervision and the text generation method based on natural language processing. The construction report content includes basic information of the construction site, existing problems, hidden danger levels and emergency response measures;

[0063] The report storage and push module is used to store the generated standardized construction reports in the database and push them to relevant managers so that they can understand the construction site and power grid status in a timely manner.

[0064] What is disclosed above is only a preferred embodiment of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. A person skilled in the art can understand that all or part of the processes of the above embodiments and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A power grid security supervision method, characterized in that: The method comprises: According to the construction operation scenario, the construction business data is obtained, the camera equipment is called to collect the image data of the construction site, and the collected image data is subjected to real-time target detection and recognition to obtain the real-time target detection and recognition results; Based on the real-time target detection and recognition results, a knowledge graph of the construction scene is formed, and the knowledge graph is used as the basis for reasoning to generate preliminary hidden danger analysis results; Based on the preliminary hidden danger analysis results and the construction site business rules, the hidden dangers are classified and rated according to their severity to obtain the hidden danger level classification results; Determine the hidden danger level according to the hidden danger level classification result, the hidden danger level is divided into level I, level II, level III, and level IV, and send a corresponding response instruction sequence to the construction equipment controller based on the hidden danger level; Obtain a variety of on-site safety and supervision results data, including law enforcement device data, camera equipment data and sensor data, and combine real-time processing records to perform feature extraction and fusion, integrate data from different sources, and generate a fused multi-modal feature vector; Generate a summary report of power grid supervision based on the fused multi-modal feature vector. The summary report includes the problems existing at the construction site, the level of hidden dangers, and emergency response measures. According to the summary report of the power grid supervision, a text generation method based on natural language processing is used to generate a standardized construction report that meets the construction site management requirements. The construction report includes basic information of the construction site, existing problems, hidden danger levels and emergency response measures; The generated standardized construction reports are stored in the database and pushed to relevant managers so that they can understand the status of the construction site and power grid in a timely manner.

2. The method according to claim 1, characterized in that The method of acquiring construction business data according to the construction operation scene, calling the camera equipment to collect the image data of the construction site, performing real-time target detection and recognition on the collected image data, and obtaining the real-time target detection and recognition results includes: Using a positioning calibration rod to obtain the construction site reference coordinates, the reference coordinates are used to establish a camera equipment calibration coordinate system, and the image data is subjected to distortion correction through the calibration coordinate system to obtain standard image data; Extracting construction target features using a residual neural network based on the standard image data, extracting position coordinate data from the target features and calculating the three-dimensional position data of the construction target space using the position coordinate data of the preset calibration points; If the three-dimensional position data of the construction target space exceeds the pre-established construction area boundary coordinate range, an alarm is triggered and the motion trajectory data of the construction target is collected; The motion trajectory data of the construction operation target is smoothed by a Kalman filter, and the smoothed motion trajectory data is classified and identified by a support vector machine to obtain a construction operation target operation state determination result.

3. The method according to claim 1, characterized in that The method forms a knowledge graph of the construction scene based on the real-time target detection and recognition results, and uses the knowledge graph as the basis for reasoning to generate preliminary hidden danger analysis results, including: Obtaining real-time target detection results at the construction site, the target detection results including construction equipment location coordinate data and construction personnel clothing marking data; A knowledge graph node is constructed in a triplet manner, wherein the knowledge graph node generates a device node and a personnel node according to the target detection result, and the device node and the personnel node are associated with each other according to a mapping relationship of a rule base; According to the knowledge graph nodes, a graph attention network is used to calculate the semantic association weights between nodes, and the semantic association weights are used to generate an action sequence graph structure based on the spatial distance between nodes and the degree of temporal association; The action sequence graph structure is used to trigger the Agent perception module to perform path reasoning, and the path reasoning is based on node semantic association and risk level threshold comparison to obtain preliminary hidden danger analysis results.

4. The method according to claim 1, characterized in that: According to the preliminary hidden danger analysis results, combined with the construction site business rules, the hidden dangers are classified and rated according to the severity of the hidden dangers, and the hidden danger level classification results are obtained, including: Retrieving construction rule nodes according to the knowledge graph to obtain quantitative index data, constructing a hidden danger determination matrix through the quantitative index data, and calculating the first rule violation quantitative score through the hidden danger determination matrix; Acquire historical hidden danger record data, use a deep neural network to perform feature extraction on the historical hidden danger record data to obtain a hidden danger feature vector, and use a feature extractor to generate a second rule violation quantization score for the hidden danger feature vector; Based on the first rule violation quantization score and the second rule violation quantization score, a long short-term memory network predictor is input, and the predictor calculates a state transition probability matrix according to a scene state transition rule to obtain a hidden danger prediction score; According to the average of the quantified score of the first rule violation and the quantified score of the second rule violation and the hidden danger prediction score, the hidden dangers are graded according to the preset scoring criteria.

5. The method according to claim 1, characterized in that The hidden danger level is determined according to the hidden danger level classification result, and the hidden danger levels are divided into level I, level II, level III, and level IV. Based on the hidden danger level, a corresponding response instruction sequence is sent to the construction equipment controller, including: Extract response rules according to the hidden danger level classification results, obtain hidden danger level threshold determination conditions from the response rule library, determine the hidden danger level based on the hidden danger level threshold determination conditions, and obtain hidden danger level determination results; Using a deep neural network to process the hidden danger level determination result, obtaining a response instruction sequence, wherein the response instruction sequence includes an operation area control instruction, a personnel evacuation instruction, and an equipment shutdown instruction; The construction site monitoring data stream is processed through the rule reasoner to determine the evolution trend of hidden dangers. If the increase in the hidden danger value exceeds the threshold and lasts for more than the sampling period, the response instruction upgrade is triggered; An industrial control bus is used to send a response instruction sequence to the construction site control terminal. After the response instruction sequence is executed, the disposal result is verified through the construction site video monitoring data.

