An artificial intelligence-based power engineering safety construction supervision system

Abstract

The application discloses an artificial intelligence-based power engineering safety construction supervision system and relates to the technical fields of power engineering construction safety supervision and artificial intelligence; solves the problems that traditional supervision methods rely on manual work, are low in efficiency and are difficult to realize real-time accurate early warning and tracing; the application realizes real-time collection of on-site audio, vibration, thermal imaging and electromagnetic signals and evaluation of fidelity through a multi-modal edge perception module; realizes abnormal fusion reasoning and cause-effect tracing through a cause-effect cognition and diagnosis module, and outputs a research and judgment report; realizes abnormal visualization and virtual pre-operation simulation through a digital twin collaborative interaction module which constructs a synchronous three-dimensional model; and generates distributed safety state verification data packets according to risk disposal effects through a closed-loop verification and data tracing module. The application realizes intelligent monitoring of a power engineering construction site, accurate risk diagnosis, visualized early warning and closed-loop control of incentive safety management.

Description

Technical Field

[0001] This invention belongs to the field of power engineering construction safety supervision and artificial intelligence technology, specifically an artificial intelligence-based power engineering safety construction supervision system. Background Technology

[0002] Power engineering construction sites are characterized by complex environments, numerous equipment, and overlapping processes, involving various high-risk scenarios such as high-voltage, high-altitude, and live-line work. Traditional safety management heavily relies on regulations, personnel experience, and regular inspections. With the development of sensor technology, the Internet of Things (IoT), and communication technologies, construction sites have gradually introduced online monitoring methods such as video surveillance and environmental sensing, achieving remote visualization and data collection of certain conditions. In recent years, artificial intelligence technologies, especially machine learning and pattern recognition, have demonstrated enormous potential in areas such as industrial anomaly detection and risk prediction, providing new technological pathways for the intelligent improvement of construction safety.

[0003] Meanwhile, digital twin technology, as a bridge connecting the physical and information worlds, has been widely applied in fields such as industrial manufacturing and smart cities. By constructing virtual mappings of physical entities, it enables state simulation, process extrapolation, and interactive intervention, providing a new paradigm for real-time monitoring and decision support of complex systems.

[0004] Currently, technological development is evolving from single-point monitoring to collaborative sensing, from post-event tracing to pre-event prediction and in-event intervention, and from passive alarm to proactive early warning and closed-loop management. How to systematically integrate cutting-edge sensing fusion, causal inference, virtual simulation, and reliable verification technologies to build an integrated regulatory platform capable of real-time and accurate sensing, intelligent diagnosis and tracing, intuitive interactive guidance, and closed-loop management has become an important direction for improving the inherent safety level of power engineering.

[0005] The following problems exist in the existing technology: Existing monitoring systems mostly use single or a few types of sensors, which cannot comprehensively characterize complex construction risks. Sensors are susceptible to environmental interference (such as noise, electromagnetic fields, temperature, and humidity), resulting in severe signal distortion, high false alarm and false alarm rates, and insufficient early warning accuracy. Existing methods are mostly based on threshold alarms or simple pattern recognition, which can only determine that "an anomaly has occurred" but cannot answer key questions such as "why is it abnormal," "what is the root cause," and "how might it develop subsequently." They lack the ability to deeply explore and deduce the underlying physical causal mechanisms of abnormal events. Traditional early warning systems often present information in the form of sound, light, or simple pop-up messages, which are abstract and not intuitive. On-site personnel find it difficult to quickly locate risk points, understand the risk evolution path, and conduct safety verification and rehearsals before taking actual measures. Decision-making relies on personal experience, which poses safety hazards. Traditional safety supervision processes often stop at alarms or handling, lacking quantitative verification of the effectiveness of risk handling, systematic assessment, and reliable evidence storage and traceability of data throughout the process, making it difficult to form a management loop for continuous improvement. Summary of the Invention

[0006] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes an artificial intelligence-based power engineering safety construction supervision system to solve the above-mentioned technical problem.

[0007] The first aspect of this invention provides an artificial intelligence-based power engineering safety construction supervision system, comprising the following modules: Multimodal edge sensing module: Deploy sensor arrays at key risk nodes in power engineering construction sites to simultaneously collect audio signals, vibration signals, thermal imaging signals and electromagnetic environment signals, dynamically evaluate the fidelity of each sensing channel, and output primary anomaly signal vectors. Causal cognition and diagnosis module: Based on the primary abnormal signal vector and the dynamic fidelity of each sensor channel, the primary abnormal signal vector is fused and reasoned through a causal fusion algorithm to generate a structured description of the abnormal event; combined with the built-in physical knowledge graph and historical equipment data, the abnormal event is traced back to its causal origin and the impact chain is deduced, and an analysis report containing the root cause, risk level, handling suggestions and causal chain is output, where the causal chain represents the nodes of the abnormal event and the causal transmission relationship between them; Digital Twin Collaborative Interaction Module: Constructs and maintains a 3D digital twin model that is updated synchronously with the power engineering site; generates a causal potential energy field in the model and visualizes anomalies based on the causal chain in the assessment report; performs virtual pre-operation simulation on the proposed handling measures and verifies the safety margin; generates an interactive early warning task containing real-world guidance and pushes the task to the real-world terminals of relevant personnel; receives operation response data, handling confirmation, and on-site verification information from the real-world terminals. Closed-loop verification and data traceability module: Based on the causal chain and risk level in the assessment report, a risk transmission diagram is established. Combined with the on-site verification information, the final risk decay value and risk entropy difference caused by the interactive early warning task are calculated. The risk decay value and risk entropy difference are fused through the risk entropy reduction verification model to generate an effectiveness index and encapsulate it into a safety status verification data packet, which is then stored in a distributed storage network.

[0008] Preferably, the multimodal edge sensing module includes: Sensor array deployment and acquisition unit: The sensor array is deployed at key risk nodes in the power engineering construction site. Each sensing unit in the sensor array integrates an audio sensor, a vibration sensor, a thermal imaging sensor and an electromagnetic sensor, forming multiple sensing channels for synchronous acquisition of audio, vibration, thermal imaging and electromagnetic environmental signals. Dynamic fidelity calculation unit: Calculates the dynamic fidelity of each sensing channel in real time based on the signal acquired by each sensing channel. The calculation formula is as follows: in, For signal quality factor, This is the sensor coupling state factor. As an environmental noise pollution factor, , , , These are the parameters of the pre-trained model; Anomaly Scoring and Vector Construction Unit: Based on the signals acquired by each sensor channel, calculate the anomaly score for each channel. ; Exception scoring Its corresponding dynamic fidelity Weighted fusion yields a weighted outlier score. Construct a primary anomalous signal vector, which contains at least the weighted anomaly scores of each sensing channel. and the dynamic fidelity of each sensor channel .

[0009] Preferably, in the causal cognition and diagnosis module, based on the primary abnormal signal vector and the dynamic fidelity of each sensing channel, the primary abnormal signal vector is fused and inferred using a causal fusion algorithm to generate a structured description of the abnormal event, including: Causal fusion unit: based on the weighted anomaly score of each sensor channel This forms a weighted anomaly score sequence, and the fidelity-weighted transfer entropy between any two sensing channels X and Y is calculated. Its calculation formula is: , where, and These are the dynamic fidelity coefficients for channels X and Y, respectively. The transit entropy is calculated based on the weighted anomaly score sequence. For time delay parameters; Cause-effect graph construction unit: Determines the optimal time delay parameters and causal direction by maximizing fidelity weighted transfer entropy, and constructs a dynamic cause-effect graph by combining preset physical causal rules and device topology connection relationships; Event description generation unit: Based on the dynamic cause-effect graph, it identifies the root cause node as the origin of the anomaly and the causal propagation path from the root cause node to other nodes; and generates a structured anomaly event description based on the root cause node and the causal propagation path.

[0010] Preferably, in the causal cognition and diagnosis module, the built-in physical knowledge graph and historical equipment data are combined to perform causal tracing and impact chain deduction of abnormal events, and output an assessment report including the root cause, risk level, handling suggestions, and causal chain. The causal chain represents the nodes of the abnormal event and the causal transmission relationship between them, including: Fault mode identification unit: Maps the root cause nodes in the structured abnormal event description to the corresponding physical entities in the built-in physical knowledge graph, and obtains the set of fault modes corresponding to the entity through multi-hop reasoning; Posterior probability calculation unit: for each fault mode The posterior probability is calculated based on Bayes' theorem. Where U is the set of observed multimodal anomaly evidence, and H is the historical data of the equipment; Impact Chain Inference Unit: Based on the failure mode with the highest posterior probability, it uses physical rules in the physical knowledge graph to infer the impact chain. The impact chain represents a series of consequence events caused by an anomaly starting from the root cause node and passing through the corresponding correlation. During the inference process, the uncertainty of the current load and environmental parameters is considered, and multiple risk evolution trajectories are generated through Monte Carlo simulation. Each risk evolution trajectory is a random realization of an impact chain. Risk entropy calculation and classification unit: Calculate risk entropy based on multiple risk evolution trajectories. The risk level is dynamically classified, and the calculation formula is as follows: ,in, Let k be the probability of the k-th influence chain occurring. The level of the most severe consequence caused by the k-th influence chain. For the k-th influence chain to evolve from the current state to... The estimated time for the consequences of the level; The assessment report generation unit generates an assessment report containing the root cause, risk level, handling recommendations, and causal chain based on risk entropy and its risk level classification, the failure mode with the highest posterior probability, the selected key impact chain as the inferred causal chain, and the handling rules in the physical knowledge graph; the causal chain consists of abnormal event nodes and the causal transmission relationship between them.

[0011] Preferably, the digital twin collaborative interaction module includes constructing and maintaining a three-dimensional digital twin model that is synchronously updated with the power engineering site, comprising: Twin Model Construction Unit: Constructs a three-dimensional digital twin model corresponding to the power engineering site. The model includes a geometric model layer representing the geometric structure of the site and a physical model layer representing the physical laws of the equipment. The physical model layer is a parameterized physical simulation model embedded in the key equipment. The parameters of the parameterized physical simulation model are obtained by Bayesian parameter inversion based on the historical operating data of the key equipment in the power engineering site and continuously calibrated. Real-time synchronization unit: This unit achieves real-time synchronization between the 3D digital twin model and the physical site of the power engineering project through a state prediction-correction synchronization algorithm. The algorithm includes: Prediction Step: Based on the parametric physical simulation model and current commands from the power engineering field control system and status Input, predict the state at the next time step. ,in, This refers to the process noise term; Calibration step: When the measured data of the received sensor signal is obtained. At that time, the confidence weights of the prediction results of the parametric physical simulation model are calculated. Confidence weights of measured data : ,in To predict the error covariance, To measure the noise variance; Hybrid update step: The predicted state and the measured data are merged according to the calculated weights to update the current state. .

