Whole-process standardized closed-loop safety management and control method and system for power enterprise

Abstract

The invention discloses a whole-process standardized closed-loop safety management and control method and system for a power enterprise, and relates to the technical field of safety management and control of the power enterprise, and the method comprises the steps: collecting image features and sensor features in real time based on multi-source heterogeneous data, and executing unified coding. And constructing an operation and equipment node graph based on the synchronous time sequence, and carrying out hidden danger trend modeling by adopting a space-time neural network. And based on a hidden danger prediction result, triggering preventive intervention in a grading and classifying manner, and performing data backflow updating after intervention is completed. According to the method, the risk prediction capability of dynamic evolution is established, so that a safety management and control system can recognize potential high-risk nodes and key propagation paths in advance, and the perspectiveness and systematicness of hidden danger prevention and control are remarkably improved. A power enterprise safety management and control mechanism which is dynamically optimized in the whole process and standardized in a closed loop is formed, sustainable evolution integrating hidden danger recognition, intervention, feedback and optimization is achieved, and the stability, intelligence and reliability of the system in a complex operation environment are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of safety management and control of electric power enterprises, and specifically to a method and system for standardized closed-loop safety management and control of the entire process of electric power enterprises. Background Art

[0002] With the continuous expansion of power companies and the increasing complexity of production operations, production safety management and control systems are gradually evolving from traditional manual inspections and paper records to digital and intelligent systems. Multi-source data acquisition technologies based on video surveillance, IoT sensors, and job log systems have been implemented in some power plants. Combined with basic alarm rules and risk assessment models, these technologies enable on-site environmental monitoring and preliminary identification of hidden dangers. Simultaneously, with the application of technologies such as deep learning and graph neural networks in the field of industrial intelligence, research on power equipment fault prediction and operational behavior recognition continues to advance, providing new means to enhance the production safety assurance capabilities of power companies. In recent years, some systems have begun to explore preliminary mechanisms for multimodal data fusion and hidden danger trend prediction, but overall, these systems are still primarily focused on single-point monitoring and static response, lacking dynamic closed-loop optimization capabilities.

[0003] While existing power company safety management and control technologies have made some progress in monitoring, perception, and risk identification, they still face numerous limitations. Insufficient real-time fusion and standardized processing of multi-source data leads to information fragmentation during the time-series alignment and feature extraction processes for different types of data, such as images, sensors, and text, impacting the accuracy of subsequent hazard identification. Traditional hazard prediction models are generally based on static historical data and lack dynamic feedback and adaptive adjustment to the actual evolution of hazards after intervention measures. Model prediction performance gradually degrades with environmental changes, making it difficult to support real-time risk perception in complex operational scenarios. Existing technologies generally rely on fixed rule thresholds for risk grading and lack the ability to jointly model node graph structures and dynamic temporal evolution patterns, making it impossible to effectively capture the potential correlation and changing trends between equipment, operating units, and personnel behavior. Therefore, existing technologies have yet to establish a standardized closed-loop safety management and control system that can dynamically optimize hazard trends based on the feedback of actual post-intervention effects. The standardized closed-loop safety control method for the entire process proposed in the present invention realizes a dynamic safety management closed loop by constructing a unified fusion mechanism for multimodal heterogeneous data, spatiotemporal feature modeling based on node graphs, hidden danger trend prediction and graded intervention, and adaptive incremental learning of intervention effect reflux. It effectively overcomes the problems of data fragmentation, static model degradation and delayed risk management response in existing technologies. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problems solved by the present invention are: the existing power safety management and control methods have problems such as insufficient multi-source data fusion and time series synchronization, lack of intervention feedback and dynamic correction in hidden danger trend prediction, rigid risk classification and response mechanism, and how to achieve full-process standardization and dynamic closed-loop optimization of power enterprise safety management and control.

