Intelligent risk identification method and system for energy increasing station
By deploying a sensor network within the refueling station, collecting and processing multi-source signals, mapping them to energy flow fingerprints, and constructing a causal network, the problem of early perception and closed-loop intervention for risk identification in refueling stations is solved, enabling accurate risk location and intelligent management.
Patent Information
- Application Number
- CN202511876638.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-12
- Publication Date
- 2026-03-17
AI Technical Summary
The operational scenarios and equipment of refueling stations are becoming increasingly complex. Existing technologies are unable to effectively integrate heterogeneous data streams, achieve early perception of risk characteristics and physically interpretable modeling, and lack a complete decision-making chain from anomaly identification to closed-loop intervention.
By deploying a sensor network within the power station, multi-source signals are collected and windowed to extract multi-scale time-frequency features and physical response features. These features are then mapped to energy flow fingerprints using a hybrid projection model. Combined with differential baselines, self-calibration is performed to identify candidate anomaly patterns. A local causal network with physical semantic constraints is constructed to quantify risk levels and execute a tiered intervention action chain.
It enables early warning of potential risks, improves the accuracy and interpretability of risk identification, ensures the necessity, safety and feasibility of intervention measures, forms a risk management and control system with self-evolution capabilities, and enhances the intelligent level of safety management at the refueling station.
Smart Images

Figure CN121684640A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial Internet of Things and edge computing technology, specifically to a method and system for intelligent risk identification of power stations. Background Technology
[0002] With the continuous advancement of digital and intelligent transformation in the energy industry, the operational safety and management efficiency of refueling stations, as important energy infrastructure, are receiving increasing attention. In recent years, the continuous development and integration of sensor technology, edge computing, and artificial intelligence methods have provided new technical paths for real-time acquisition of multi-physics field signals within the station, in-depth extraction of operational characteristics, and intelligent analysis.
[0003] CN120217127A discloses a real-time monitoring method and system for thermal power equipment based on edge computing. The method includes: deploying an array of edge computing nodes to collect multi-source heterogeneous data from thermal power equipment; preprocessing the data through data filtering, multi-scale adaptive filtering, and wavelet packet decomposition; extracting features using a conditional variational autoencoder; generating hybrid feature representations by combining a hierarchical attention mechanism and a deep feature fusion network; performing refined distribution estimation and abnormal pattern recognition of equipment status; and performing collaborative monitoring modeling based on a multi-scale spatiotemporal feature fusion network and a hierarchical deep deterministic strategy gradient network to generate an optimal monitoring scheme. This method can effectively improve the accuracy and efficiency of real-time monitoring of thermal power equipment, reduce equipment failure rate, and ensure the safe and stable operation of thermal power plants.
[0004] Currently, refueling stations are developing into comprehensive energy supply stations, and their operational scenarios and equipment are becoming increasingly complex, placing higher demands on the real-time, accurate, and systematic nature of risk management. The continuous advancement of technologies such as multi-sensor information fusion, edge intelligent computing, and causal inference has provided new technical paths for building a more intelligent and proactive safety protection system. Under this trend, how to effectively integrate heterogeneous data streams, achieve early perception and physically interpretable modeling of risk characteristics, and form a complete decision-making chain from anomaly identification to closed-loop intervention has become an important issue in the industry's technological development. Summary of the Invention
[0005] The purpose of this invention is to address the problems existing in the background technology by proposing an intelligent risk identification method and system for energy stations.
[0006] The technical solution of this invention: an intelligent risk identification method for energy refueling stations, comprising the following specific implementation steps: S1. By deploying a sensor network in the power station and synchronously acquiring signals, the multi-source signals are windowed to extract multi-scale time-frequency features and physical response features. The heterogeneous features are mapped to energy flow fingerprints through a hybrid projection model, and self-calibration enhancement is performed using differential baselines. S2. Decode the energy flow fingerprint and map it into semantic features. Identify candidate anomaly patterns through short-term trend analysis and multi-source cross-correlation, and calculate local anomaly scores. S3. Construct a local causal network with physical semantic constraints based on candidate anomaly patterns, deduce the anomaly propagation path and quantify the risk contribution of each node, classify the risk level and locate the key risk source. S4. Encapsulate risk information, including key risk sources, risk levels, and contribution levels, into structured risk tokens. Match corresponding handling strategies from a pre-set strategy library based on the risk tokens, and execute a transactional action chain consisting of several actions according to the principle of hierarchical intervention. Simultaneously, monitor the execution effect of the action chain in real time and support effect rollback and strategy upgrade.
[0007] Preferably, step S1, which involves deploying a sensor network and synchronously acquiring signals, specifically includes: Sensor arrays are deployed in the refueling area, tank area, charging area, equipment room, and personnel flow passage of the refueling station; Each sensor group includes at least a chemical gas sensor, an acoustic sensor, an infrared temperature measurement unit, a differential pressure sensor, a current sampling unit, and a behavior capture device; All sensors are connected to the edge node and timestamp synchronization is ensured using hardware clocks and network time protocols; Edge nodes employ an overlapping sliding window strategy to slice the signal, and perform high-pass filtering to remove low-frequency drift and low-order filtering or wavelet denoising to remove high-frequency noise, resampling signals with different sampling rates to a unified reference frequency.
[0008] Preferably, step S1, which extracts multi-scale time-frequency features and physical response features, specifically includes: Perform time-frequency transformation on the sensor signal in each window, and calculate its energy integral value in multiple preset frequency bands as the basic quantity; The transient peak value, envelope rate of increase, and spectral centroid of the signal are calculated as descriptors; By combining the sensor's own response time constant, nominal sensitivity value, local hot spot area, temperature gradient, or frequency band distribution physical parameters, a window-level original feature vector is formed for each sensor; The original feature vectors of all sensors are concatenated into the overall feature vector.