[0012] Preferably, in the digital twin collaborative interaction module, based on the causal chain in the assessment report, a causal potential energy field is generated in the model and anomalies are visualized; virtual pre-operation simulation is performed on the handling suggestions and the safety margin is verified, an interactive early warning task containing real-world guidance is generated, and the task is pushed to the real-world terminals of relevant personnel; operation response data, handling confirmation, and on-site verification information are received from the real-world terminals, including: Causal Chain Visualization Unit: Maps the causal chain in the assessment report to a causal potential energy field in the three-dimensional space of the three-dimensional digital twin model. For each anomalous event node m in the causal chain, its spatial location is defined. Causal potential energy generated at the location : in, The amplitude is determined by the risk level of this abnormal event. To determine the location of this anomalous event in the 3D digital twin model, For the scope of influence, For dynamic evolution frequency, For phase; The causal potential energy field is the vector superposition of the potential energies of all anomalous event nodes. ,in It is a causal direction unit vector, and is visualized in a three-dimensional digital twin model using volume rendering techniques in computer graphics; Virtual pre-operation simulation unit: In a 3D digital twin model, virtual pre-operation simulations are performed on the handling recommendations in the assessment report, and safety margin indicators are calculated. To assess operational safety: ,in, For the nth safety indicator, For safety limits, The current value, This is a normal value; if If the value is below the preset threshold, the handling plan will be adjusted and the simulation will be repeated until the safety margin index meets the requirements. Interactive task generation and push unit: Based on the verified handling plan, generate an interactive early warning task containing step-by-step augmented reality guidance and push it to the real terminals of relevant personnel; Feedback Receiving and Verification Unit: After the interactive early warning task is pushed, it receives operation response data, handling confirmation, and on-site verification information from the actual terminal; based on the on-site verification information, it calculates the verification confidence score. : in, The first on-site verification Item index value, These are the predicted values ​​from the twin model. To allow for error, For weights.

[0013] Preferably, in the closed-loop verification and data traceability module, a risk transmission diagram is established based on the causal chain and risk level in the assessment report, and the final risk attenuation value caused by the interactive early warning task is calculated in conjunction with the on-site verification information, including: Risk Graph Modeling Unit: Models the causal chains in the assessment report into a risk transmission graph. =(V,E,W), where node V represents an anomalous event node in the causal chain, edge E represents the causal transmission relationship, and edge weight W represents the risk transmission intensity. Risk attenuation calculation unit: node After completing the suggested measures in the interactive early warning task, calculate the effect of these measures on the overall risk mitigation, i.e., the risk attenuation value. for: in, Represents a node The set of all downstream nodes; For nodes The inherent risk intensity is determined by the risk level in the assessment report. The probability of occurrence of the influence chain to which this node belongs. Decision, that is , From arrive The set of all paths; Let e ​​be the blocking efficiency of edge e; Spatiotemporal calibration unit: for risk attenuation values Spatiotemporal calibration is performed to obtain the final risk attenuation value. : ,in, The time decay factor, This represents the environmental difficulty factor.

[0014] Preferably, in the closed-loop verification and data traceability module, the risk entropy difference caused by the interactive early warning task is calculated; the final risk decay value and the risk entropy difference are fused through a risk entropy reduction verification model to generate an effectiveness index and encapsulate it into a security status verification data packet, which is then stored in a distributed storage network, including: Risk entropy calculation unit: calculates the risk entropy before the disposal measures are completed. , and the risk entropy after completion Where S is the set of risk states, and These represent the probability of risk state s occurring before and after the completion of the disposal measures; Risk Entropy Difference Calculation Unit: Calculates the risk entropy difference. ; Confidence verification calculation unit: based on the risk entropy difference and the final risk attenuation value The effectiveness index of the response measures is calculated using a risk entropy reduction verification model. : ,in, , To contribute weighting coefficients, As the benchmark risk level; Data storage generation unit: Generates a security status verification data package containing an effectiveness index, final risk decay value, risk entropy difference, associated interactive early warning task identifier, personnel identifier, and timestamp; Network Record Unit: The security status verification data packet is recorded in the distributed storage network using a practical Byzantine fault-tolerant consensus algorithm to form a security contribution record.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention deploys an integrated multimodal sensor array to simultaneously collect multi-dimensional signals of sound, vibration, heat, and electromagnetic fields. It also introduces a dynamic fidelity calculation model to perform real-time quality assessment of each sensing channel, and weights and fuses fidelity with anomaly scores to effectively filter out interference. This significantly improves the reliability and comprehensiveness of primary anomaly signals, laying a solid foundation for subsequent accurate diagnosis. This invention constructs a dynamic causal graph based on fidelity-weighted transfer entropy, and combines it with physical knowledge graphs and historical equipment data to achieve causal tracing and impact chain deduction of abnormal events. It uses Bayesian inference to identify the most likely failure modes and uses Monte Carlo simulation to predict multiple risk evolution trajectories, ultimately generating an assessment report that includes the root cause, causal chain, and quantified risk level, thus upgrading the early warning from "superficial alarm" to "root cause diagnosis and trend prediction". This invention uses a three-dimensional digital twin model that is synchronized with the site in real time to visualize the abstract causal chain through a causal potential energy field, thereby achieving precise risk positioning and intuitive presentation. By performing virtual pre-operation simulations on disposal suggestions and calculating safety margins, it provides on-site personnel with visual and verifiable augmented reality guidance, improving the scientific nature of decision-making and operational safety. This invention establishes a risk transmission diagram and calculates the difference between risk decay value and risk entropy to quantitatively assess the effectiveness of response measures, generating a safety status verification data package containing an effectiveness index. By storing this data package in a distributed storage network, the effectiveness of the response is measurable, verifiable, and traceable throughout the entire process, forming a complete management closed loop of "perception-diagnosis-interaction-response-verification," which effectively improves the enforceability of safety regulations. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Please see Figure 1 This invention is an artificial intelligence-based power engineering safety construction supervision system, comprising the following modules: Multimodal edge sensing module: Deploy sensor arrays at key risk nodes in power engineering construction sites to simultaneously collect audio signals, vibration signals, thermal imaging signals and electromagnetic environment signals, dynamically evaluate the fidelity of each sensing channel, and output primary anomaly signal vectors. Causal cognition and diagnosis module: Based on the primary abnormal signal vector and the dynamic fidelity of each sensor channel, the primary abnormal signal vector is fused and reasoned through a causal fusion algorithm to generate a structured description of the abnormal event; combined with the built-in physical knowledge graph and historical equipment data, the abnormal event is traced back to its causal origin and the impact chain is deduced, and an analysis report containing the root cause, risk level, handling suggestions and causal chain is output, where the causal chain represents the nodes of the abnormal event and the causal transmission relationship between them; Digital Twin Collaborative Interaction Module: Constructs and maintains a 3D digital twin model that is updated synchronously with the power engineering site; generates a causal potential energy field in the model and visualizes anomalies based on the causal chain in the assessment report; performs virtual pre-operation simulation on the proposed handling measures and verifies the safety margin; generates an interactive early warning task containing real-world guidance and pushes the task to the real-world terminals of relevant personnel; receives operation response data, handling confirmation, and on-site verification information from the real-world terminals. Closed-loop verification and data traceability module: Based on the causal chain and risk level in the assessment report, a risk transmission diagram is established. Combined with the on-site verification information, the final risk decay value and risk entropy difference caused by the interactive early warning task are calculated. The risk decay value and risk entropy difference are fused through the risk entropy reduction verification model to generate an effectiveness index and encapsulate it into a safety status verification data packet, which is then stored in a distributed storage network.

[0019] Specifically, firstly, an array integrating audio, vibration, thermal imaging, and electromagnetic sensors is deployed at key risk nodes in the power engineering construction site to collect multimodal signals in real time. The system dynamically evaluates the signal fidelity of each sensor channel and fuses it with the calculated anomaly score to form a weighted primary anomaly signal vector. Subsequently, the causal cognition and diagnosis module, based on this vector and the fidelity of each channel, uses causal fusion algorithms such as fidelity weighted transfer entropy to infer and construct a dynamic causal graph to identify the root cause and propagation path of the anomaly. Combining physical knowledge graphs and historical data, Bayesian inference and Monte Carlo simulation are used to identify fault modes and extrapolate risk evolution, ultimately generating a structured assessment report containing the root cause, risk level, handling recommendations, and causal chain. Next, the digital twin collaborative interaction module maps the causal chain in the report to a three-dimensional digital twin model synchronized with the physical site in real time, visualizing the anomaly through a causal potential energy field. Virtual pre-operation simulations are performed on the handling recommendations to verify the safety margin, and then an interactive early warning task containing augmented reality guidance is generated, pushed to relevant personnel terminals, and their operation responses, handling confirmations, and on-site verification information are received. Finally, the closed-loop verification and data traceability module establishes a risk transmission diagram based on the causal chain in the assessment report, calculates the risk attenuation value and risk entropy difference brought about by the disposal operation by combining the on-site verification information, and integrates the two through the risk entropy reduction verification model to generate an effectiveness index, which is then packaged into a safety status verification data package and stored in a distributed storage network to form a complete trusted security control closed loop.

[0020] In one embodiment of the present invention, the multimodal edge sensing module includes: Sensor array deployment and acquisition unit: The sensor array is deployed at key risk nodes in the power engineering construction site. Each sensing unit in the sensor array integrates an audio sensor, a vibration sensor, a thermal imaging sensor and an electromagnetic sensor, forming multiple sensing channels for synchronous acquisition of audio, vibration, thermal imaging and electromagnetic environmental signals. Dynamic fidelity calculation unit: Calculates the dynamic fidelity of each sensing channel in real time based on the signal acquired by each sensing channel. The calculation formula is as follows: in, For signal quality factor, This is the sensor coupling state factor. As an environmental noise pollution factor, , , , These are the parameters of the pre-trained model; Anomaly Scoring and Vector Construction Unit: Based on the signals acquired by each sensor channel, calculate the anomaly score for each channel. ; Exception scoring Its corresponding dynamic fidelity Weighted fusion yields a weighted outlier score. Construct a primary anomalous signal vector, which contains at least the weighted anomaly scores of each sensing channel. and the dynamic fidelity of each sensor channel .

[0021] Specifically, at power engineering construction sites, key risk nodes are identified and determined. Key risk nodes include, but are not limited to, high-voltage equipment operating areas, high-altitude work areas, large lifting machinery working areas, areas with high temporary power consumption, and areas storing flammable and explosive materials. A sensing unit is deployed at each key risk node. This sensing unit is an integrated sensing device, internally containing four types of sensors: an audio sensor, a vibration sensor, a thermal imaging sensor, and an electromagnetic sensor. These sensors constitute four independent sensing channels. The audio sensor is used to collect sound wave signals from the construction site; the vibration sensor is used to collect mechanical vibration signals from equipment or structures; the thermal imaging sensor is used to collect infrared thermal radiation image sequences; and the electromagnetic sensor is used to collect electromagnetic field strength signals in specific frequency bands. All sensors are connected to an edge computing node via a unified hardware interface (such as RS-485 or Ethernet), and this node controls the synchronous acquisition of signals from the four channels. The sampling frequency is preset according to each signal type.