[0006] To address the above-mentioned technical problems, the present invention provides the following technical solution: a standardized closed-loop safety management and control method for the entire process of a power enterprise, comprising real-time acquisition of image and sensor features based on multi-source heterogeneous data and unified encoding. A graph of operations and equipment nodes is constructed based on synchronized time series, and a spatiotemporal neural network is used to model hidden danger trends. Preventive interventions are triggered based on hierarchical classification of hidden danger prediction results, and data is then re-updated after the intervention. This post-intervention data re-updation includes, after the intervention is executed, re-collecting image features, sensor features, and operation log text for the intervention node and its adjacent nodes, comparing the data changes before the intervention, extracting differential features before and after the intervention, and dynamically annotating the actual hidden danger evolution trends based on these differential features. Nodes whose hidden danger risk scores decrease after the intervention are recorded as effective intervention cases and stored in an empirical sample library. Nodes whose risk does not decrease or increases after the intervention are marked as difficult-to-predict samples and stored in an abnormal sample library. After completing every fifty closed-loop interventions, the empirical and abnormal samples are summarized, and the spatiotemporal neural network model parameters are updated to optimize hidden danger prediction capabilities through incremental learning.

[0007] As a preferred embodiment of the standardized closed-loop safety management and control method for power enterprises throughout the entire process described in the present invention, the real-time acquisition of image and sensor features based on multi-source heterogeneous data includes deploying video acquisition equipment to continuously collect video stream data from the power enterprise's operating area and extract key frames. Based on the extracted key frames, image features are detected and extracted, including the worker's safety attire status, abnormal operating movements, and changes in the operating environment structure.

[0008] As a preferred embodiment of the standardized closed-loop safety management and control method for power enterprises throughout the entire process described in the present invention, the unified coding includes collecting environmental parameters and key equipment operating status parameters within the power operation environment, aligning all data along a unified timeline, and normalizing different sensor features to a unified feature space dimension. These features are then combined with image features to form a multimodal fusion input.

[0009] As a preferred solution of the method for standardized closed-loop safety management and control of the entire process of electric power enterprises described in the present invention, the method comprises: constructing a job and equipment node graph based on a synchronous time series, collecting the text of the electric power job log, extracting the job unit, personnel identification, equipment identification and job category as node entities. Graph connection edges are established between nodes that have physical associations and job collaboration relationships, wherein the physical proximity relationship is automatically derived according to the equipment layout diagram, and the collaboration relationship is generated by reasoning based on the job process log. Node features include real-time sensor data features, image anomaly detection features and job text encoding features. Each node feature is combined into a feature vector sequence according to the node number and time step. All node graph structures are dynamically updated with job arrangements and environmental changes, and the adjacency matrix is ​​automatically and synchronously adjusted when the node changes.

[0010] As a preferred solution of the method for standardized closed-loop safety management and control of the entire process of power enterprises described in the present invention, the method includes: the use of a spatiotemporal neural network to model hidden danger trends includes, based on the constructed operation and equipment node graph, using a spatial convolutional network to perform graph convolution operations on adjacent nodes to extract the spatial feature relationships between nodes. In the time dimension, a long short-term memory network or a time series modeling network based on a self-attention mechanism is used to extract the dynamic evolution pattern of each node over time. The spatial convolution output and the time modeling output are feature spliced, and finally the hidden danger risk score for each node in the next three days is output, and the risk score ranges from 0 to 1.

[0011] As a preferred embodiment of the standardized closed-loop safety management and control method for power enterprises throughout the entire process described in the present invention, the preventive intervention triggered by grading and categorizing hidden danger prediction results includes generating instructions to suspend operations, arrange on-site safety re-inspections, and archive re-inspection records for nodes with risk scores exceeding a high-risk threshold. For nodes with risk scores in the medium-risk range, instructions to increase the frequency of manual inspections and restrict the operation approval process are generated, with approval records forming a complete track within the system.