[0009] Preferably, step S1, mapping heterogeneous features to energy flow fingerprints using a hybrid projection model, specifically includes: The overall eigenvectors are projected onto the physical basis using a linear transformation matrix with physical priors, resulting in a linear projection result. A single-layer small neural network is used to perform nonlinear compensation on the linear projection results; The compensation results are normalized to obtain a normalized energy flow vector. The uncertainty vector of the normalized energy flow vector is obtained based on the sensor confidence level and characteristic noise estimation.
[0010] Preferably, the self-calibration enhancement using differential baseline in step S1 specifically includes: Maintain a short-term baseline and a long-term baseline. The short-term baseline is the moving average of a specific number of past windows, and the long-term baseline is the moving average of a longer historical window. Calculate the differences between the current normalized energy flow vector and the short-term baseline and the long-term baseline, respectively, to obtain the short-term difference vector and the long-term difference vector; The short-term and long-term difference vectors are weighted and combined to form the final difference response; The long-term baseline is updated using an exponential moving average method with a forgetting factor; The weighted norm of the final difference response is calculated as the local anomaly score; The normalized energy flow vector and the final differential response are quantized and hashed to generate a compact fingerprint packet.
[0011] Preferably, step S2, which decodes the energy flow fingerprint and maps it to semantic features, specifically includes: The received compact fingerprint packet is dequantized to recover an approximate energy flow vector; The approximate energy flow vector is mapped to the semantic space using a semantic projection matrix to form a semantic vector. Constructing short-term sequences based on semantic vectors from continuous time windows; Calculate the first and second differences of the semantic vectors in this short-term sequence; Calculate the mean, variance, and maximum gradient of this short-term series; The first-order difference, second-order difference, mean, variance, and maximum gradient value are combined to form a short-term anomaly feature vector.
[0012] Preferably, step S2, which identifies candidate anomaly patterns through short-term trend analysis and multi-source cross-correlation, specifically includes: Calculate the short-time correlation coefficient matrix for short-term semantic sequences from different sensors or regions; Introducing lag time to calculate short-time correlation coefficients with lag terms to capture delay effects; The short-term correlation coefficient, the correlation coefficient with lag term, and the short-term anomaly feature vector are combined to form a local pattern vector; Local anomaly scores are obtained by weighting the local pattern vectors. The local anomaly score is compared with a threshold adapted based on the exponential moving average, and if it exceeds the threshold, it is marked as a candidate anomaly pattern.
[0013] Preferably, the construction of a local causal network with physical semantic constraints in step S3 specifically includes: For candidate anomaly patterns, identify relevant equipment, sensors, or regional indicators within the power station as causal nodes; Establish a conditional probability matrix between nodes based on historical data to characterize causal relationships; Imposing physical semantic constraints on the conditional probability matrix ensures that causal relationships conform to actual physical or energy flow paths; Construct a propagation weight matrix based on the conditional probability matrix and local anomaly scores; The diffusion process of the anomaly along the propagation weight matrix is simulated by iterative calculation to obtain the comprehensive anomaly influence vector of each node; The risk contribution coefficient of each node is calculated based on the comprehensive anomaly impact vector.
[0014] Preferably, step S3, which involves classifying risk levels and identifying key risk sources, specifically includes: An adaptive risk threshold vector is set based on historical statistics and equipment differences. The threshold vector includes a low-risk threshold, a medium-risk threshold, and a high-risk threshold. The risk contribution coefficient of each node is compared with the threshold vector to classify them into low-risk, medium-risk, or high-risk levels, and a risk level matrix is generated. High-risk nodes are sorted in descending order according to their risk contribution coefficient; High-risk nodes whose risk contribution coefficient exceeds a preset contribution threshold are identified as key risk sources. The system generates intelligent alerts that include the cause of the anomaly, its potential impact, and recommended actions, based on the location, level, and spread of key risk sources.
[0015] The technical solution of the present invention: an intelligent risk identification system for refueling stations, which is used to execute the above-mentioned intelligent risk identification method for refueling stations, comprising: The multi-layer sensing and energy flow fingerprinting module is used to deploy and synchronously control the sensor network within the power station, collect multi-source signals, perform windowing preprocessing on the signals, extract multi-scale time-frequency features and physical response features, generate a normalized energy flow vector through a hybrid projection model, perform self-calibration using differential baseline, and generate a compact energy flow fingerprint flow. The candidate anomaly pattern recognition and local scoring module is used to decode the energy flow fingerprint and map it into a semantic vector, construct a short sequence and calculate a short anomaly feature vector, perform multi-source cross-correlation analysis to generate a local pattern vector, calculate the local anomaly score and identify candidate anomaly patterns. The multi-source causal inference and risk level determination module is used to construct a local causal network with physical semantic constraints based on candidate anomaly patterns, infer anomaly propagation paths and calculate node risk contribution coefficients, classify node risk levels based on adaptive thresholds, locate key risk sources and generate intelligent prompts. The risk strategy adaptation and intelligent intervention module is used to encapsulate key risk sources, risk levels and context information into risk tokens, match and sort disposal strategies from the strategy library, execute transactional action chains according to the hierarchical principle of prompting reminders, soft intervention, automatic isolation to emergency activation, evaluate intervention measures through sandbox simulation, monitor the execution effect in real time and support rollback and strategy upgrade.