6. The method according to claim 1, characterized in that The method obtains a variety of on-site safety and supervision result data, including law enforcement instrument data, camera equipment data and sensor data, and combines real-time processing records to perform feature extraction and fusion; integrates data from different sources to generate a fused multi-modal feature vector, including: The video frames, image data and sensor values ​​are collected by the body camera, the camera and the sensor respectively, and the video frames, image data and sensor values ​​are stored in the data cache table according to the timestamp index; Using a convolutional neural network to process the image data in the data cache table to obtain an image feature vector, calculating statistics on the sensor values ​​to obtain a sensor feature vector, and aligning and splicing the image feature vector with the sensor feature vector in a time dimension to form a spliced ​​feature vector; The concatenated feature vector is processed by a feature fusion device, and the feature fusion device transforms the feature by a mapping matrix and then weights the transformed feature according to a weight coefficient to obtain a fused feature, thereby obtaining a fused multimodal feature vector.

7. The method according to claim 1, characterized in that The summary report of the power grid supervision is generated according to the fused multi-modal feature vector. The content of the summary report includes the problems existing at the construction site, the level of hidden dangers, and the emergency response measures, including: The construction rule graph nodes are retrieved according to the multimodal feature vectors, and the cosine similarity between the feature vectors and the attribute values ​​of the rule nodes is calculated using a deep neural network to obtain a rule violation degree data table; After performing density clustering processing on the rule violation degree data table, the corresponding hidden danger level is obtained; Obtain response and disposal nodes, use graph convolutional neural network to perform similarity matching on the response and disposal nodes, and determine the disposal nodes corresponding to the hidden danger level interval; A summary report is generated according to the disposal node, and the summary report includes four chapters: project overview, construction site problems, hazard level identification and emergency measures recommendations.

8. The method according to claim 1, characterized in that: According to the summary report of the power grid supervision, a text generation method based on natural language processing is used to generate a standardized construction report that meets the construction site management requirements. The construction report includes basic information of the construction site, existing problems, hidden danger levels and emergency response measures, including: Obtain the text information of the summary report, extract the structural features of the text information through the long short-term memory network, and construct a structured feature matrix including the construction object dimension, the work content dimension, and the supervision requirement dimension; Extracting the problem description text, the hidden danger determination text and the disposal measure text according to the structured feature matrix, and performing word segmentation processing on the text using a construction professional word segmentation dictionary; For the text after word segmentation, the matching degree between the text semantics and the specification requirements is calculated through a bidirectional attention network. If the matching degree is lower than the matching degree threshold, semantic replacement is performed according to the construction professional knowledge base; The semantically replaced text is filled into the corresponding chapter position according to the standard report format requirements, and the professional terminology standardization, sentence fluency and logical consistency of the text are verified through the grammar checker to generate a standardized construction report.

9. The method according to claim 1, characterized in that: The generated standardized construction report is stored in the database and pushed to relevant managers so that they can understand the construction site and power grid status in a timely manner, including: Storing standardized construction reports in a distributed database, the distributed database includes a main table for recording basic report information and a subtable for recording detailed content, and storing construction professional terms through a reverse index table; According to the construction professional terms, a long short-term memory network is used to extract information from the content of the report to obtain a construction progress percentage label, a quality problem classification label, and a safety hazard level label; For the construction progress percentage label, quality problem classification label, and safety hazard level label, the standardized construction report is extracted in blocks through the report content filtering rules to obtain the content of the construction management position, the content of the safety supervision position, and the content of the power grid operation and maintenance position; According to the content of the construction management position, the safety supervision position and the power grid operation and maintenance position, the construction report information is sent to the management personnel terminal through the industrial data bus. If the terminal returns a confirmation package, the message delivery time is recorded. If the confirmation package is not received, the message retransmission is triggered.

10. The power grid security monitoring system is characterized by: The system comprises: The data acquisition and real-time target detection module is used to obtain construction business data according to the construction operation scene, call the camera equipment to collect the image data of the construction site, perform real-time target detection and recognition on the collected image data, and obtain the real-time target detection and recognition results; The preliminary hidden danger analysis module is used to form a knowledge graph of the construction scene based on the real-time target detection and recognition results, and use the knowledge graph as the basis for reasoning to generate preliminary hidden danger analysis results; The hidden danger level classification module is used to classify and rate hidden dangers according to the severity of hidden dangers based on the preliminary hidden danger analysis results and the construction site business rules, and obtain hidden danger level classification results; A response instruction generation and execution module is used to determine the hidden danger level according to the hidden danger level classification result. The hidden danger levels are divided into level I, level II, level III, and level IV, and send a corresponding response instruction sequence to the construction equipment controller based on the hidden danger level; The multimodal data fusion module is used to obtain a variety of on-site safety and supervision results data, including law enforcement instrument data, camera equipment data and sensor data, and combine real-time processing records to perform feature extraction and fusion, integrate data from different sources, and generate a fused multimodal feature vector; The supervision summary report generation module is used to generate a summary report of power grid supervision based on the fused multi-modal feature vector. The summary report includes the problems existing at the construction site, the level of hidden dangers, and the emergency response measures. The standardized construction report generation module is used to generate a standardized construction report that meets the construction site management requirements based on the summary report of the power grid supervision and the text generation method based on natural language processing. The construction report content includes basic information of the construction site, existing problems, hidden danger levels and emergency response measures; The report storage and push module is used to store the generated standardized construction reports in the database and push them to relevant managers so that they can understand the construction site and power grid status in a timely manner.

Images