[0012] As a preferred solution of the method for standardized closed-loop safety management and control of the entire process of power enterprises described in the present invention, wherein: the data reflux update after the intervention is completed includes:

[0013] After implementing an intervention, image features, sensor features, and log text from the intervention node and its adjacent nodes are recollected and compared with pre-intervention data. Differential features between nodes before and after the intervention are extracted based on node number and time series order, dynamically annotating the actual hidden danger evolution trend. If the hidden danger risk score decreases after the intervention compared to before the intervention, the node and its associated data are classified into the empirical sample library as a valid intervention case. If the hidden danger risk score does not decrease or increases after the intervention, the node and its associated data are classified into the abnormal sample library as a difficult-to-predict case. After every fifty intervention actions, the collected samples for each intervention node are arranged in chronological order, and the empirical samples and abnormal samples are classified and stored separately. The feature vectors extracted from all empirical and abnormal samples are normalized according to a unified time step length and node number format. Incremental training is performed, using mini-batch stochastic gradient descent to update the parameters of the spatiotemporal neural network model. The prediction error is gradually calculated using each mini-batch of sample data as input, and the network weights and biases are adjusted through backpropagation. During training, the node number mapping relationship is maintained, and the time series order is strictly increasing. After all incremental samples are trained, the updated model replaces the original spatiotemporal neural network model.

[0014] Another object of the present invention is to provide a standardized closed-loop safety management and control system for the entire process of an electric power enterprise, which can be used to trigger preventive intervention based on the hierarchical classification of hidden danger prediction results through an optimization module, and the data can be refluxed and updated after the intervention is completed, thereby solving the problems of insufficient multi-source data fusion and time series synchronization in existing power safety management and control methods, the problem of lack of dynamic correction of intervention feedback in hidden danger trend prediction, the problem of rigid risk classification and response mechanism, and the problem of how to achieve full-process standardization and dynamic closed-loop optimization of power enterprise safety management and control.

[0015] As a preferred solution of the full-process standardized closed-loop safety management and control system for electric power enterprises described in the present invention, it includes: a data acquisition and preprocessing module, a modeling and calculation module, and an optimization module.

[0016] The data acquisition and preprocessing module is used to acquire image features and sensor features in real time based on multi-source heterogeneous data and perform unified coding.

[0017] The modeling and calculation module is used to construct an operation and equipment node graph based on a synchronous time series, and adopt a spatiotemporal neural network to perform hidden danger trend modeling.

[0018] The optimization module is used to trigger preventive intervention based on the hierarchical classification of hidden danger prediction results, and the data is updated after the intervention is completed.

[0019] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program as a step to implement a standardized closed-loop safety management and control method for the entire process of an electric power enterprise.

[0020] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a standardized closed-loop safety management and control method for the entire process of an electric power enterprise.

[0021] The beneficial effects of the present invention are as follows: the standardized closed-loop safety management and control method for the entire process of electric power enterprises provided by the present invention collects image features and sensor features in real time based on multi-source heterogeneous data, and executes unified coding to ensure high consistency of basic data for hidden danger prediction and safety management throughout the entire process, eliminates information islands, and improves the model's sensitivity and response accuracy to actual changes in the operating environment.

[0022] Based on the synchronous time series, the operation and equipment node diagrams are constructed, and the space-time neural network is used to model the hidden danger trends to establish the risk prediction capability of dynamic evolution, so that the safety management and control system can identify potential high-risk nodes and key transmission paths in advance, significantly improving the foresight and systematic nature of hidden danger prevention and control.

[0023] Preventive intervention is triggered based on the hierarchical classification of hidden danger prediction results. After the intervention is completed, the data is refluxed and updated to form a safety management and control mechanism for power companies with dynamic optimization and standardized closed loop throughout the entire process. This has achieved the continuous evolution of hidden danger identification - intervention - feedback - optimization, greatly improving the stability, intelligence and reliability of the system in complex operating environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 This is an overall flow chart of a standardized closed-loop safety management and control method for the entire process of an electric power enterprise provided in the first embodiment of the present invention.