[0016] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: This invention designs an intelligent risk identification method and system for refueling stations. It constructs a closed-loop risk identification and response system from bottom-level perception to top-level decision-making, significantly improving the intelligence level and proactive defense capabilities of refueling station safety management. By fusing multi-source heterogeneous sensor data and converting it into physically interpretable energy flow fingerprints, the system can keenly capture weak abnormal signals from complex background noise, achieving early warning of potential risks. The introduction of a causal inference mechanism based on physical semantic constraints allows anomaly analysis to go beyond superficial correlations, inferring propagation paths and root causes, greatly improving the accuracy and interpretability of risk location. The system employs a hierarchical adaptive strategy matching and sandbox simulation pre-playing mechanism to ensure the necessity, safety, and feasibility of intervention measures, maximizing business continuity while controlling risks. The entire solution achieves a balance between data processing and decision response efficiency through edge computing and cloud collaboration, and optimizes the strategy library and models through a continuous learning mechanism, forming a risk management system with self-evolution capabilities, providing reliable technical support for the long-term safe and stable operation of refueling stations. Attached Figure Description
[0017] Figure 1 This is a flowchart of a method for intelligent risk identification of energy stations proposed in this invention; Figure 2 This is a system architecture diagram of an intelligent risk identification system for energy stations proposed in this invention. Detailed Implementation
[0018] Example 1, as Figure 1 As shown, the specific implementation steps of the intelligent risk identification method for energy stations proposed in this invention are as follows: S1. By constructing a multi-layer sensing grid at the power station, windowing multi-source signals, extracting multi-scale time-frequency features, and combining lightweight physical projection and differential baseline self-calibration, the raw sensor data is transformed into a compact, physically interpretable energy flow fingerprint representation. The specific implementation process is as follows: S11. By deploying a sensor network in key areas of the power station and synchronously acquiring data with a unified time baseline, a parallel overlapping sliding window strategy of short and long windows is used to slice, denoise, and resample multi-source signals. This achieves data preprocessing that considers both transient and trend data, making signals from different sensors such as gas, temperature, acoustic, and electrical sensors comparable and fused within the same time window. This provides a unified and clean input for subsequent feature extraction. Specifically: Sensor groups are arranged in the refueling station according to functional areas (refueling area, storage tank area, charging area, equipment room, and personnel passage). Each group includes at least: a chemical gas sensor (or catalytic / semiconductor), a set of acoustic / ultrasonic sensors, an infrared temperature measurement / thermal imaging unit, a pair of differential pressure / flow sensors, a set of current / harmonic sampling units, and a set of low-resolution behavior cameras or attitude micro-motion capture devices. Each sensor group is connected to the nearest edge node (SoC / FPGA gateway), and the timestamp synchronization error is guaranteed to be less than half of the sampling window (e.g., ≤10ms, depending on the service settings) using a hardware clock or NTP / PTP. Edge nodes employ an overlapping sliding window strategy: a short window Ws (e.g., 0.5-2s) is used to capture transient events, and a long window W... l (e.g., 10-60s) is used to capture trends and diffusion processes; the two types of windows use different step sizes. and Slide; for each sensor i, form the original vector within window k. (Time truncated and timestamped); High-pass filtering (or first-order differential de-trending) is used for low-frequency drift; low-order filters or wavelet denoising are used for high-frequency noise; if the sensor sampling rates are different, sampling interpolation (linear or polynomial) is used to resample all signals to a unified reference frequency. in, This represents the original time-domain signal (including timestamp, truncated to the window) of the i-th sensor within the k-th window; and These represent the sliding step size (time interval between window movements) for the short window and the long window, respectively. In this embodiment, it is set to half the window length to ensure a balance between coverage and computation. S12. Within each window, extract features from the sensor signal, including multi-scale spectrum, time-frequency energy, transient peak, envelope change, and physical response indicators. Combine these features with the sensor's own response constant and operating parameters to form a rich descriptive vector, ensuring that the features both reflect the signal change pattern and retain physical meaning. Specifically: Acquiring continuous signals Frequency band energy distribution: ; Calculate the energy of each frequency band as a basic quantity: ; Calculate descriptors such as transient peak value, envelope rate of increase, and spectral centroid to form a vector set. ; It should be noted that, for gas sensors, the response-recovery time constant and nominal sensitivity (obtained from the sensor manufacturer / experimental calibration) are recorded; for thermal imaging, the local hotspot area and temperature gradient are extracted; for acoustic imaging, transient energy and frequency band distribution are extracted. The above time-frequency and physical response descriptions are concatenated into a window-level original feature vector: ; in, This represents the window function used for localized time-frequency transformation; in this embodiment, a Hamming window is used. The function value represents the time-frequency representation of the i-th sensor in window k, and the function value varies with frequency f and time shift τ. This represents the frequency demodulation term, which is the basic kernel function that converts the time-domain signal into its various frequency components. This indicates that the i-th sensor window k is in frequency band B. b The energy integral value; This represents a set of time-frequency and physical response characteristics extracted by the i-th sensor in window k; This indicates that all sensors The overall feature vector (with dimension M) is obtained by concatenating the data sequentially; N represents the total number of sensors. S13. By constructing a linear-nonlinear hybrid projection model with physical priors, heterogeneous feature vectors are mapped to a unified energy flow feature space. This allows the resulting energy flow vectors to not only express the core properties of different energy processes (thermal, acoustic, concentration diffusion, current perturbation), but also retain sensitivity to minute energy flow changes while compressing dimensions. This forms a core structure for energy flow fingerprints that is compact, interpretable, and highly discriminative. Specifically: A hybrid projection operator with physical priors is used to provide an energy flow vector for the window features: ; right Perform component normalization and generate an uncertainty vector (This embodiment is based on sensor confidence level w) i (Obtained from characteristic noise estimation): ; ; in, Represents a linear transformation matrix with shape . The original features are projected onto an m-dimensional physical basis; m represents the dimension of the projected energy flow vector (the dimension