[0022] Edge computing nodes perform real-time processing on the raw signals acquired by each sensing channel, including signal denoising, filtering, normalization, and feature extraction, and calculate the dynamic fidelity of that channel at the current time t. Dynamic fidelity is an indicator of the reliability of a channel's signal; its value ranges from 0 to 1, with values ​​closer to 1 indicating higher reliability. The calculation formula is: ;in, The signal quality factor is calculated by normalizing the current signal's signal-to-noise ratio (SNR), sharpness (for audio), or image contrast (for thermal imaging) to the [0,10] range; for example, when the audio signal's SNR is greater than 40dB... The value is close to 9.5. This is the sensor coupling state factor, reflecting the degree of performance degradation caused by loose installation, aging, or electromagnetic interference from other sensors. It is obtained through self-testing and cross-validation of the sensor output, and its value ranges from [0, 10], where 0 indicates complete failure and 10 indicates optimal performance. For example, a newly installed and functioning sensor... The initial value is approximately 9.0. This study identifies environmental noise pollution factors and quantifies the impact of environmental factors (such as extreme temperature, humidity, strong winds, and strong electromagnetic interference) on signals. The impact is calculated by monitoring the deviation of environmental parameters from the signal baseline, with values ​​ranging from [0,10], where 0 represents no interference and 10 represents extremely strong interference. For example, in a windless, normal temperature environment... It is approximately 0.5, but rises to over 7.0 under strong electromagnetic interference. , , , These are the parameters of the pre-trained model, and their typical value range is... ∈[−2,2], ∈[0.1,0.3], ∈[0.05,0.15], ∈[0.2,0.4]; these parameters are obtained by training with logistic regression using historical normal and abnormal data (e.g., normal construction audio, normal equipment vibration baseline, temperature distribution images, electromagnetic field strength records, and corresponding data under abnormal conditions), with the aim of making... It can accurately reflect the reliability of the signal.

[0023] For each sensing channel, an anomaly score is calculated based on the real-time signal it acquires. This reflects the degree to which the current signal deviates from the normal reference; the specific calculation method varies depending on the sensor type. For the audio channel, the 13-dimensional Mel-frequency cepstral coefficient (MFCC) feature sequence of the signal is extracted within a time window (e.g., 1 second). MFCC reference sequence corresponding to the pre-stored normal construction environment sound model A comparison is performed; the minimum cumulative distance between the two is calculated using the Dynamic Time Warping (DTW) algorithm. The specific steps are as follows: constructing the distance matrix. ,in (Euclidean distance), where i ranges from 1 to... The value of j ranges from 1 to... Initialize the cumulative distance matrix cost(1,1)=D(1,1), and recursively calculate the cumulative distance: DTW distance is Divide the DTW distance by the preset maximum allowable distance. (For example =200 (obtained based on historical normal and abnormal samples) and amplitude limiting is applied to obtain a normalized anomaly score: For example, if the DTW distance between the current MFCC sequence and the reference sequence... =85.3, and =200.0, then =0.4265. For the vibration channel, calculate the root mean square (RMS) or peak factor of the signal in a specific frequency band (e.g., 1Hz-1000Hz), and compare it with the vibration threshold for normal equipment operation (this threshold is based on the equipment's factory standard or historical normal operation data statistics, for example, the RMS threshold is set to 0.5g). The excess percentage is taken as... ,Right now For the thermal imaging channel, abnormally high-temperature areas in the image are identified, and their area, maximum temperature, and temperature rise rate are calculated. Combined with a preset temperature rise alarm threshold model (e.g., an alarm is triggered if the area is >100cm² and the temperature is >80℃ or the temperature rise rate is >5℃ / min), a comprehensive overheating anomaly score is generated by weighted summation and normalization to the [0,1] interval. For electromagnetic channels, the field strength in a specific frequency band (such as the power frequency 50Hz and its harmonics) is monitored and compared with historical baselines (such as the rolling average of the past 24 hours) and safety limits (such as 100V / m). The degree of field strength exceeding the limit is calculated as... ,Right now Anomaly scores for all channels All values ​​were normalized to the [0,1] interval, where 0 represents complete normality and 1 represents severe abnormality.

[0024] Anomaly scoring based on each sensing channel and its corresponding dynamic fidelity The weighted anomaly score of the sensing channel is obtained. : Subsequently, a primary anomaly signal vector V(t) is constructed. This vector is a data structure with fixed dimensions, suitable for deployments with... A system with 1 sensor node (4 channels per node) has a dimension of 5. The primary anomaly signal vector contains, in sequence, the weighted anomaly scores for each sensing channel. and the dynamic fidelity of each sensor channel The specific arrangement is as follows: This vector summarizes the abnormal states and reliability information of all sensing channels, providing structured input for subsequent causal analysis and risk assessment.

[0025] In one embodiment of the present invention, the causal cognition and diagnosis module, based on the primary abnormal signal vector and the dynamic fidelity of each sensing channel, fuses and infers the primary abnormal signal vector using a causal fusion algorithm to generate a structured abnormal event description, including: Causal fusion unit: based on the weighted anomaly score of each sensor channel This forms a weighted anomaly score sequence. The fidelity-weighted transfer entropy between any two sensing channels X and Y is calculated using the following formula: Where, and are the dynamic fidelity coefficients for channels X and Y, respectively. The transit entropy is calculated based on the weighted anomaly score sequence. For time delay parameters; Cause-effect graph construction unit: Determines the optimal time delay parameters and causal direction by maximizing fidelity weighted transfer entropy, and constructs a dynamic cause-effect graph by combining preset physical causal rules and device topology connection relationships; Event description generation unit: Based on the dynamic cause-effect graph, it identifies the root cause node as the origin of the anomaly and the causal propagation path from the root cause node to other nodes; and generates a structured anomaly event description based on the root cause node and the causal propagation path.

[0026] Specifically, it receives a primary anomalous signal vector output from the multimodal edge sensing module, which contains a weighted anomalous score for each sensing channel. For a containing The system has multiple sensing channels. The weighted anomaly score of each channel is extracted and arranged in chronological order to form a time series of length L. This is called the weighted anomaly score sequence; here, L is the length of the sliding time window, typically ranging from 60 to 300 sampling points, corresponding to 1 to 5 minutes of physical time (assuming a sampling interval of 1 second). Before constructing the sequence, the weighted anomaly score of each sensor channel is detrended, i.e., the moving average of that sensor channel over a recent period (e.g., the previous 10 minutes) is subtracted to eliminate the influence of slowly changing background noise and highlight sudden anomalies.

[0027] Fidelity-weighted transitive entropy is a metric that measures the improvement in the predictive power of one time series over another, used to infer causal direction. It defines any two distinct sensor channels X and Y, and calculates the traditional transitive entropy based on the weighted anomaly score sequences of sensor channels X and Y. The calculation process is as follows: embedding the time series into... A 3D state space is typically taken as =2 or 3, meaning that the weighted anomaly score at the current moment and its previous one or two historical states are considered; given a time delay parameter Calculate the conditional probability (using the number of sampling intervals): ,in, This indicates that the weighted anomaly score of channel Y, starting from time t, has a total of [number missing] points. The set of values ​​at consecutive moments (usually taken as...) =1 or 2), This indicates that the weighted anomaly score of channel X, starting from time t, has a total of [number missing] points. The set of values ​​at consecutive moments (usually taken as...) =1 or 2), the probability distribution p is obtained by histogram statistics or kernel density estimation of the time series states within the sliding window, and the time delay parameter. Within a preset range (typical range is) ,in =1, The search is performed within a time window of 10 seconds (i.e., from 1 to 10 seconds). The average dynamic fidelity of sensor channels X and Y within the current analysis time window is obtained from the primary anomaly signal vector, denoted as... and Its fidelity-weighted transitive entropy The calculation formula is: The physical meaning of this formula is that it applies only when both channels have high fidelity (i.e., ... When the value is close to 1, the calculated transfer entropy is fully trusted; if the fidelity of any channel signal is low, its transfer entropy value will be suppressed accordingly, thereby reducing the interference of noise or malfunctioning sensors on causal inference.

[0028] To determine the causal strength from sensing channel X to Y, it is necessary to find the time delay parameter that maximizes the weighted transfer entropy. ,Right now: Simultaneously, calculate the opposite direction. Fidelity-weighted transfer entropy And similarly, find its maximum time delay. Compare the maximum weighted transfer entropy values ​​in the two directions; if If the significance exceeds a preset significance threshold δ (typically 0.05 to 0.2 bits), a significant causal influence from X to Y is considered to exist; otherwise, reverse causality or no significant causality exists. This process traverses all sensor channel pairs and initially constructs a weighted directed graph (i.e., a preliminary causal graph) based on the transfer entropy analysis results. The nodes of the graph are sensor channels, the directions of the edges are causal directions, and the weights are the corresponding maximum weighted transfer entropy values.

[0029] The system incorporates a physical knowledge graph, which defines the topological connections between devices (e.g., the output of transformer A is connected to switchgear B) and basic physical causal rules (e.g., abnormal vibration may be caused by electrical overload, which further leads to a temperature rise). The system then fuses the preliminary causal graph with the physical knowledge graph. For edges present in the preliminary causal graph, the system checks whether there is a direct physical connection or conforms to a known causal pattern between the physical entities (devices or locations) monitored by the connected sensing channels in the physical knowledge graph. If such a connection exists, the edge is retained, and its weight is multiplied by the confidence level of the physical rule (a value between 0 and 1, e.g., the confidence level for a direct electrical connection can be set to 0.9) to obtain the fused causal strength. If an edge in the preliminary causal graph has no basis in the physical knowledge graph, the edge is marked as "to be verified," and its weight is reduced (e.g., multiplied by a decay factor less than 1, such as 0.3). Simultaneously, if certain strong causal relationships existing in the physical knowledge graph are not detected in the initial causal graph (possibly due to signal delay or noise), the system will automatically add an edge in the corresponding direction based on the causal rules defined in the physical knowledge graph (e.g., a short circuit in a switchgear cabinet will inevitably lead to a protection trip), and assign a basic weight (e.g., 0.7) based on the rule confidence level. This ultimately generates a dynamic causal graph. ,in It is a set of nodes (each node corresponds to a sensing channel). It is a set of directed edges (representing the direction of causal influence). It is a set of edge weights (representing the strength of causal influence, with values ​​between 0 and 1); the specific weight of each edge is determined by the fusion of two factors: first, the "causal influence strength" calculated based on data-driven computation, which is the maximum value of the fidelity-weighted transitive entropy; second, the "causal rule strength" defined based on domain knowledge, which is the confidence level of the rule corresponding to the causal relationship path obtained from the physical knowledge graph (e.g., the confidence level of a direct electrical connection is 0.9); the fusion method is as described above (e.g., multiplication or weighted average). This dynamic causal graph is updated periodically (e.g., every 30 seconds) as new data windows are input.