[0026] Figure 2 A system flow chart of a standardized closed-loop safety management and control method for the entire process of an electric power enterprise provided in the second embodiment of the present invention. DETAILED DESCRIPTION

[0027] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0028] Example 1, reference Figure 1 , which is an embodiment of the present invention, provides a standardized closed-loop safety management and control method for the entire process of a power enterprise, including:

[0029] S1: Collect image features and sensor features in real time based on multi-source heterogeneous data and perform unified encoding.

[0030] Deploy video capture equipment to continuously collect video stream data from power company operation areas and extract key frames. Based on these key frames, detect and extract image features, including the safety clothing status of workers, abnormal operation movements, and changes in the working environment structure.

[0031] Furthermore, after pre-processing, the extracted key frames are used to identify the workers in the picture and detect their safety clothing status, including whether they are wearing a safety helmet and safety protective clothing, and to detect whether there are any abnormalities in the work actions, such as illegal operations, erroneous actions, etc. At the same time, structural change features in the working environment are extracted, such as abnormal phenomena such as increased obstacles and equipment displacement.

[0032] In power operation environments, environmental parameters and key equipment operating status parameters are collected and aligned along a unified timeline. Different sensor features are normalized and mapped to a unified feature space dimension, which is then combined with image features to form a multimodal fusion input.

[0033] Furthermore, environmental parameters include temperature, humidity, smoke concentration, vibration amplitude, etc., and key equipment operating status parameters include speed, load, current, voltage, etc. These sensor data are collected in real time according to their respective set sampling frequencies, and a timestamp in a unified format is attached during collection to ensure that the data can be aligned in the time dimension.

[0034] It should be noted that during data preprocessing, all collected sensor features are normalized to a uniform interval (typically 0 to 1) within the reasonable range of each parameter to eliminate scale differences between different physical quantities. These normalized sensor features are then synchronized with image features obtained through image processing using a unified time base, forming a unified multimodal feature input dataset.

[0035] S2: Build job and equipment node graphs based on synchronized time series, and use spatiotemporal neural networks to model hidden danger trends.

[0036] The power operation log text is collected, and the operation unit, personnel ID, equipment ID, and operation category are extracted as node entities. Graph edges are established between nodes with physical associations and operation collaboration relationships. Physical proximity relationships are automatically derived from the equipment layout diagram, and collaboration relationships are inferred from the operation process log. Node features include real-time sensor data features, image anomaly detection features, and operation text encoding features. Each node feature is combined into a feature vector sequence based on the node number and time step. The entire node graph structure is dynamically updated with operation schedules and environmental changes, and the adjacency matrix is ​​automatically and synchronously adjusted when a node changes.

[0037] After the graph structure is constructed, a spatial convolutional neural network is used to extract spatial features based on the node graph. Graph convolution is used to perform weighted aggregation of features from adjacent nodes, taking each node as the center and combining them with the connectivity relationships determined by the adjacency matrix. The aggregation weights are determined by the edge weights between the nodes. Subsequently, in the temporal dimension, the time series feature vectors corresponding to each node are processed using a long short-term memory (LSTM) temporal modeling network to extract the dynamic features of each node over time. The input order of the time series is strictly increasing, with no skips or missing data. Through this spatiotemporal modeling process, spatial feature vectors and temporal dynamic feature vectors are generated, respectively. These two vectors are concatenated in the feature dimension to form a joint feature representation for each node. Finally, the joint feature vector for each node is input into the hidden danger score prediction layer, which outputs a hidden danger risk score between 0 and 1, where 0 represents no risk and 1 represents extremely high risk. The predicted output covers the risk trend at each time step over the next three days. This risk score serves as an important basis for subsequent intervention decisions and is periodically updated with each new data input.

[0038] It should be noted that S2 constructs a graph of operations and equipment nodes by synchronizing time series. Combining graph convolution with time series modeling, it extracts spatial correlations and temporal evolution characteristics between nodes to form a hazard trend prediction model. This step dynamically captures changes in the operating environment, improving the accuracy and foresight of hazard identification and providing real-time decision-making for subsequent intervention measures.

[0039] S3: Trigger preventive intervention based on the classification of hidden danger prediction results, and update the data after the intervention is completed.