of the energy flow feature space). This represents a low-parameter nonlinear compensation function for the insufficient portion of the linear projection (in this embodiment, a small one-layer neural network is used), with the following parameters: ; This represents the m-dimensional energy flow feature vector obtained after projection, which represents the physical quantity signature of this window; Indicates to Normalize the vectors to eliminate dimensional differences; Indicates and The corresponding measurement / projection uncertainty estimates, with each component corresponding to a variance; The reliability weight of a single sensor or channel is determined by historical performance and calibration results; denoted as zero constant; p represents the norm type used in vector normalization or norm calculation; Indicates variance; S14. Using short-term and long-term baseline fingerprints as references, differential encoding and self-calibration are performed on each real-time energy flow vector to amplify extremely weak drift signals and suppress seasonal environmental disturbances. Simultaneously, a lightweight hash compression and update strategy at the edge level is employed to generate the final transmittable energy flow fingerprint stream. This ensures the fingerprint is highly sensitive without increasing network load, providing stable and continuously updated foundational fingerprint evidence for subsequent risk identification modules. Specifically: Preserve short-term baseline (This embodiment is set to past T) s (Window moving average) and long-term baseline (This embodiment is set to past T) l The sliding mean of the window. ), calculate the difference vector: ; ; The final difference response is formed by weighted combination: ; Update the long-term baseline using the exponentially moving average (EMA) with a forgetting factor. To adapt to slow environmental changes: ; Define the local anomaly score (weighted norm): ; ; To reduce bandwidth and protect the original video / high-frequency data, edge nodes... and Performing quantization and hashing: ; Based on this edge node output: Compact fingerprint packet ; when At the same time, additional decoding requests are reported (cloud / operations can request complete decoding). , (Decoding or retrieving local cache fragments for causal analysis) in, and These represent the differences between the current normalized vector and the short-term / long-term baselines, respectively, reflecting the relative deviations in the short and long term. This represents a weighted difference that combines short-term and long-term differences; Indicates the weighting coefficient; This represents the updated long-term baseline; This represents the EMA forgetting factor, used for updating the long-term baseline exponential moving mean. This represents a weighted diagonal matrix of the difference vector components, with elements of... ; This represents the weight of the j-th energy flow component at time k; This represents the local anomaly score, i.e., an anomaly metric based on a weighted difference vector; This represents the edge threshold, which is used to determine... Does it significantly exceed the threshold of normal fluctuations (triggering candidate anomalies)? Represents a fingerprint hash, a compact fingerprint identifier that is quantized and hashed, used for reporting, retrieval, and indexing; This represents a hash function, specifically a lightweight hash calculated on a quantized vector. This indicates a quantization operation.
[0019] S2. Based on the physically interpretable energy flow fingerprint output in step S1, the micro-drift information is transformed into actionable semantic features. Preliminary anomaly patterns are generated through short-term trend analysis and multi-source cross-correlation, providing high-confidence candidate events for subsequent risk assessment. The specific implementation process is as follows: S21. Decode the edge quantized fingerprint into an approximate energy flow vector and map it to the semantic space to form an interpretable semantic vector, taking into account both micro-drift sensitivity and physical meaning, providing a foundation for subsequent anomaly feature extraction, specifically: Edge hashing Fast dequantization with the quantization vector recovers the approximate energy flow vector. And use local cache or edge decoding table for mapping to ensure that no critical micro-drift information is lost; Define a semantic mapping function to project the decoded vector into the semantic space: ; in, This represents the mapped semantic vector, i.e., the event semantic features of the k-th time window; This represents the semantic projection matrix, a weight matrix that maps physical vectors to the semantic space; This represents the decoded approximate energy flow vector, which contains the original energy flow value and its micro-drift information; This represents the semantic mapping bias, used to correct for long-term drift or environmental context. This represents an uncertainty mapping function that adjusts semantic values based on confidence levels. S22. Construct a continuous time window sequence, calculate short-term differences and trend indicators, and form a sensitive short-term anomaly feature vector to highlight transient anomalies and potential risk signals. Specifically: continuous time window semantic vectors Forming short-term sequence SD k ; Calculate the short-term difference vector: ; ; Statistical trend indicators: Short-term series mean short-term series variance The maximum gradient value g of short-term sequences max etc., forming short-term abnormal feature vectors ; Among them, SD k Represents a short-term semantic sequence matrix, consisting of the most recent T s semantic vectors of each window Build; T s Indicates the length of the short-term time window; The first-order difference of the semantic vector reflects instantaneous changes; The second-order difference of the semantic vector reflects the acceleration of change; The short-term anomaly feature vector is composed of first / second-order differences, mean, variance, and maximum gradient. S23. Utilize semantic sequences from different sensors or regions to calculate short-term correlations and hysteresis effects, and combine differential, trend, and multi-source correlations to form local pattern vectors to distinguish between single-source false alarms and cross-source consistency anomalies. Specifically: semantic sequence , For each pair of sensors / regions i and j, calculate the short-time correlation coefficient matrix: ; Introducing lag terms Capture delay effect: ; The correlation coefficient, lag, and short-term anomaly features are combined to form a local pattern vector: ; in, and These represent short semantic sequences from sensor or monitoring areas i and j, respectively. Represents the short-term correlation coefficient matrix; Indicates the lag time, used to capture delayed events; Represents the correlation coefficient function; This represents the short-time correlation coefficient with lag terms; Represents the covariance function; and Representing sequences respectively , Standard deviation; Represents a local pattern vector; S24. Anomaly scores are generated based on local pattern weighted quantization. When a score exceeds an adaptive threshold, it is marked as a candidate anomaly pattern, providing input for subsequent causal analysis and intelligent intervention. Specifically: Define a local anomaly scoring function: ; when (Threshold is adaptive, based on edge EMA), marked as candidate anomaly patterns Otherwise, mark it as normal. in, This represents a local anomaly score, quantifying the anomaly intensity of the current window; This represents the local pattern weighting matrix, reflecting the credibility and importance of each feature; This represents the threshold for determining local anomalies, which is adaptively calculated based on the exponential moving average (EMA).