[0030] In dynamic cause-effect diagrams In the process, the source of the abnormal propagation, namely the root cause node, is identified; among which, in the graph There are no significant incoming edges (i.e., the weights of all edges pointing to this node are below a threshold). ,typical =0.1), but has significant outgoing edges (at least one outgoing edge has a weight greater than the threshold). ,typical Nodes with a weight of 0.2 or less are identified as root cause nodes. This means that the anomalous changes of these nodes cannot be well explained by the changes of other nodes in the graph, but their changes can explain the changes of other nodes. Starting from each identified root cause node, a depth-first search or breadth-first search is performed, traversing all reachable downstream nodes along directed edges with edge weights greater than the path threshold (e.g., 0.15). During the traversal, the sequence of nodes visited is recorded, forming one or more causal propagation paths. Each path represents the hypothesis chain that the anomaly starts from the root cause node, passes through a series of intermediate events, and ultimately leads to other nodes observing the anomaly.

[0031] Based on the identified root cause nodes and causal propagation paths, a structured description of the anomalous event (e.g., in JSON format) is generated. This description includes at least: a unique identifier for the sensor channel corresponding to the root cause node; the physical entity monitored by that sensor channel (e.g., the A-phase winding of the #3 main transformer, mapped from the physical knowledge graph); the anomalous type corresponding to that sensor channel (e.g., overheating, excessive vibration, arcing, etc.); and a list containing all significant causal propagation paths (each path is itself a dictionary containing a sequence of channel IDs and the geometric mean of the weights of all edges on the path (as the overall confidence level)). The event trigger timestamp; where the anomaly type is obtained by matching the sensor type of the sensing channel and the key features used in the anomaly scoring process (such as the RMS value of the vibration channel, the highest temperature of the thermal imaging channel, etc.) with a pre-set anomaly pattern library. The anomaly pattern library is a mapping table that defines the anomaly type labels corresponding to the threshold range of various key features under different sensor types (for example, when the RMS value in the vibration channel is >0.4g, it is mapped as "vibration exceeds the standard", and when the highest temperature in the thermal imaging channel is >80℃ and the temperature rise rate is >5℃ / min, it is mapped as "overheating", etc.

[0032] In one embodiment of the present invention, the causal cognition and diagnosis module combines a built-in physical knowledge graph and historical equipment data to perform causal tracing and impact chain deduction of abnormal events, and outputs an assessment report including the root cause, risk level, handling suggestions, and causal chain. The causal chain characterizes the nodes of the abnormal event and the causal transmission relationship between them, including: Fault mode identification unit: Maps the root cause nodes in the structured abnormal event description to the corresponding physical entities in the built-in physical knowledge graph, and obtains the set of fault modes corresponding to the entity through multi-hop reasoning; Posterior probability calculation unit: for each fault mode The posterior probability is calculated based on Bayes' theorem. Where U is the set of observed multimodal anomaly evidence, and H is the historical data of the equipment; Impact Chain Inference Unit: Based on the failure mode with the highest posterior probability, it uses physical rules in the physical knowledge graph to infer the impact chain. The impact chain represents a series of consequence events caused by an anomaly starting from the root cause node and passing through the corresponding correlation. During the inference process, the uncertainty of the current load and environmental parameters is considered, and multiple risk evolution trajectories are generated through Monte Carlo simulation. Each risk evolution trajectory is a random realization of an impact chain. Risk entropy calculation and classification unit: Calculate risk entropy based on multiple risk evolution trajectories. The risk level is dynamically classified, and the calculation formula is as follows: ,in, Let k be the probability of the k-th influence chain occurring. The level of the most severe consequence caused by the k-th influence chain. For the k-th influence chain to evolve from the current state to... The estimated time for the consequences of the level; The assessment report generation unit generates an assessment report containing the root cause, risk level, handling recommendations, and causal chain based on risk entropy and its risk level classification, the failure mode with the highest posterior probability, the selected key impact chain as the inferred causal chain, and the handling rules in the physical knowledge graph; the causal chain consists of abnormal event nodes and the causal transmission relationship between them.

[0033] Specifically, its built-in physical knowledge graph is a graph database containing entities, attributes (such as rated current, insulation class, installation location, etc.), and relationships. Entities represent specific equipment (such as transformers, circuit breakers), components (such as windings, insulators), or abstract concepts (such as overload conditions) in power engineering. Each entity is associated with one or more predefined fault modes. A fault mode is a description of a specific failure mechanism or abnormal state, such as inter-turn short circuit in transformer windings, increased contact resistance due to oxidation of circuit breaker contacts, flashover on the surface of insulators, etc. The physical knowledge graph connects entities with fault modes through relationships such as "cause" or "manifests as". When a structured abnormal event description containing the root cause node and its corresponding physical entity is received, the fault mode identification unit performs multi-hop reasoning. First, it locates the physical entity node in the physical knowledge graph, and then traverses the fault mode nodes directly connected to it as a first-level fault mode set. Further, based on the inheritance relationship (such as belonging to a certain type of equipment) or co-occurrence relationship (such as often occurring together) defined in the physical knowledge graph, it queries the fault modes associated with other entities of the same type or related to this entity, and merges them to form the final fault mode candidate set. ,in This represents the total number of candidate failure modes. Its specific value depends on the number of modes associated with the entity in the graph. Typically, a physical entity may be associated with 3 to 10 possible failure modes.

[0034] For each fault mode in the fault mode candidate set , Calculate its posterior probability under the currently observed multimodal anomaly evidence U. The calculation is based on Bayes' theorem: Wherein, U represents the set of observed multimodal anomaly evidence, directly derived from the structured anomaly event description and the weighted anomaly score sequence of all channels in the primary anomaly signal vector. U is quantified into a feature vector, including numerical evidence such as vibration energy in a specific frequency band, maximum temperature value, temperature rise rate, and specific harmonic distortion rate. H represents the equipment's historical data, including historical curves of operating parameters, historical fault maintenance records, and mean time between failures (MTBF) of similar equipment over a past period (e.g., 90 days). Let H be the prior probability, representing the failure mode based on historical data H. The probability of occurrence; for example, the prior probability of "insulation aging" failure modes is higher for older equipment than for newer equipment; the prior probability can be obtained by statistically analyzing historical failure frequencies or using equipment reliability prediction models (such as the Weibull distribution). The initial value is set based on equipment factory data, industry standards, or historical statistical information, and the typical range is within... Between 0.1 and 0.1. Let be the likelihood probability, representing the probability when the fault mode is... When it actually occurs, the probability of observing the current evidence set U; its calculation is specifically as follows: for each failure mode A parameterized probability distribution model (e.g., multivariate Gaussian distribution) is established in advance. This model is trained based on a historical failure case library and simulation data, in which... and These are the mean vector and covariance matrix of the distribution model, respectively. By assigning... The likelihood value is obtained by performing maximum likelihood estimation on historical samples; during inference, the feature vector U of the current observed evidence is substituted into the probability density function of the distribution to calculate the likelihood value: This value is the likelihood probability. For non-Gaussian features, mixture models or kernel density estimation can be used for estimation. The historical fault case library is derived from the historical operation and maintenance records (alarm logs, inspection reports, maintenance reports) of the same or similar equipment, and is constructed by extracting fault modes and corresponding multi-sensor feature data through structured processing. Simulation data is generated by setting fault parameters in simulation software to supplement the missing severe fault samples in the historical cases, forming a training set together with the historical data. Posterior probability. The calculation result is a value between 0 and 1, and the sum of the posterior probabilities of all candidate failure modes is 1; the failure mode with the highest posterior probability is determined to be the most likely root cause.

[0035] Select the failure mode with the highest posterior probability As the starting point for influence chain deduction, this deduction is conducted within a physical knowledge graph. This graph not only contains static entities and fault relationships but also defines dynamic physical propagation rules. For example, a rule might be stated as: "If a transformer experiences winding overheating (Event A) and the cooling system fails (Condition C), then the insulating oil temperature will exceed the threshold (Event B)." Within the physical knowledge graph, [the following will be implemented / implemented]. The mapping is to its corresponding initial abnormal state node. For example, the fault mode "transformer winding inter-turn short circuit" is mapped to the state node "transformer A-phase winding short circuit"; the deduction unit starts from... Starting from the corresponding initial abnormal state node, the propagation process of the abnormal state is simulated according to the propagation rules of physics, electricity, thermodynamics, etc. defined in the physics knowledge graph.

[0036] To quantify uncertainty, Monte Carlo simulation is used in the simulation process. Specifically, the simulation process involves randomly sampling current load parameters (such as current and power), environmental parameters (such as ambient temperature and wind speed), and uncertainty parameters involved in the rules (such as fault development rate and equipment tolerance threshold) according to their probability distribution (usually assumed to be normally or uniformly distributed) to form a parameter set. For example, the ambient temperature is sampled normally around the current measured value with a standard deviation of ±2°C. Based on the sampled parameter set, physical propagation rules are deterministically applied to generate a chain that starts from the root cause event and triggers a series of subsequent consequence events (such as protection alarms and equipment tripping). The simulation is terminated when the simulation state reaches a stable point (such as the protection action clearing the fault), exceeds a preset time range (such as extrapolating the next hour), or all possible subsequent effects have occurred. This simulation process is repeated. Second-rate, Typical values ​​are 1000 to 10000 times, thus obtaining Each risk evolution trajectory is a random realization of the possible future development path of an abnormal event.

[0037] Monte Carlo simulation Statistical analysis was performed on the risk evolution trajectory to calculate the risk entropy. A clustering algorithm based on sequence similarity (such as dynamic time warping distance) was used to cluster all trajectories, and the following clustering patterns were summarized: There are 1 representative key influence chain, and each key influence chain k (k=1,2,..., This corresponds to a typical accident evolution scenario; calculate the probability of each key influence chain k occurring. Its value is the proportion of the number of simulation trajectories that generated this scenario mode to the total number of simulations. The proportion. Determine the level of the most severe consequence that each chain k could potentially cause. The consequence level is a predefined discrete value, for example, divided into 5 levels: Level 1 (minor anomaly), Level 2 (alarm), Level 3 (equipment damage), Level 4 (partial power outage), and Level 5 (large-scale power outage or personal injury). This level is determined based on the severity of the event at the end of the chain, referring to a pre-defined consequence level mapping table. For example, the mapping table could specify that if the end event is "critical equipment temperature exceeds 150℃," then the corresponding... =3 (Equipment failure). Estimate the progression of each chain k from its current state to its most severe consequence level. The estimated time required for the corresponding state (Unit: minutes) Take all the simulation trajectories belonging to this chain to achieve The average time required to reach a certain level. Risk entropy. The calculation formula is: Risk entropy It is a positive real number that comprehensively measures the uncertainty, severity, and urgency of risks in a power engineering system; the larger the value, the higher the overall risk faced by the system. According to... The risk level is dynamically divided based on the numerical range, for example: ∈[0,2) indicates low risk (blue warning). ∈[2,5) is considered medium risk (yellow warning). The range [5,10) is considered high-risk (orange alert). ∈[10,+∞) represents extremely high risk (red alert); the threshold range can be calibrated based on historical data from different engineering scenarios.