[0040] For nodes with risk scores above the high-risk threshold, a suspension order is generated, an on-site safety re-inspection is arranged, and the re-inspection records are filed. For nodes with risk scores in the medium-risk range, an order to increase the frequency of manual inspections and a limited operation approval process is generated, with approval records forming a complete track within the system.

[0041] Furthermore, in the present invention, the high-risk threshold is set to 0.7, the medium-risk interval is set to 0.4-0.7, and the rest are low-risk nodes.

[0042] After implementing the intervention measures, the image features, sensor features and job log text of the intervention node and its adjacent nodes are re-collected and compared with the data collected before the intervention. The difference features of each node before and after the intervention are extracted based on the node number and time series order, and the actual hidden danger evolution trend is dynamically marked.

[0043] Furthermore, after implementing an intervention, we recollect image features, sensor features, and job log text for the intervention node and all its first-order neighboring nodes, maintaining consistency with the pre-intervention sampling data in terms of timestamp and node number. We encode the feature data before and after the intervention into vectors, aligning them based on node number and time steps, and calculate the change in hidden danger risk score before and after the intervention.

[0044] When the hidden danger risk score decreases after the intervention compared with that before the intervention, the node and its associated data are classified into the experience sample library as an effective intervention case. When the hidden danger risk score does not decrease or increase after the intervention, the node and its associated data are classified into the abnormal sample library as a difficult-to-predict case.

[0045] Define the hidden danger score of each node before intervention as r m,pre , the hidden danger score after intervention is r m,post , m represents the mth node, and the formula for calculating the score change difference is:

[0046] Δr m =r m,pre -r m,post

[0047] If Δr m If the value is greater than 0, it means the intervention is effective, and the node and its data are included in the experience sample library. Otherwise, it means the intervention is ineffective or the risk is increased, and the node and its data are included in the abnormal sample library.

[0048] An optimal solution for dynamically marking the evolution trend of actual hidden dangers specifically includes: based on the node number and the corresponding time series, the method of dynamically marking the evolution trend of actual hidden dangers is to compare the change trend of the hidden danger score sequence at different time steps of the same node, and use a sliding window linear fitting method to perform a first-order linear regression on a data segment with a window length of three steps. If the regression slope is less than 0, it is marked as a downward risk trend; if the regression slope is greater than 0, it is marked as an upward risk trend; if the absolute value of the regression slope is less than 0.5, it is marked as a stable risk trend.

[0049] After completing every fifty intervention actions, the collected samples from each intervention node are arranged in chronological order, and empirical samples and abnormal samples are categorized and stored separately. The feature vectors extracted from all empirical and abnormal samples are standardized according to a unified time step length and node numbering format. Incremental training is performed, using mini-batch stochastic gradient descent to update the parameters of the spatiotemporal neural network model. Using each mini-batch of sample data as input, the prediction error is gradually calculated, and the network weights and biases are adjusted through backpropagation.

[0050] Furthermore, the incremental model training process specifically involves normalizing the feature vectors extracted from the empirical and abnormal samples to a uniform scale and inputting them into the mini-batch training process. Mini-batch stochastic gradient descent (Mini-batch SGD) is used to update the spatiotemporal neural network model parameters. The detailed calculation scheme is as follows:

[0051] Define the sample set of each small batch as Q b,t , where b represents the batch number, t represents the tth round of training, and the model parameter is Θ t , the update formula is:

[0052]

[0053] Among them, Θ t+1 X is the model parameter at the t+1th round of training. u,b,t is the feature vector of the u-th sample in the small batch. J(Θ t ,X u,b,t ) is the parameter Θ t and sample X u,b,t Calculate the loss function value. ξ u,b,t Represents the dynamic weight given to the u-th sample according to the change of risk score. u,b,t is the modulation factor constructed based on the sampling time difference. t is the learning rate of the tth round. |Q b,t | represents the number of mini-batch samples.