[0020] S3. Based on the candidate anomaly patterns and local anomaly scores identified in step S2, a local causal network with physical semantic constraints is constructed to deduce the propagation path of anomalies within the refueling station and quantify the impact of each node on the overall system risk. A hierarchical risk level determination is performed based on the propagation contribution, ultimately identifying key risk sources and generating intelligent prompts. This achieves a closed-loop deduction from characteristic anomalies to actionable risk conclusions. The specific implementation process is as follows: S31. Construct a set of local causal nodes based on candidate anomaly patterns, and derive the conditional probabilities between nodes through historical data, energy flow direction, and regional physical connections to form a causal relationship matrix with physical semantic constraints. This matrix ensures that anomaly propagation only unfolds along actual physical or energy flow paths, avoiding spurious associations, and providing an interpretable and structurally sound causal basis for subsequent propagation deduction. Specifically: For each candidate anomaly pattern Construct a set of local causal nodes: ; Where node n i This corresponds to specific equipment, sensors, or regional indicators in the energy station (such as energy storage unit temperature, gas concentration, current fluctuations, etc.); Local causal matrix Characterizing the causal probability between nodes: ; For matrix Add physical semantic constraints: This ensures that causal deduction will not result in unreasonable cross-regional abnormal propagation; in, This represents a set of nodes in a local causal network, where each node corresponds to a specific sensor reading, equipment indicator, or sub-region of the station (e.g., tank temperature, filling gun status, hydrogen concentration point, current channel, etc.). Represents a local causal relationship matrix; Indicates when node n i When an anomaly occurs, node n j The conditional probability of an anomaly occurring ( ); S32. Construct a propagation weight model using a causal relationship matrix and combine it with local anomaly scoring. Simulate the diffusion process of anomalies between nodes through multiple iterations to obtain the comprehensive anomaly impact vector after propagation. Then, calculate the node risk contribution coefficient based on the cumulative propagation value to measure the relative impact of each node on the overall system risk, specifically: Define the weighting matrix : ; Calculate the anomaly propagation path : ; Define the risk contribution coefficient: ; in, represents the propagation weight matrix, i.e., the matrix used for iterative propagation calculation; I represents the identity matrix; Indicates the propagation intensity adjustment coefficient; This represents the comprehensive impact vector of abnormal propagation, which is obtained by accumulating and diffusing the initial score through propagation iterations. This represents the number of propagation iterations / propagation depth, controls the maximum number of propagation iterations (or the number of terms truncated in the power series), and limits the propagation depth / duration of anomalies in the topology; The node risk contribution coefficient represents the normalized risk contribution of node i at time k, which measures the proportion of the overall propagation impact of that node. Represents matrix exponentiation, i.e., matrix power. The power of t represents the influence diffusion pattern after t steps of propagation; p represents the total number of nodes. S33. Based on the risk contribution coefficient, an adaptive risk threshold combining historical statistics and equipment differences is introduced, and each node is divided into three risk levels: low, medium, and high, generating a structured risk level matrix. This matrix can be used for alarm triggering and policy linkage, realizing the transformation from quantitative scoring to hierarchical management, specifically: Define a risk threshold vector: ; Risk level assessment: ; Anomaly nodes are classified into different levels based on their severity, and a risk report matrix is generated. Each column corresponds to a low, medium, or high risk level. This hierarchical classification can directly drive intelligent alarm strategies, operational decisions, or automatic protection measures. in, Indicates the contribution coefficient used to normalize the normalized contribution coefficient. The threshold set is mapped to three levels: "low / medium / high". This indicates the set low threshold. This indicates the set medium threshold. This indicates the set high threshold. This represents a risk reporting matrix, which is a structured representation of low / medium / high risk levels by node (each column corresponds to a level); S34. Select key risk sources from high-risk nodes based on their contribution, and generate intelligent prompts by combining information such as node location, risk level, and propagation impact range. These prompts can be directly used by operation and maintenance, scheduling, or automated protection modules. The prompts include the cause of the anomaly, potential impact, and suggested actions, achieving a closed loop of risk identification and decision support. Specifically: For high-risk nodes, consider their contribution to the spread. Sort: ; Select those whose contribution exceeds the threshold. The node is a key source of risk. ; Generate intelligent prompts for key nodes : ; in, Indicates the priority of key risks, in order of The index sequence after sorting high-risk nodes in descending order; This represents the set of key risk sources, i.e., the set of key risk nodes that have been selected. This indicates intelligent prompts, which are the alarm message structures generated for key risk sources; This refers to the set alarm generation function, which maps key node information, level, time, location, etc., into an easy-to-understand and executable prompt text and operation suggestions. This function or rule engine is used to do so.