[0038] The final result is an analysis report that includes the root cause, risk level, remedial recommendations, and causal chain; specifically, the root cause is the failure mode with the highest posterior probability. And its corresponding physical entity. The risk level is specifically determined based on the calculated risk entropy. The assigned interval is labeled with the current risk level (e.g., "high risk") and a specific entropy value is provided. The causal chain specifically involves filtering out the probability of occurrence from the key influence chains generated by the Monte Carlo simulation. The event level is higher than the threshold (this threshold is set to, for example, 0.05, a typical value obtained based on historical data analysis, used to filter low-probability events) and the consequence level The top 3 to 5 chains are output as typical causal chains derived from the deduction, with each chain presented as a text description and a sequence of event nodes. The specific recommendations are based on the identified root cause. Based on the risk level, the system queries the pre-built handling rule base in the physical knowledge graph. The handling rule base contains standardized operation guidelines and emergency plans for different fault modes and risk levels. For example, for the fault mode "transformer winding overheating" and the risk level "high risk", the handling suggestions may include: immediately reduce the transformer load to 50% of the rated capacity, start all backup cooling devices, and notify maintenance personnel to arrive within 30 minutes. These suggestions are combined and instantiated to generate an actionable list of step-by-step handling suggestions.

[0039] In one embodiment of the present invention, the digital twin collaborative interaction module constructs and maintains a three-dimensional digital twin model that is synchronously updated with the power engineering site, including: Twin Model Construction Unit: Constructs a three-dimensional digital twin model corresponding to the power engineering site. The model includes a geometric model layer representing the geometric structure of the site and a physical model layer representing the physical laws of the equipment. The physical model layer is a parameterized physical simulation model embedded in the key equipment. The parameters of the parameterized physical simulation model are obtained by Bayesian parameter inversion based on the historical operating data of the key equipment in the power engineering site and continuously calibrated. Real-time synchronization unit: This unit achieves real-time synchronization between the 3D digital twin model and the physical site of the power engineering project through a state prediction-correction synchronization algorithm. The algorithm includes: Prediction Step: Based on the parametric physical simulation model and current commands from the power engineering field control system and status Input, predict the state at the next time step. ,in, This refers to the process noise term; Calibration step: When the measured data of the received sensor signal is obtained. At that time, the confidence weights of the prediction results of the parametric physical simulation model are calculated. Confidence weights of measured data : ,in To predict the error covariance, To measure the noise variance; Hybrid update step: The predicted state and the measured data are merged according to the calculated weights to update the current state. .

[0040] Specifically, a three-dimensional digital twin model is constructed that perfectly corresponds to the target power engineering site. Structurally, this model consists of two complementary layers. The first layer is the geometric model layer. This layer primarily reconstructs the three-dimensional geometric shape of the entire construction area and all equipment and facilities based on the site's design blueprints, Building Information Modeling (BIM) data, and high-precision laser scanning point clouds using computer graphics technology. This model accurately reflects the spatial dimensions, shape, and relative positional relationships of objects, typically requiring a geometric accuracy of less than 2 centimeters from the physical entity to support precise spatial positioning and visualization. The second layer is the physical model layer, the core of the digital twin's simulation and prediction capabilities. It embeds corresponding parametric physical simulation models for key equipment identified in the geometric model (such as main transformers, high-voltage circuit breakers, diesel generators, etc.). These models are programmed components described by mathematical equations that can simulate the dynamic physical behavior of the equipment. Taking an oil-immersed air-cooled transformer as an example, its embedded physical simulation model typically includes a set of differential equations describing its thermal dynamics. The state variables of this model at least include the top oil temperature. Hotspot temperature and winding temperature The control input includes at least the load current. and ambient temperature The behavior of the model is determined by a set of parameters. These parameters, as determined by the parameters, are the parameters of the parametric physical simulation model.

[0041] For the example of the oil-immersed air-cooled transformer thermal model, the parameters Specifically, this includes the thermal resistance from the winding to the oil. (Typical value range 0.02-0.15 K / W), thermal resistance from oil to environment (Typical value range: 0.1-0.6 K / W, depending on cooling method), winding heat capacity (typical value) to J / K) and the heat capacity of oil (typical value) to J / K), etc.; parameters of the physical simulation model The specific implementation method of Bayesian parameter inversion is as follows: collect multi-source time series data of the device during a period of stable and normal operation in history. For example, load current, ambient temperature, and oil temperature measured directly or indirectly, recorded every five minutes for two weeks; assuming parameters There exists a prior probability distribution P(θ) based on the design values, for example, assuming that each parameter follows a Gaussian distribution with a standard deviation of 20% of the initial value around its typical value; given parameters Under conditions of historical input sequences (such as load current and ambient temperature), the physical simulation model is run to generate predicted temperature sequences for the corresponding time points; the predicted temperature sequences are then compared with the actual observed sequences. Construct the likelihood function Applying Bayes' theorem, the posterior probability distribution of the parameters is obtained. The Markov Chain Monte Carlo (MCMC) sampling algorithm is used to approximate the posterior distribution. Samples of the posterior distribution of the parameters are obtained through MCMC sampling, and their mean is taken as the optimal estimate. (Usually the mean of the posterior distribution is taken) and its uncertainty interval.

[0042] To achieve real-time synchronization between the state of the 3D digital twin model and the actual state of the physical site (synchronization period) Typically set to 1 second), a confidence-weighted state prediction-correction synchronization algorithm is employed, which is executed cyclically at each discrete time step t. The first step is the prediction step: at time t, the digital twin model holds the best estimates of all current state variables (such as the temperature, voltage, vibration amplitude, etc. of each device), denoted as the state vector. Simultaneously, the current control commands and operating condition inputs are obtained from the Supervisory Control and Data Acquisition (SCADA) system, denoted as the input vector. (e.g., circuit breaker status, generator output setpoint); utilizing embedded physical simulation models and inverted calibrated parameters. Predict the state at the next time step t+1; the prediction formula is as follows: The state transition function f is obtained by discretizing the differential equation or difference equation of the physical simulation model. The process noise vector is used to characterize the stochastic uncertainty introduced by the physical simulation model due to simplifying assumptions, parameter errors, and unmodeled dynamics. It is modeled as a vector with zero mean and a covariance matrix of... Gaussian white noise process, i.e. Covariance matrix The magnitude and correlation of this uncertainty were quantified, and they were obtained primarily through two methods: one is based on historical data, estimating it by analyzing the statistical characteristics of the model's prediction error sequence under normal operating conditions; the other is based on prior engineering knowledge. For example, statistical analysis of historical prediction errors can be used to set the parameters. Specifically, it can be set as a diagonal matrix, with its diagonal elements taking a certain proportion (such as 5%) of the variance of the corresponding state variable in the historical normal fluctuation range.

[0043] The second step is the correction step: at time t+1, the latest measured data from the field sensor network is actually received. When (e.g., the temperature value of an infrared thermometer or the effective value of acceleration from a vibration sensor) is measured, the calibration process begins. The core of this step is calculating the relative weights of the model's predicted values ​​and the measured values ​​in the state update, i.e., the confidence weights; the confidence weights of the model predictions... Confidence weights of measured data The calculation formula is: .in, It is the prediction error covariance matrix, which incorporates the uncertainty of the model prediction. Its calculation depends on the state estimation error covariance of the previous time step. By adjusting the state transition function f in the current state The Jacobian matrix obtained by taking the partial derivative at point 1 and process noise covariance At the initial time of the system (t=0), given the initial state estimate and its corresponding initial state estimation error covariance matrix (Usually set as a diagonal matrix, the values ​​of its diagonal elements are determined by the estimation of the initial uncertainties of each state variable.) By performing the first prediction step, the initial prediction error covariance matrix is ​​calculated. The calculation formula is as follows: ,in, Let f be the Jacobian matrix of the initial state transition function f. The initial process noise covariance matrix; thereafter, According to the formula at every moment Iterative updates will be performed. It is a matrix The trace, that is, the sum of all its diagonal elements, quantitatively characterizes the magnitude of the uncertainty of the overall predicted state. This is the current measurement noise variance, based on the dynamic fidelity of this sensor channel provided by the multimodal edge sensing module. Dynamically adjusted; specific relationships can be as follows: ,in This is the variance of the sensor's fundamental measurement error under ideal conditions; therefore, when the sensor signal quality is high ( When it is close to 1), Small, measured data weight Increase; conversely, when the model's prediction uncertainty is high ( When the weights are large, the model predicts the weights. reduce. and All are dimensionless numbers between 0 and 1.

[0044] The third step is the hybrid update step: based on the calculated confidence weights, the predicted state and the measured data are fused to obtain the final state estimate of the 3D digital twin model at time t+1. : After the state is updated, the state estimation error covariance matrix also needs to be updated synchronously. To prepare for the next prediction, the formula is updated as follows: ,in This is the noise covariance matrix of the current measurement, with its diagonal elements representing the measurement noise variance of each sensor channel. .

[0045] By continuously running the state prediction-correction synchronization algorithm, the 3D digital twin model can integrate physical simulation predictions with real-time sensing data, maintain high-fidelity synchronization with the power engineering construction site, and provide a dynamic simulation environment for subsequent visualization, simulation and task generation.

[0046] In one embodiment of the present invention, the digital twin collaborative interaction module generates a causal potential energy field and visualizes anomalies in the model based on the causal chain in the assessment report; performs virtual pre-operation simulation and verifies the safety margin of the handling suggestions, generates an interactive early warning task containing real-world guidance, and pushes the task to the real-world terminals of relevant personnel; and receives operation response data, handling confirmation, and on-site verification information from the real-world terminals, including: Causal Chain Visualization Unit: Maps the causal chain in the assessment report to a causal potential energy field in the three-dimensional space of the three-dimensional digital twin model. For each anomalous event node m in the causal chain, its spatial location is defined. Causal potential energy generated at the location : in, The amplitude is determined by the risk level of this abnormal event. To determine the location of this anomalous event in the 3D digital twin model, For the scope of influence, For dynamic evolution frequency, For phase; The causal potential energy field is the vector superposition of the potential energies of all anomalous event nodes. ,in It is a causal direction unit vector, and is visualized in a three-dimensional digital twin model using volume rendering techniques in computer graphics; Virtual pre-operation simulation unit: In a 3D digital twin model, virtual pre-operation simulations are performed on the handling recommendations in the assessment report, and safety margin indicators are calculated. To assess operational safety: ,in, For the nth safety indicator, For safety limits, The current value, This is a normal value; if If the value is below the preset threshold, the handling plan will be adjusted and the simulation will be repeated until the safety margin index meets the requirements. Interactive task generation and push unit: Based on the verified handling plan, generate an interactive early warning task containing step-by-step augmented reality guidance and push it to the real terminals of relevant personnel; Feedback Receiving and Verification Unit: After the interactive early warning task is pushed, it receives operation response data, handling confirmation, and on-site verification information from the actual terminal; based on the on-site verification information, it calculates the verification confidence score. : in, The first on-site verification Item index value, These are the predicted values ​​from the twin model. To allow for error, For weights.