[0054] Furthermore, in order to be more sensitive to high-risk node predictions, the loss function cannot only use the ordinary mean square error (MSE) or cross entropy (CE). The present invention uses the basic error term to fit the hidden danger score, the risk weighting term to emphasize the importance of high-risk samples, and the time evolution regularization term to encourage the model output trend to be continuous and smooth, which can be expressed as:

[0055]

[0056] Among them, μ u,b,t The dynamic risk weight corresponding to the u-th sample in the b-th small batch of training in the t-th round. The model of the b-th mini-batch and the u-th sample is trained in the t-th round to predict the hidden danger risk score. The true hidden danger risk score of the u-th sample in the b-th small batch of the t-th training round. ρ u,b,t is the time evolution smoothing weight. τ u,b,t The time step number for the u-th sample in the b-th mini-batch of the t-th training round.

[0057] During the training process, the node number mapping relationship is kept unchanged, and the time series order is strictly increased until all incremental samples are trained, and the updated model replaces the original spatiotemporal neural network model.

[0058] It should be noted that S3 manages node risk grading based on hazard prediction results, dynamically collects node features before and after intervention, extracts risk score changes and categorizes them into empirical and abnormal samples, uses sliding window regression to annotate hazard evolution trends, and then dynamically updates the spatiotemporal neural network model through small-batch incremental training using standardized feature vectors. Compared to existing technologies that rely solely on static data modeling, this step uses the actual effects of intervention to flow back, enabling the model to adaptively optimize based on the actual evolution results of the field. This overcomes the problems of traditional models' inability to respond to environmental changes in a timely manner and the degradation of prediction accuracy over time, significantly improving the real-time, accuracy, and adaptability of safety hazard trend prediction.

[0059] Example 2, reference Figure 2 , which is an embodiment of the present invention, provides a standardized closed-loop safety management and control system for the entire process of an electric power enterprise, including a data acquisition and preprocessing module 100, a modeling and calculation module 200, and an optimization module 300.

[0060] S4: The data acquisition and preprocessing module 100 is used to acquire image features and sensor features in real time based on multi-source heterogeneous data and perform unified coding.

[0061] The data acquisition and preprocessing module 100 includes an image data acquisition submodule 101 and a sensor data normalization submodule 102 .

[0062] Furthermore, the image data acquisition submodule 101 deploys high-resolution monitoring equipment to continuously capture video stream data from the power company's operating area at a rate of 15 frames per second, extracting key frames every five seconds. These extracted key frames undergo preprocessing to identify characteristics such as worker safety attire, abnormal operating movements, and changes in the operating environment. The sensor data normalization submodule 102 collects parameters such as temperature, humidity, vibration, smoke concentration, and equipment operating status from the operating environment. These parameters are then standardized according to their actual value ranges and synchronized with timestamps for subsequent unified feature fusion.

[0063] It should be noted that the image data acquisition submodule 101 is the starting point for the data acquisition and preprocessing module 100, providing real-time image feature information of the work area. The sensor data normalization submodule 102 performs standardized preprocessing on multi-source environmental and equipment data to ensure synchronization with image features, forming a unified data input foundation and providing standardized input for the subsequent construction of the work and equipment node graph in the modeling and calculation module 200.

[0064] S5: The modeling and calculation module 200 is used to construct an operation and equipment node diagram based on a synchronous time series, and use a spatiotemporal neural network to perform hidden danger trend modeling.

[0065] The modeling and calculation module 200 includes a node graph structure construction submodule 201 and a hidden danger trend prediction submodule 202 .

[0066] Furthermore, the node graph structure construction submodule 201 is used to extract information such as work unit, personnel identification, and equipment identification based on the image features, sensor features, and work log text output by the data acquisition and preprocessing module 100, establish node entities, and establish graph connection edges based on the physical proximity of equipment and work collaboration relationships, dynamically generating work and equipment node graphs. The hidden danger trend prediction submodule 202 is used to extract spatial features between nodes based on the node graph structure using a spatial convolutional network. It then combines a long short-term memory network or a time series modeling network based on a self-attention mechanism to extract the dynamic features of nodes as they evolve over time, ultimately outputting a hidden danger risk score for each node within the next three days.