[0021] S4. Construct a cross-scenario transferable linkage triggering system, enabling the system to automatically identify abnormal triggering chains under different load states, equipment operating conditions, and operational conditions. Through comprehensive deduction of multi-source constraints, adaptive environmental variables, and policy dependencies, it generates a layered policy package that fits the field's executability. The core of this mechanism lies in utilizing the temporal closed-loop relationship of energy flow anomalies to uniformly drive policy selection, conflict resolution, and action-level implementation, thereby achieving dynamic and self-consistent responses to complex industrial sites. The specific implementation process is as follows: S41. Encapsulate the key risk sources, contribution coefficients, and level information from step S3 into structured risk tokens, and match the optimal handling strategy according to the strategy library. Simultaneously, evaluate the strategy priority and business impact to achieve auditable and quantifiable risk decisions. Specifically: For each candidate event Generate risk token: ; ; Maintain a policy library P (local / operations-side synchronized), and each policy p j Includes: applicable conditions (risk level, node type, environmental constraints), sequence of actions, level of authority, estimated impact and fallback measures; Define the strategy matching function: ; Return a set of candidate strategies ; Candidate strategies are scored and ranked according to the following formula: ; in, This represents a risk token object, used to encapsulate candidate risk event information and serve as the input unit for policy matching; ID represents the unique identifier of the risk token; t k Indicates the timestamp of the risk event; This indicates an evidence pointer, which points to locally cached collected evidence (sensor fingerprints, snapshots, causal networks, etc.). It represents the context information of risk events, including environmental parameters (such as wind speed, operational status, and personnel presence), which are used for strategy matching; represents the set of node risk contribution coefficients; P represents the strategy library, which contains all executable intervention strategies, and each strategy includes triggering conditions, action sequence, permission level, and estimated impact. This represents the set of candidate strategies obtained by matching risk tokens, i.e., the set obtained by matching token information with the rules in the strategy library; Indicates the strategy priority score; It represents the utility of a strategy, that is, the estimated effect of the strategy on reducing overall risk; This indicates the business impact cost introduced by the strategy, and the interruption or loss that the implementation of the strategy may cause to the business (such as refueling, charging); This indicates the complexity / cost of policy execution, i.e., the required manpower, permissions, or operational complexity. , and Indicates the weight of the scoring function; S42. Based on strategy priority, implement intervention actions according to the principle of minimum necessary intervention, starting with low-impact soft interventions, and escalating to mandatory isolation or emergency measures when necessary. Combine edge sandbox simulation and transactional actions to ensure execution security and business continuity. Specifically: Level Classification: Level A (Tips / Reminders): Send visual / voice prompts, record them, and suggest manual review; Level B (Soft Intervention / Local Control): Short-term changes to operating parameters (such as reducing charging power by 10%, enabling local ventilation, and temporarily limiting the refueling flow rate) to maintain business continuity; Level C (Automatic Isolation / Forced Shutdown): Cuts off local power supply, isolates storage tanks, or disconnects charging circuit; Level D (Emergency Activation / Remote Linkage): Triggers a station-wide shutdown, links fire / emergency contacts, and switches to the emergency plan; Before execution, the edge controller first performs a rapid simulation in the local sandbox model (near real-time model or playback of similar historical scenarios) to evaluate the impact of the preferred strategy on key indicators (temperature, concentration, personnel safety). If the simulation passes, the release is allowed; otherwise, the suboptimal strategy is selected or manual confirmation is required. Sandbox simulation is based on an approximate physical model or the most recent intervention-response history and is modified with runtime context to ensure timeliness and feasibility. Execution follows transactional steps with rollback points: each action a t Includes expected results, monitoring indicators, and timeout / rollback conditions; action chain. Release sequentially upon successful confirmation; Concurrency control: If multiple highly correlated policies exist at the same time, the execution scheduler performs parallel or serial scheduling according to priority, scope of impact and resource constraints to avoid secondary risks caused by mutual exclusion actions; in, This represents an intervention action chain, which is a set of actions arranged in the order of the strategy to perform a complete intervention; S43. During the intervention process, key indicators are monitored in real time to assess the effectiveness of the action. If any abnormalities occur or expectations are not met, the process is automatically rolled back or escalated. Simultaneously, complete records of execution evidence and monitoring data are maintained to ensure safety, traceability, and auditability. Specifically: Define the set of key monitoring metrics M for the casting action a (For example, changes in fingerprints from relevant sensors, device current / temperature, and personnel location information), the edge controller monitors M using high-frequency sampling. a And calculate the response metrics: ; like If the expected threshold is reached or the value remains within the safe range, the marking action is considered successful. If an unexpected reaction is detected (e.g., deterioration of relevant indicators, conflict of field operations), the rollback action sequence R(a) (based on the definition of action reversibility) is immediately triggered. If the action does not produce the expected improvement within the time limit and the risk indicators continue to rise, the next level of action will be automatically triggered and the on-duty and operations teams will be notified. Every action, monitoring data point, and decision (including sandbox simulation results) is recorded with a signature and linked to a risk token T. k This forms a complete audit package: ; The audit package can be used for post-event analysis, liability determination, and model retraining; Among them, M a Represents the set of monitoring indicators; The response norm is used to measure the actual effect of an intervention on key indicators at time t. This represents the response weight matrix, which assigns different weights to different monitoring indicators and is used to calculate the overall response. and represents the indicator vector, which represents the key monitoring indicator vectors before and after the intervention; Indicates an audit log package; This represents the simulation results of the sandbox action in the local sandbox model, used to verify the safety of the strategy before actual execution. This refers to the action execution log, which records the specific execution of intervention actions on-site, ensuring that each step is traceable; This refers to the execution monitoring log, which is the real-time monitoring data of key system indicators after the intervention action is performed. It is used to evaluate the effect of the action and trigger rollback / upgrade. S44. Quantify the execution effect into indicators, update the strategy library's estimated utility and cost, and supplement the knowledge base with success / failure cases, sandbox simulation results, and manual confirmation results to achieve adaptive strategy optimization and operational collaboration support, providing intelligent guidance for future risk management. Specifically: Define the effect evaluation vector ; according to Update the estimated utility of the strategy library. With execution costs This, in turn, affects future priority scoring; online updates with a forgetting factor are employed. ; Success / failure cases, sandbox simulation parameters, and manual confirmation results are written into the knowledge base for future strategy matching and causal model updates; recurring patterns can be automatically generated into "quick response scripts" and then incorporated into the strategy library after manual review. For medium- to high-level events, the system automatically generates structured reports and pushes them to on-duty engineers and operations decision-makers; it provides a "one-click replay" function (based on token-based evidence pointers) to help humans quickly judge and decide whether to intervene or roll back; any manual confirmation or unauthorized operation is written into the audit chain and triggers compliance alerts; in, This represents the intervention effect evaluation vector, which is used to assess the effectiveness of action execution and provide feedback to the strategy library; This indicates the amount of risk reduction, specifically the decrease in the overall quantitative risk indicator of the system after intervention. This indicates the business impact, specifically the quantification of business interruptions or losses caused by intervention actions. This indicates the system stabilization time, which is the time required from the implementation of intervention to the recovery of system indicators to a safe range. This indicates a strategy utility update, which updates the estimated utility of strategies in the strategy library online. Represents the historical utility of a strategy, and the original utility estimate in the strategy library; This represents the current observed utility, the actual observed utility value after this action is performed; This represents the forgetting factor, which controls the weight of historical and current observational utility to prevent over-reliance on historical experience.