[0047] Specifically, the analysis and judgment report is analyzed to extract the causal chain information contained therein. A causal chain consists of multiple abnormal event nodes and the directed causal transmission relationship between the nodes. Each abnormal event node is associated with a three-dimensional entity with a clear spatial location in the three-dimensional digital twin model. In order to intuitively demonstrate the spatiotemporal correlation and propagation trend of abnormal events, the causal chain is mapped as a dynamic causal potential energy field with precise geometric coordinates and scale overlaid on the three-dimensional space defined by the digital twin model for visualization.

[0048] For each anomalous event node m in the causal chain (including a total of...), (an abnormal event node), its location in three-dimensional space. Causal potential energy generated at the location The calculation formula is: .in, The spatial coordinates (in meters) of the abnormal event node m in the 3D digital twin model are obtained by adding the local offset of the abnormal event occurrence determined by the geometric model of the equipment to the center coordinates of the equipment model; for example, for transformer winding overheating, Located at the geometric center of the winding inside the transformer casing. The potential energy amplitude (dimensionless) is determined by the risk level of this anomalous event. The mapping relationship between risk level and amplitude is predefined; for example, low risk (blue) corresponds to... =0.5, medium risk (yellow) corresponds to =1.0, High Risk (Orange) =1.5, extremely high risk (red) corresponds to =2.0; The value range is generally from 0.1 to 5.0 to ensure that different risk levels are clearly distinguished in the visualization. The influence range parameter (unit: meters) of the anomalous event determines the causal potential energy. The rate of decay in space characterizes the physical radius of influence of the anomalous event; its value is related to the type of anomaly and the size of the equipment, and a typical calculation method is as follows. ,in These are the characteristic dimensions of the equipment that is malfunctioning (such as the average length, width, and height of a transformer). It is a proportionality coefficient between 0.5 and 3.0, determined by experience or historical case statistics; for example, the impact range of a localized hotspot is relatively small ( =0.5), while a severe short-circuit current surge can affect the entire bay ( =3.0). The dynamic evolution frequency (unit: radians / second) controls the dynamic fluctuation rate of the potential energy field, used to characterize the urgency or volatility of an abnormal state. Its value is determined by the development rate of the anomaly, and the calculation formula is as follows: ,in This is the estimated time of change of abnormal state characteristics (in seconds); for example, for a rapidly developing arc fault, 0.1 seconds ≈62.8rad / s), while for slow insulation aging, For 3600 seconds ( ≈0.0017 rad / s). Phase (unit: radians) is used to distinguish the starting point of fluctuations in different anomalous events at the same time. It is usually randomly initialized to a value in the interval [0, 2π), or calculated based on the trigger timestamp of the anomalous event, for example... . The current time (in seconds).

[0049] Total causal potential field It is the vector superposition of the potential energy of all abnormal event nodes: ;in It is the causal direction unit vector, determined by the direction pointing to the upstream node of node m in the causal chain; if node m is the root cause node (without an upstream node), then... Take it as the zero vector; for non-root cause nodes, It is set as the direction vector pointing from its main upstream node to the location of node m, and normalized to a unit length. The calculated total 3D space... The potential energy value is converted into a 3D texture data; a volume rendering algorithm based on ray casting is used to assign specific colors and transparency to potential energy values ​​of different sizes and directions; typically, high potential energy areas are rendered as a striking red with low transparency, while low potential energy areas are rendered as cool tones with high transparency; directional information is represented by drawing streamlines or arrow icons embedded in the volume data; this visualization effect will change over time. It updates dynamically in response to changes in abnormal states.

[0050] In the 3D digital twin model, a virtual pre-operation simulation is conducted on the preliminary handling recommendations given in the assessment report to verify the safety and effectiveness of the operation and optimize the operation steps. Before the simulation begins, the control inputs or parameters θ of the relevant models are modified in the physical model layer of the digital twin model according to the definition of the handling recommendations; for example, if the handling recommendation is "disconnect the 35kV bus tie switch", then the disconnection state of the corresponding switch model is used as the input in the simulation. Subsequently, a real-time simulation is started based on the state vector (including variables such as temperature and voltage of each device) of the physical simulation model and the 3D digital twin model at the start of the simulation; the simulation runs at a faster speed than real time (e.g., 10 times faster) to predict the state evolution over a future period of time (e.g., the next 30 minutes); during the simulation, a set of predefined safety indicators are continuously monitored. These indicators correspond to key physical quantities, such as temperature, voltage, current, vibration amplitude, and pressure at critical nodes. After the simulation is completed, a comprehensive safety margin index is calculated for the operational sequence corresponding to each proposed action. This is used to quantify the safety buffer space of the operational plan: ;in, This is the safety limit for the nth safety indicator, i.e., the threshold that cannot be exceeded. This value is derived from equipment technical specifications, industry standards, or safety regulations. For example, for the top oil temperature of a transformer, the safety limit is... ; This is the predicted value of the nth safety indicator obtained by running the physical simulation model and performing a deduction under a given disposal operation input after the virtual pre-operation simulation is completed (or under the worst-case scenario during the simulation process); This is a typical value for this safety indicator under completely normal operating conditions, used as a benchmark. For example, the top oil temperature during normal transformer operation might be... .ratio This represents the ratio of the current state's distance from the safe limit to the normal state's distance from the safe limit; the closer this value is to 1, the more sufficient the safety margin; 0 indicates the limit has been reached; less than 0 indicates the limit has been exceeded. The final safety margin indicator. The minimum value of this ratio among all monitored safety indicators is taken, following the "barrel principle," where the weakest link determines the overall safety margin. A preset safety margin threshold is then established. Typical values ​​are 0.2 or 0.3. If the calculated... If the risk level falls below this threshold, the proposed treatment plan is considered high and requires adjustment. The process of adjusting the treatment plan follows a rule-based iterative optimization logic, identifying factors that could lead to this risk. The lowest critical safety indicator (i.e., the weakest link) is used to select and execute targeted adjustment operations based on the type of the critical indicator, the associated equipment, and a pre-set optimization rule base. The optimization rule base defines typical optimization actions and their applicable conditions for different types of safety indicators (such as temperature, voltage, vibration, etc.). For example, if the critical safety indicator is transformer winding temperature, the corresponding optimization rules might include "reducing the load current to X% of the rated value" or "activating the backup cooling device"; if it's a bus voltage exceeding the limit, it might include "adjusting the output of the reactive power compensation device" or "switching the transformer tap changer." Based on the rules, a new adjusted handling plan is generated, and a virtual pre-operation simulation is performed again for calculation. This process is repeated until... The solution is considered valid only after reaching a preset maximum number of iterations (e.g., 5 times), at which point an interactive warning task can be generated.

[0051] Based on the validated response plan, an interactive early warning task is generated for on-site personnel. This task is a structured data package containing at least the following core elements: task overview, augmented reality (AR) guidance, and task metadata. The task overview concisely describes the abnormal event, its root cause, risk level, and core response objectives. The AR guidance binds each step of the response plan to the spatial location of the 3D digital twin model and the equipment model, generating a series of step-by-step AR instructions. These instructions include: the visual identifier of the equipment to be operated in the real-world environment (located via image recognition or QR code), the operational actions (e.g., "rotate this valve to the closed position"), the operating standards (e.g., "continue until a click is heard"), and safety precautions (e.g., "wear insulated gloves"). The task metadata includes a unique interactive early warning task ID, creation time, estimated completion time, and information on the responsible personnel or work group. The generated interactive early warning task data is proactively pushed to the real-world terminal devices of relevant personnel, such as tablets, AR smart glasses, or smartphones, via a secure wireless network (such as a 5G private network or Wi-Fi 6). The push uses a priority queue, with higher risk levels receiving higher push priority, and can be accompanied by multimodal alerts such as sound and vibration.

[0052] After an interactive early warning task is pushed out, a feedback receiving window is opened to receive information from on-site personnel terminals in real time, including operation response data, handling confirmation, and on-site verification information. Operation response data specifically includes questions and supplementary observation information in text or voice format. Handling confirmation specifically indicates the completion status of each AR guidance step (e.g., not started, in progress, completed). On-site verification information specifically refers to the actual values ​​of key verification indicators collected and uploaded by on-site personnel after the handling is completed, using sensors on their terminal devices (such as cameras, infrared thermal imagers, multimeter interfaces) or manual input, such as temperature and voltage readings after the fault point is repaired and returning to normal.

[0053] Based on the uploaded on-site verification information, a verification confidence score is calculated. To quantify the degree of consistency between the predictions of the digital twin model and the actual results on site: ;in, The total number of metrics used for validation. ∈[1, Typically, three to five of the most critical, measurable indicators are selected. The first one obtained from on-site verification The actual measured value of the indicator; To calculate the predicted value of the same index at the same time based on the physical simulation model and deduction results in the three-dimensional digital twin model after the virtual pre-operation simulation is completed. For the first The permissible error range for each indicator is determined based on the accuracy of the measuring instrument and the engineering tolerances; for example, for temperature measurement, if an accuracy of ± Infrared thermal imagers, Set as . For the first The weight coefficients of the indicators satisfy the following conditions: Weights are allocated based on the importance of the indicators; for example, indicators that directly reflect whether a fault has been eliminated have a higher weight. Typical values ​​could be 0.4 or 0.3, etc. Confidence score verification. It is a value between 0 and 1; A higher score indicates a better match between the virtual simulation and the actual situation on-site, and more reliable its diagnosis and prediction; this score will be recorded. If the data falls below a preset confidence threshold (e.g., 0.7, which can be set based on the consistency of historical validation data or an acceptable false alarm rate), an alarm will be triggered, indicating that there is model mismatch or a new unknown failure mode, requiring manual intervention for analysis.

[0054] In one embodiment of the present invention, the closed-loop verification and data traceability module establishes a risk transmission diagram based on the causal chain and risk level in the assessment report, and calculates the final risk attenuation value caused by the interactive early warning task in conjunction with the on-site verification information, including: Risk Graph Modeling Unit: Models the causal chains in the assessment report into a risk transmission graph. =(V,E,W), where node V represents an anomalous event node in the causal chain, edge E represents the causal transmission relationship, and edge weight W represents the risk transmission intensity. Risk attenuation calculation unit: node After completing the suggested measures in the interactive early warning task, calculate the effect of these measures on the overall risk mitigation, i.e., the risk attenuation value. for: in, Represents a node The set of all downstream nodes; For nodes The inherent risk intensity is determined by the risk level in the assessment report. The probability of occurrence of the influence chain to which this node belongs. Decision, that is , From arrive The set of all paths; Let e ​​be the blocking efficiency of edge e; Spatiotemporal calibration unit: for risk attenuation values Spatiotemporal calibration is performed to obtain the final risk attenuation value. : ,in, The time decay factor, This represents the environmental difficulty factor.