[0067] It should be noted that the node graph structure construction submodule 201 is the foundation of the modeling and calculation module 200, ensuring a clear logical structure between work units, equipment, and personnel nodes. The hidden danger trend prediction submodule 202 accurately captures the evolution trend of hidden dangers through spatiotemporal feature fusion modeling, providing a risk classification basis for the subsequent optimization module 300.

[0068] S6: The optimization module 300 is used to trigger preventive intervention based on the hierarchical classification of hidden danger prediction results, and the data is refluxed and updated after the intervention is completed.

[0069] The optimization module 300 includes a risk grading intervention submodule 301 and an intervention effect reflux update submodule 302 .

[0070] Furthermore, the risk-grading intervention submodule 301 is used to implement graded management based on the hidden danger trend prediction score, with the high-risk threshold set at 0.7 and the medium-risk range at 0.4 to 0.7. These submodules generate instructions to suspend operations or limit the operation approval process, respectively, and form a complete approval trajectory record. The intervention effect reflow update submodule 302 is used to recollect data on the intervention node and its adjacent nodes after the intervention is completed, compare the changes in hidden danger risk scores before and after the intervention, and classify node samples as empirical samples or abnormal samples based on the change trend. After completing every fifty intervention closed loops, the spatiotemporal neural network is incrementally trained and optimized using a small-batch stochastic gradient descent method to update the model parameters.

[0071] It should be noted that the risk-level intervention submodule 301 dynamically adjusts operational processes and safety measures based on hazard levels. The intervention effect feedback update submodule 302 implements adaptive model evolution based on the actual effects of interventions, ensuring that the hazard prediction system maintains high accuracy and environmental adaptability.

[0072] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program code.

[0073] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0074] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0075] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having logic gate circuits for implementing logical functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc. It should be noted that the above embodiments are merely illustrative of the technical solutions of the present invention and are not intended to be limiting. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced with equivalents without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications should be encompassed by the claims of the present invention.

Claims

1. A method for standardized closed-loop safety management and control of the entire process of an electric power enterprise, characterized in that: include: Collect image features and sensor features in real time based on multi-source heterogeneous data and perform unified encoding; Build operation and equipment node graphs based on synchronized time series, and use spatiotemporal neural networks to model hidden danger trends; Trigger preventive intervention based on the classification of hidden danger prediction results, and update data after the intervention is completed; The data reflux update after the intervention is completed includes, after the intervention measures are executed, re-collecting the image features, sensor features and job log text of the intervention node and its adjacent nodes, comparing the data changes before the intervention, extracting the difference features before and after the intervention, and dynamically marking the actual hidden danger evolution trend according to the difference features; for nodes whose hidden danger risk scores decrease after the intervention, they are recorded as effective intervention cases and stored in the experience sample library; for nodes whose risks do not decrease or increase after the intervention, they are marked as difficult-to-predict samples and stored in the abnormal sample library; after completing each fifty intervention closed loops, the experience samples and abnormal samples are summarized, the spatiotemporal neural network model parameters are updated, and the hidden danger prediction ability is optimized through incremental learning.

2. The method for standardized closed-loop safety management and control of the entire process of a power enterprise according to claim 1, characterized in that: The real-time acquisition of image features and sensor features based on multi-source heterogeneous data includes: Deploy video acquisition equipment to continuously collect video stream data from the power company's operating area and extract key frames; Image features are detected and extracted based on the extracted key frames, including the safety clothing status of the workers, abnormal information of working actions and changes in the working environment structure.

3. The method for standardized closed-loop safety management and control of the entire process of a power enterprise according to claim 1 or 2, characterized in that: The execution of unified coding includes: In the power operation environment, environmental parameters and key equipment operating status parameters are collected and all data are aligned according to a unified timeline; Different types of sensor features are normalized and mapped to a unified feature space dimension, and then combined with image features to form a multimodal fusion input.