[0022] Example 2: The present invention proposes an intelligent risk identification system for refueling stations, which is applied to the intelligent risk identification method for refueling stations proposed in Example 1. The system includes: a multi-layer sensing and energy flow fingerprinting module, a candidate anomaly pattern recognition and local scoring module, a multi-source causal inference and risk level determination module, and a risk strategy adaptation and intelligent intervention module.
[0023] The multi-layer sensing and energy flow fingerprinting module is responsible for high-frequency acquisition of various physical and environmental parameters within the refueling station, including but not limited to energy storage unit temperature, pressure, current, voltage, gas concentration, operating status, and personnel distribution. After multi-layer sensor fusion and preliminary cleaning, the acquired data is processed through feature extraction and vectorization to form an energy flow fingerprint, which is used to quantify the operating status and potential abnormal signals of each node. The module supports multi-source heterogeneous data fusion, dynamic data storage, and provides standardized vector output to upper-layer modules for subsequent anomaly identification and analysis. The module also has local rapid preprocessing capabilities, which can perform preliminary screening and label spatiotemporal features of sudden signals to ensure data timeliness and integrity. The candidate anomaly pattern recognition and local scoring module combines historical patterns and equipment physical characteristics to identify anomaly patterns. It generates candidate anomaly patterns through multi-dimensional feature clustering and statistical anomaly detection methods, and calculates a local anomaly score for each pattern, reflecting the degree of anomaly within a node or region. The module supports multi-source causal feature analysis, enabling the identification of complex coupled anomalies or weak signal anomalies. It also provides interpretable annotations for anomaly patterns, facilitating subsequent causal deduction and risk level determination. The module's internal pattern library can be dynamically updated to adapt to equipment aging or environmental changes, achieving adaptive anomaly recognition. The multi-source causal inference and risk level determination module analyzes the propagation path of anomalies within the refueling station and quantifies the contribution of each node to the overall risk. The module combines the causal inference results with adaptive risk thresholds to perform hierarchical risk level determination for each node, generating a list of key risk sources and a risk level matrix, providing a quantitative basis for intelligent intervention. The module can process multi-source information, including physical parameters, environmental conditions, historical events, and operation records, to achieve interpretable risk analysis, while providing a highly available, real-time risk monitoring interface. The risk strategy adaptation and intelligent intervention module generates structured risk tokens based on key risk sources and risk level information, and matches the optimal handling strategy in the strategy library. The module executes tiered interventions, ranging from low-impact alerts and soft controls to necessary isolation or shutdown actions, and performs real-time response monitoring, rollback control, and evidence recording during execution to ensure safe and controllable handling. The module supports sandbox simulation to evaluate intervention effects, transactional action orchestration, multi-strategy concurrent scheduling, and online strategy updates. Execution results and monitoring data are fed back to the knowledge base to achieve a learning loop and operational collaboration, providing experience data and strategy optimization basis for future events.
[0024] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.
Claims
1. A method for intelligent risk identification of an energy loading station, characterized in that, The specific implementation steps include the following: S1, by deploying a sensor network inside the energy station and synchronously collecting signals, windowing processing is performed on multi-source signals, multi-scale time-frequency features and physical response features are extracted, heterogeneous features are mapped into energy flow fingerprints through a hybrid projection model, and self-calibration enhancement is performed through a differential baseline; S2, the energy flow fingerprints are decoded and mapped into semantic features, candidate abnormal patterns are identified through short-term trend analysis and multi-source cross-correlation, and local anomaly scores are calculated; S3, a local causal network with physical semantic constraints is constructed based on the candidate abnormal patterns, abnormal propagation paths are deduced and the risk contribution of each node is quantified, risk levels are divided and key risk sources are located; S4, risk information including key risk sources, risk levels and contribution degrees is packaged into a structured risk token, a corresponding disposal strategy is matched from a pre-set strategy library according to the risk token, and a transactional action chain composed of a plurality of actions is executed according to the hierarchical intervention principle, and the execution effect of the action chain is monitored in real time and effect rollback and strategy upgrade are supported.
2. The method of claim 1, wherein, In step S1, deploying a sensor network and synchronously collecting signals specifically includes: Sensor groups are arranged in the refueling area, storage tank area, charging area, equipment room and human flow channel of the energy station; Each group of sensors at least includes a chemical gas sensor, an acoustic sensor, an infrared temperature measurement unit, a differential pressure sensor, a current sampling unit and a behavior capturer; All sensors are connected to edge nodes and time stamp synchronization is ensured by hardware clock and network time protocol; The edge nodes use an overlapping sliding window strategy to slice the signals, and high-pass filtering is performed on the signals to remove low-frequency drift, and low-order filtering or wavelet denoising is performed to remove high-frequency noise, and signals with different sampling rates are resampled to a unified reference frequency.