[0055] Specifically, the causal chain described in the analysis report is transformed into a structured mathematical model, namely a directed acyclic graph, called a risk transmission graph. =(V,E,W); the set of nodes in the graph. For all anomalous event nodes identified in the causal chain, each node ( ∈[1, The event corresponds to a specific abnormal event, such as the temperature of the A-phase winding of a transformer exceeding the standard. The edge set E represents the causal transmission relationship between nodes of the abnormal event. The occurrence of this directly led to the abnormal event. If the occurrence of [something] (or a significant increase in its probability and severity) occurs, then there exists a directed edge [something]. from point to Since the causal chains occur sequentially in time and do not form cycles, the graph is acyclic. The set of edge weights W corresponds one-to-one with each edge. There is a weight value ∈[0,1] indicates from node To the node The weight is determined based on the results of the causal relationship discovery phase, and its value combines the "causal influence strength" of data-driven inference with the "causal rule strength" defined in the physical knowledge graph. Typically, strong causal relationships (such as fault propagation caused by direct electrical connections) correspond to a weight... >0.7, weight corresponding to weak causal relationship (such as indirect influence through ambient temperature) <0.3.

[0056] Suppose that on-site personnel, based on the suggestions from the interactive early warning task, target a specific abnormal event node in the risk transmission diagram. (For example, an initial overheating fault point) is addressed with measures (such as emergency cooling); the goal of the risk attenuation calculation unit is to quantify the “blocking” effect of the measures on the risk transmission map.

[0057] Abnormal event nodes After the measures take effect, they are considered to be able to completely or partially block the flow from the anomalous event node. The risk of propagating to all its downstream nodes. Risk decay value. Defined as the cumulative period-weighted risk caused by all downstream nodes if this measure is not taken, its calculation formula is as follows: ;in, In the risk transmission diagram In, from the node Starting from the set of all downstream nodes reachable through directed edges, this is obtained using a graph depth-first search or breadth-first search algorithm. For nodes The inherent risk intensity represents the magnitude of the original risk that would result if the anomalous event were to occur. It is determined by two factors: the risk level assessed for this event in the analysis report. and its probability of occurrence The calculation formula is: Risk level This is a numerical value categorized by the severity of the consequences, typically using discrete integer levels. For example, the severity of consequences could be mapped to levels 1 to 5, corresponding to Level 1 (minor anomaly), Level 2 (alarm), Level 3 (equipment damage), Level 4 (partial power outage), and Level 5 (large-scale power outage or personal injury); probability of occurrence. Without considering upstream nodes When the event has been blocked, the prior probability of the node event occurring independently under the current operating conditions is derived from the equipment's historical fault records, the operation statistics of similar equipment, or the failure probability predicted based on the equipment's physical simulation model, and the value ranges from [0,1]. Refers to the source node to the target downstream node The set of all possible paths may have only one path in a simple chain causality, but may have multiple paths in a branching causality. Let be the blocking efficiency of edge e, representing the proportion by which its mitigation measures reduce the risk transmitted through that edge. The value range is [0,1]; if the remedial measures can completely eliminate the causal relationship (e.g., severing the physical connection), then =1; if it can only partially alleviate the problem, then 0 < <1; if invalid, then =0; In the initial stage of system operation, if there is no relevant historical data, It can be set to a conservative default value (such as 0.5), and its value will quickly converge to the actual level as data accumulates.

[0058] Directly calculated risk decay value This is a theoretical value. To more reasonably reflect the actual contribution of the response measures in the real spatiotemporal environment, it needs to be calibrated to obtain the final risk attenuation value. : .in, The time decay factor takes into account the timeliness of risk management; risks are time-dependent, and the earlier they are managed, the greater the potential losses that can be mitigated. It is a function of the response time t (the time interval from issuing the warning to personnel confirmation of the completion of the response, in minutes); a typical definition is a piecewise function or an exponentially decaying function, for example: ,in It is the decay time constant, for example, set to 30 minutes. This means that the contribution value of a process that takes longer to complete will be discounted over time. This is an environmental difficulty factor used to calibrate the objective difficulty differences in performing the same disposal operation under different environmental conditions, so that operations performed in high-difficulty environments receive higher weight. It is a coefficient greater than or equal to 1, and multiple environmental dimensions are considered during the calculation: ;in, , , It is an indicator function that takes the value of 1 when the corresponding adverse conditions (such as thunderstorms, nighttime, or high-altitude operations) are met when the incident occurs, and 0 otherwise. , , These are the difficulty coefficients for each dimension, set by management rules, with typical values ​​between 0.1 and 0.3; for example, setting... =0.2, =0.15, =0.25, then the difficulty factor for high-altitude operations during thunderstorms is =1+0.2+0.15+0.25=1.6.

[0059] In one embodiment of the present invention, the closed-loop verification and data traceability module calculates the risk entropy difference caused by the interactive early warning task; it then fuses the final risk decay value and the risk entropy difference through a risk entropy reduction verification model to generate an effectiveness index, encapsulates it into a security status verification data packet, and stores it in a distributed storage network, including: Risk entropy calculation unit: calculates the risk entropy before the disposal measures are completed. , and the risk entropy after completion Where S is the set of risk states, and These represent the probability of risk state s occurring before and after the completion of the disposal measures; Risk Entropy Difference Calculation Unit: Calculates the risk entropy difference. ; Confidence verification calculation unit: based on the risk entropy difference and the final risk attenuation value The effectiveness index of the response measures is calculated using a risk entropy reduction verification model. : ,in, , To contribute weighting coefficients, As the benchmark risk level; Data storage generation unit: Generates a security status verification data package containing an effectiveness index, final risk decay value, risk entropy difference, associated interactive early warning task identifier, personnel identifier, and timestamp; Network Record Unit: The security status verification data packet is recorded in the distributed storage network using a practical Byzantine fault-tolerant consensus algorithm to form a security contribution record.

[0060] Specifically, risk entropy is the application of the concept of information entropy in the field of risk assessment, used to measure the uncertainty of the risk state of the entire system. Risk entropy needs to be calculated before and after the implementation of mitigation measures to quantify the degree to which the mitigation actions eliminate risk uncertainty. A set S of risk states of a system is defined, containing all possible discretized final consequence states that can occur during the influence chain deduction in the causal cognition and diagnosis module; for example, a typical set could be... These can represent five mutually exclusive and complete risk outcomes: normal status, minor abnormality, protection alarm, partial equipment damage, and partial power outage.

[0061] In the instant before the response measures are completed, based on the current assessment report and dynamic causal diagram, Monte Carlo simulation is used to extrapolate the risk evolution over a future period (e.g., the next hour); the final state of all risk extrapolation trajectories is statistically analyzed to determine which risk state it falls into. The frequency of each risk state is normalized to obtain a probability estimate of its occurrence, denoted as . For example, through 1000 simulations, we can obtain... ("Partial equipment damage") = 0.15 ("Causing a partial power outage") = 0.05, and the probabilities of other states can be deduced similarly, and the following conditions are met: Based on this probability distribution, the risk entropy before treatment is calculated. : Risk entropy It is a non-negative real number. The larger its value, the higher the uncertainty of the future risk outcome of the system, that is, the more unpredictable the risk state; typically, The value range may be from 0 (deterministically in a certain state) to... (Any state is possible). After the response measures are completed and their effectiveness is verified on-site, the state of the 3D digital twin model will be updated, and a rapid risk simulation will be performed again based on the new state (the number of simulations can be reduced, for example, 500 times), thereby obtaining the probability distribution of the new risk state after the response. Subsequently, the risk entropy after treatment was calculated using the exact same formula. : .

[0062] Risk entropy difference This represents the reduction in overall system risk uncertainty due to the successful implementation of mitigation measures, and its calculation formula is as follows: Since effective safety measures aim to converge the system to a safer, more deterministic state, there are typically... < ,thereby >0; The larger the value, the more significant the contribution of the action to reducing systemic risk uncertainty; for example, a single action might reduce the risk entropy from 1.2 bits to 0.3 bits. =0.9 bits.

[0063] Effectiveness Index It is a single numerical value that ultimately quantifies the total contribution of the safety response measures, directly quantifying the risk attenuation value. Difference between risk entropy and information theory dimension Combination; effectiveness index The formula, calculated using the risk entropy reduction verification model, is as follows: ;in, This is the final risk attenuation value after spatiotemporal calibration; This represents the risk entropy difference. and The contribution weighting coefficients are used to balance the final risk decay value (direct effect) and the risk entropy difference (uncertainty reduction effect) in the overall effectiveness index. The proportions in; these two coefficients are set by the administrator according to the security management policy, and are both positive real numbers, usually satisfying + =1 or adjusted according to actual management rules; typical values ​​may be =0.6, =0.4 indicates a greater emphasis on the direct risk elimination effect. The baseline risk level is a constant used for normalization, whose value is set as the average risk entropy or risk intensity value assessed by the system during typical normal operation or under a certain baseline state; for example, the average risk entropy calculated from all "low-risk" states of the system over the past month can be used as the baseline risk level. ;typical The value is between 0.5 and 2.0.

[0064] The data storage and generation unit receives the validity index, the final risk decay value, and the risk entropy difference. Simultaneously, it retrieves from the system context the globally unique identifier (ID) of the interactive early warning task associated with this closed-loop response, the identifier of the personnel or team performing the response, and the response completion timestamp. All the above data fields, along with the encrypted hash digest of the causal chain data in the original assessment report that triggered the process (used for subsequent tamper-proof tracing), are assembled into a standardized data structure. This structure is defined and encoded using an efficient binary serialization format (such as Protocol Buffers) to ensure compact data and fast parsing speed. The serialized data constitutes the original content of the security status verification data packet. To ensure the integrity and source credibility of the data packet, the unit calculates the encrypted hash value (such as SHA-256) of the original content as the data digest. Subsequently, using an asymmetric cryptographic private key belonging to this regulatory system or project (securely stored in the hardware security module), a digital signature operation is performed on the hash value to generate a signature result. Finally, the serialized original data, its digital digest, and the digital signature (if used) are packaged together into the final security status verification data packet. Subsequently, the network recording unit submits the security status verification data packet as evidence of the transaction to a permissioned distributed storage network jointly maintained by project participants (such as the owner, contractor, and supervisor); the network employs the Practical Byzantine Fault Tolerance (PBFT) consensus algorithm. The core process is as follows: the network recording unit submits the security status verification data packet to a consensus node; the node verifies the format and broadcasts it; the consensus node group achieves consistency in the order and state of a batch of transactions through the PBFT protocol (including proposal, voting, and submission phases), generating a new block; once a block is confirmed by more than two-thirds of the nodes, it is finalized; subsequently, the block data is persistently stored in a distributed storage engine (such as a storage system based on IPFS or erasure coding). Thus, the data packet is recorded immutably, forming a globally verifiable and traceable security contribution record.