4. The method for standardized closed-loop safety management and control of the entire process of a power enterprise according to claim 3, characterized in that: The construction of the job and device node graph based on the synchronized time series includes: Collect power operation log text, extract operation unit, personnel identification, equipment identification and operation category as node entities; Graph edges are established between nodes that have physical associations and operational collaboration relationships. Physical proximity relationships are automatically derived from the equipment layout diagram, and collaboration relationships are inferred based on the operational process log. Node features include real-time sensor data features, image anomaly detection features, and job text encoding features. Each node feature is combined into a feature vector sequence according to the node number and time step. All node graph structures are dynamically updated with job arrangements and environmental changes, and the adjacency matrix is ​​automatically and synchronously adjusted when the node changes.

5. The method for standardized closed-loop safety management and control of the entire process of a power enterprise according to any one of claims 1, 2 or 4, characterized in that: The use of spatiotemporal neural network to model hidden danger trends includes: Based on the constructed job and device node graph, a spatial convolutional network is used to perform graph convolution operations on adjacent nodes to extract the spatial feature relationships between nodes. In the time dimension, a long short-term memory network or a time series modeling network based on the self-attention mechanism is used to extract the dynamic evolution pattern of each node over time; The spatial convolution output and the temporal modeling output are feature spliced, and the hidden danger risk score of each node in the next three days is finally output. The risk score ranges from 0 to 1.

6. The method for standardized closed-loop safety management and control of the entire process of a power enterprise according to claim 5, characterized in that: The preventive intervention triggered by the classification of hidden danger prediction results includes: For nodes with risk scores greater than the high-risk threshold, a suspension instruction is generated, an on-site safety re-inspection is arranged, and the re-inspection records are filed; For nodes whose risk scores are in the medium-risk range, instructions to increase the frequency of manual inspections and limit the operation approval process are generated, and the approval records form a complete track within the system.

7. The method for standardized closed-loop safety management and control of the entire process of a power enterprise according to any one of claims 1, 2, 4 or 6, characterized in that: The data reflux update after the intervention is completed includes: After implementing the intervention measures, the image features, sensor features, and operation log text of the intervention node and its adjacent nodes are recollected and compared with the data collected before the intervention. The difference features of each node before and after the intervention are extracted based on the node number and time series order, and the actual hidden danger evolution trend is dynamically marked; When the hidden danger risk score after intervention decreases compared with that before intervention, the node and its associated data will be included in the experience sample library as an effective intervention case. When the hidden danger risk score does not decrease or increase after intervention, the node and its associated data are classified into the abnormal sample library as a difficult-to-predict case; After completing fifty intervention actions, the collected samples of each intervention node are arranged in chronological order, and the experience samples and abnormal samples are classified and stored separately; The feature vectors extracted from all empirical samples and abnormal samples are standardized according to the unified time step length and node number format; Perform incremental training and use the mini-batch stochastic gradient descent method to update the parameters of the spatiotemporal neural network model. Take each mini-batch of sample data as input, gradually calculate the prediction error, and adjust the network weights and biases through backpropagation. During the training process, the node number mapping relationship is kept unchanged, and the time series order is strictly increased until all incremental samples are trained, and the updated model replaces the original spatiotemporal neural network model.

8. A standardized closed-loop safety management and control system for the entire process of an electric power enterprise, characterized by: It includes a data acquisition and preprocessing module (100), a modeling and calculation module (200), and an optimization module (300); The data acquisition preprocessing module (100) is used to acquire image features and sensor features in real time based on multi-source heterogeneous data and perform unified coding; The modeling and calculation module (200) is used to construct an operation and equipment node graph based on a synchronous time series, and to perform hidden danger trend modeling using a spatiotemporal neural network; The optimization module (300) is used to trigger preventive intervention based on the hierarchical classification of hidden danger prediction results, and to update data after the intervention is completed.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for standardized closed-loop safety management and control of the entire process of an electric power enterprise as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for standardized closed-loop safety management and control of the entire process of an electric power enterprise as described in any one of claims 1 to 7 are implemented.

Images