3. The method of claim 2, wherein, In step S1, the multi-scale time-frequency features and physical response features are extracted, which specifically includes: Time-frequency transformation is performed on the sensor signals in each window, and the energy integral values of the signals in a plurality of preset frequency bands are calculated as basic quantities; The transient peak value, envelope amplitude rate and spectral centroid of the signal are calculated as descriptors; The response time constant, sensitivity nominal value, local hot spot area, temperature gradient or frequency band distribution physical parameters of the sensor itself are combined to form a window-level original feature vector of each sensor; The original feature vectors of all sensors are spliced into an overall feature vector.
4. The method of claim 3, wherein, In step S1, the heterogeneous features are mapped into energy flow fingerprints through a hybrid projection model, which specifically includes: A linear transformation matrix with physical prior is used to project the overall feature vector onto a physical basis to obtain a linear projection result; A small neural network is used to perform nonlinear compensation on the linear projection result; The compensation result is subjected to component normalization processing to obtain a normalized energy flow vector; The uncertainty vector of the normalized energy flow vector is obtained according to the sensor confidence and feature noise estimation.
5. The method of claim 4, wherein, In step S1, self-calibration enhancement is performed using a differential baseline, which specifically includes: Short-term baseline and long-term baseline are maintained, the short-term baseline is the sliding average of a certain number of windows in the past, and the long-term baseline is the sliding average of a longer historical window; The difference between the current normalized energy flow vector and the short-term baseline and the long-term baseline is calculated to obtain a short-term difference vector and a long-term difference vector; The short-term difference vector and the long-term difference vector are combined to form a final difference response; The long-term baseline is updated using an exponential moving average method with a forgetting factor; The weighted norm of the final difference response is calculated as a local anomaly score; The normalized energy flow vector and the final difference response are quantized and hashed to generate a compact fingerprint package.
6. The method of claim 5, wherein, Step S2 includes: The received compact fingerprint package is dequantized to recover an approximate energy flow vector; The approximate energy flow vector is mapped to a semantic space through a semantic projection matrix to form a semantic vector; A short-term sequence is constructed based on the semantic vectors in consecutive time windows; First-order and second-order differences of the semantic vectors in the short-term sequence are calculated; The mean, variance, and maximum gradient value of the short-term sequence are calculated; The first-order and second-order differences, the mean, the variance, and the maximum gradient value are combined into a short-term anomaly feature vector.
7. The method of claim 1, wherein the method further comprises: Step S2 includes: A short-term correlation coefficient matrix is calculated for short-term semantic sequences of different sensors or regions; A lagged short-term correlation coefficient with a lag term is introduced to capture delay effects; The short-term correlation coefficient, the correlation coefficient with the lag term, and the short-term anomaly feature vector are combined to form a local pattern vector; A local anomaly score is calculated by weighted calculation of the local pattern vector; The local anomaly score is compared with an adaptive threshold based on an exponential moving average, and if it exceeds the threshold, it is marked as a candidate abnormal pattern.
8. The method of claim 7, wherein the method further comprises: Step S3 includes: For the candidate abnormal pattern, determine the related devices, sensors, or regional indicators within the energy injection station as causal nodes; Based on historical data, a conditional probability matrix is established to represent the causal relationship between nodes; Physical semantic constraints are imposed on the conditional probability matrix to ensure that the causal relationship conforms to the actual physical or energy flow path; Based on the conditional probability matrix and the local anomaly score, a propagation weight matrix is constructed; The diffusion process of the anomaly along the propagation weight matrix is simulated by iterative calculation to obtain a comprehensive anomaly impact vector for each node; The risk contribution coefficient of each node is calculated based on the comprehensive anomaly impact vector.
9. The method of claim 8, wherein, Step S3 includes: An adaptive risk threshold vector is set based on historical statistics and device differences, which includes low, medium, and high risk thresholds; The risk contribution coefficient of each node is compared with the threshold vector to divide it into low, medium, or high risk levels, and a risk level matrix is generated; High-risk nodes are sorted in descending order of their risk contribution coefficients; High-risk nodes with risk contribution coefficients exceeding a preset contribution threshold are screened as key risk sources; Intelligent prompt information including abnormal causes, potential impacts, and suggested operations is generated based on the location, level, and propagation impact range of the key risk sources.
10. An intelligent risk identification system for an energy station, which is used to execute the intelligent risk identification method of any one of claims 1-9. Multi-layer perception and energy flow fingerprinting module, used for deploying and synchronously controlling sensor networks inside the energizing station, collecting multi-source signals, windowing preprocessing the signals, extracting multi-scale time-frequency features and physical response features, generating normalized energy flow vectors through hybrid projection model, self-calibrating and generating compact energy flow fingerprint streams by using differential baseline; Candidate anomaly pattern recognition and local scoring module, used for decoding and mapping energy flow fingerprints into semantic vectors, constructing short-term sequences and calculating short-term anomaly feature vectors, generating local pattern vectors through multi-source cross-correlation analysis, calculating local anomaly scores and identifying candidate anomaly patterns; Multi-source causal reasoning and risk level determination module, used for constructing local causal networks with physical semantic constraints according to candidate anomaly patterns, deducing anomaly propagation paths and calculating node risk contribution coefficients, dividing node risk levels based on adaptive threshold, locating key risk sources and generating intelligent prompts; Risk policy adaptation and intelligent intervention module, used for encapsulating key risk sources, risk levels and context information into risk tokens, matching and sorting disposal strategies from the strategy library, executing transactional action chains according to the hierarchical principle from prompt reminder, soft intervention, automatic isolation to emergency start, evaluating intervention measures through sandbox simulation, monitoring execution effect in real time and supporting rollback and policy upgrade.
Citation Information
Patent Citations
Thermal power equipment real-time monitoring method and system based on edge calculation
CN120217127A
Cited By
Intelligent monitoring method and system for skid-mounted gas station
CN121882726A