[0065] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. An artificial intelligence-based power engineering safety construction supervision system, characterized in that, Includes the following modules: Multimodal edge sensing module: Deploy sensor arrays at key risk nodes in power engineering construction sites to simultaneously collect audio signals, vibration signals, thermal imaging signals and electromagnetic environment signals, dynamically evaluate the fidelity of each sensing channel, and output primary anomaly signal vectors. Causal cognition and diagnosis module: Based on the primary abnormal signal vector and the dynamic fidelity of each sensor channel, the primary abnormal signal vector is fused and reasoned through a causal fusion algorithm to generate a structured description of the abnormal event; combined with the built-in physical knowledge graph and historical equipment data, the abnormal event is traced back to its causal origin and the impact chain is deduced, and an analysis report containing the root cause, risk level, handling suggestions and causal chain is output, where the causal chain represents the nodes of the abnormal event and the causal transmission relationship between them; Digital Twin Collaborative Interaction Module: Constructs and maintains a 3D digital twin model that is updated synchronously with the power engineering site; generates a causal potential energy field in the model and visualizes anomalies based on the causal chain in the assessment report; performs virtual pre-operation simulation on the proposed handling measures and verifies the safety margin; generates an interactive early warning task containing real-world guidance and pushes the task to the real-world terminals of relevant personnel; receives operation response data, handling confirmation, and on-site verification information from the real-world terminals. Closed-loop verification and data traceability module: Based on the causal chain and risk level in the assessment report, a risk transmission diagram is established, and combined with the on-site verification information, the final risk decay value and risk entropy difference caused by the interactive early warning task are calculated. By fusing the risk decay value and the risk entropy difference through the risk entropy reduction verification model, an effectiveness index is generated and encapsulated into a security status verification data packet, which is then stored in a distributed storage network.

2. The artificial intelligence-based power engineering safety construction supervision system according to claim 1, characterized in that, The multimodal edge sensing module includes: Sensor array deployment and acquisition unit: The sensor array is deployed at key risk nodes in the power engineering construction site. Each sensing unit in the sensor array integrates an audio sensor, a vibration sensor, a thermal imaging sensor and an electromagnetic sensor, forming multiple sensing channels for synchronous acquisition of audio, vibration, thermal imaging and electromagnetic environmental signals. Dynamic fidelity calculation unit: Calculates the dynamic fidelity of each sensing channel in real time based on the signal acquired by each sensing channel. The calculation formula is as follows: in, For signal quality factor, This is the sensor coupling state factor. As an environmental noise pollution factor, , , , These are the parameters of the pre-trained model; Anomaly Scoring and Vector Construction Unit: Based on the signals acquired by each sensor channel, calculate the anomaly score for each channel. ; Exception scoring Its corresponding dynamic fidelity Weighted fusion yields a weighted outlier score. Construct a primary anomalous signal vector, which contains at least the weighted anomaly scores of each sensing channel. and the dynamic fidelity of each sensor channel .

3. The artificial intelligence-based power engineering safety construction supervision system according to claim 1, characterized in that, In the causal cognition and diagnosis module, based on the primary abnormal signal vector and the dynamic fidelity of each sensing channel, the primary abnormal signal vector is fused and inferred using a causal fusion algorithm to generate a structured description of the abnormal event, including: Causal fusion unit: based on the weighted anomaly score of each sensor channel This forms a weighted anomaly score sequence, and the fidelity-weighted transfer entropy between any two sensing channels X and Y is calculated. The calculation formula is as follows: ,in, and These are the dynamic fidelity coefficients for channels X and Y, respectively. The transit entropy is calculated based on the weighted anomaly score sequence. For time delay parameters; Cause-effect graph construction unit: Determines the optimal time delay parameters and causal direction by maximizing fidelity weighted transfer entropy, and constructs a dynamic cause-effect graph by combining preset physical causal rules and device topology connection relationships; Event description generation unit: Based on the dynamic cause-effect graph, it identifies the root cause node as the origin of the anomaly and the causal propagation path from the root cause node to other nodes; and generates a structured anomaly event description based on the root cause node and the causal propagation path.

4. The artificial intelligence-based power engineering safety construction supervision system according to claim 3, characterized in that, The causal cognition and diagnosis module combines a built-in physical knowledge graph and historical equipment data to perform causal tracing and impact chain deduction of abnormal events, outputting an assessment report including the root cause, risk level, handling recommendations, and causal chain. The causal chain characterizes the nodes of the abnormal event and the causal transmission relationships between them, including: Fault mode identification unit: Maps the root cause nodes in the structured abnormal event description to the corresponding physical entities in the built-in physical knowledge graph, and obtains the set of fault modes corresponding to the entity through multi-hop reasoning; Posterior probability calculation unit: for each fault mode The posterior probability is calculated based on Bayes' theorem. Where U is the set of observed multimodal anomaly evidence, and H is the historical data of the equipment; Impact Chain Inference Unit: Based on the failure mode with the highest posterior probability, it uses physical rules in the physical knowledge graph to infer the impact chain. The impact chain represents a series of consequence events caused by an anomaly starting from the root cause node and passing through the corresponding correlation. During the inference process, the uncertainty of the current load and environmental parameters is considered, and multiple risk evolution trajectories are generated through Monte Carlo simulation. Each risk evolution trajectory is a random realization of an impact chain. Risk entropy calculation and classification unit: Calculate risk entropy based on multiple risk evolution trajectories. The risk level is dynamically classified, and the calculation formula is as follows: ,in, Let k be the probability of the k-th influence chain occurring. The level of the most severe consequence caused by the k-th influence chain. For the k-th influence chain to evolve from the current state to... The estimated time for the consequences of the level; The assessment report generation unit generates an assessment report containing the root cause, risk level, handling recommendations, and causal chain based on risk entropy and its risk level classification, the failure mode with the highest posterior probability, the selected key impact chain as the inferred causal chain, and the handling rules in the physical knowledge graph; the causal chain consists of abnormal event nodes and the causal transmission relationship between them.

5. The artificial intelligence-based power engineering safety construction supervision system according to claim 1, characterized in that, The digital twin collaborative interaction module constructs and maintains a three-dimensional digital twin model that is updated synchronously with the power engineering site, including: Twin Model Construction Unit: Constructs a three-dimensional digital twin model corresponding to the power engineering site. The model includes a geometric model layer representing the geometric structure of the site and a physical model layer representing the physical laws of the equipment. The physical model layer is a parameterized physical simulation model embedded in the key equipment. The parameters of the parameterized physical simulation model are obtained by Bayesian parameter inversion based on the historical operating data of the key equipment in the power engineering site and continuously calibrated. Real-time synchronization unit: This unit achieves real-time synchronization between the 3D digital twin model and the physical site of the power engineering project through a state prediction-correction synchronization algorithm. The algorithm includes: Prediction Step: Based on the parametric physical simulation model and current commands from the power engineering field control system and status Input, predict the state at the next time step. ,in, This refers to the process noise term; Calibration step: When the measured data of the received sensor signal is obtained. At that time, the confidence weights of the prediction results of the parametric physical simulation model are calculated. Confidence weights of measured data : ,in To predict the error covariance, To measure the noise variance; Hybrid update step: The predicted state and the measured data are merged according to the calculated weights to update the current state. .

6. The artificial intelligence-based power engineering safety construction supervision system according to claim 5, characterized in that, In the digital twin collaborative interaction module, based on the causal chain in the judgment report, a causal potential energy field is generated in the model and anomalies are visualized; the handling suggestions are simulated virtually and the safety margin is verified, an interactive early warning task containing real-world guidance is generated, and the task is pushed to the real-world terminals of relevant personnel. Receive operation response data, handling confirmation, and on-site verification information from the actual terminal, including: Causal Chain Visualization Unit: Maps the causal chain in the assessment report to a causal potential energy field in the three-dimensional space of the three-dimensional digital twin model. For each anomalous event node m in the causal chain, its spatial location is defined. Causal potential energy generated at the location : in, The amplitude is determined by the risk level of this abnormal event. To determine the location of this anomalous event in the 3D digital twin model, For the scope of influence, For dynamic evolution frequency, For phase; The causal potential energy field is the vector superposition of the potential energies of all anomalous event nodes. ,in It is a causal direction unit vector, and is visualized in a three-dimensional digital twin model using volume rendering techniques in computer graphics; Virtual pre-operation simulation unit: In a 3D digital twin model, virtual pre-operation simulations are performed on the handling recommendations in the assessment report, and safety margin indicators are calculated. To assess operational safety: ,in, For the nth safety indicator, For safety limits, The current value, This is a normal value; if If the value is below the preset threshold, the handling plan will be adjusted and the simulation will be repeated until the safety margin index meets the requirements. Interactive task generation and push unit: Based on the verified handling plan, generate an interactive early warning task containing step-by-step augmented reality guidance and push it to the real terminals of relevant personnel; Feedback Receiving and Verification Unit: After the interactive early warning task is pushed, it receives operation response data, handling confirmation, and on-site verification information from the actual terminal; based on the on-site verification information, it calculates the verification confidence score. : in, The first on-site verification Item index value, These are the predicted values ​​from the twin model. Let be the allowable error, and be the weight.

7. The artificial intelligence-based power engineering safety construction supervision system according to claim 1, characterized in that, In the closed-loop verification and data traceability module, a risk transmission diagram is established based on the causal chain and risk level in the assessment report. Combined with the on-site verification information, the final risk attenuation value caused by the interactive early warning task is calculated, including: Risk Graph Modeling Unit: Models the causal chains in the assessment report into a risk transmission graph. =(V,E,W), where node V represents an anomalous event node in the causal chain, edge E represents the causal transmission relationship, and edge weight W represents the risk transmission intensity. Risk attenuation calculation unit: node After completing the suggested measures in the interactive early warning task, calculate the effect of these measures on the overall risk mitigation, i.e., the risk attenuation value. for: in, Represents a node The set of all downstream nodes; For nodes The inherent risk intensity is determined by the risk level in the assessment report. The probability of occurrence of the influence chain to which this node belongs. Decision, that is , From arrive The set of all paths; Let e ​​be the blocking efficiency of edge e; Spatiotemporal calibration unit: for risk attenuation values Spatiotemporal calibration is performed to obtain the final risk attenuation value. : ,in, The time decay factor, This represents the environmental difficulty factor.

8. The artificial intelligence-based power engineering safety construction supervision system according to claim 7, characterized in that, In the closed-loop verification and data traceability module, the risk entropy difference caused by the interactive early warning task is calculated. By fusing the final risk decay value and the risk entropy difference through a risk entropy reduction verification model, an effectiveness index is generated and encapsulated into a security status verification data packet, which is then stored in a distributed storage network, including: Risk entropy calculation unit: calculates the risk entropy before the disposal measures are completed. , and the risk entropy after completion Where S is the set of risk states, and These represent the probability of risk state s occurring before and after the completion of the disposal measures; Risk Entropy Difference Calculation Unit: Calculates the risk entropy difference. ; Confidence verification calculation unit: based on the risk entropy difference and the final risk attenuation value The effectiveness index of the response measures is calculated using a risk entropy reduction verification model. : ,in, , To contribute weighting coefficients, Baseline risk level Data storage generation unit: Generates a security status verification data package containing an effectiveness index, final risk decay value, risk entropy difference, associated interactive early warning task identifier, personnel identifier, and timestamp; Network Record Unit: The security status verification data packet is recorded in the distributed storage network using a practical Byzantine fault-tolerant consensus algorithm to form a security contribution record.

Images