An energy station digital operation system and method based on the internet of things

Abstract

This invention relates to the field of operation management technology, specifically to a digital operation system and method for refueling stations based on the Internet of Things (IoT). The system collects multi-source data through an IoT sensing layer using a dynamic priority sampling and scheduling mechanism, and performs layered lightweight feature extraction via an edge computing gateway to generate key indicator vectors. Based on a dynamic spatial grid refinement mechanism, the data is mapped to a three-dimensional spatial grid model and assembled into a spatiotemporal data cube. A composite risk association identification model is employed, performing single-modal dynamic baseline assessment and cross-modal spatiotemporal association pattern matching in parallel to calculate a comprehensive risk score. The score is compared with dynamic dual thresholds to execute an ordered set of instructions for attention warnings or linkage alarms. The digital management presentation layer achieves closed-loop event recording and model self-optimization. This invention realizes a single-map perception of the entire refueling station's status, improves the accuracy and lead time of composite risk warnings, achieves millisecond-level linkage control, and forms a continuously evolving safety operation knowledge base.

Description

Technical Field

[0001] This invention relates to the field of operation management technology, and in particular to a digital operation system and method for energy stations based on the Internet of Things. Background Technology

[0002] With the diversification of transportation energy structures, "integrated energy stations" that combine charging, battery swapping, refueling, gas filling, and hydrogen refueling have become key nodes in new urban infrastructure. These stations are characterized by diverse energy forms, complex equipment types, and overlapping operational risks.

[0003] Currently, the operation and management of most refueling stations still rely primarily on localized, independent subsystems, such as refueling / gas filling control systems, charging pile management systems, video surveillance systems, fire alarm systems, and perimeter security systems. These systems are typically provided by different vendors, using proprietary protocols, resulting in data silos. Daily inspections, equipment checks, and risk assessments by management personnel heavily depend on paper records and personal experience, lacking a comprehensive, real-time, and digital understanding of safety status. When safety incidents such as equipment malfunctions, gas leaks, temperature overloads, or personnel violations occur, the systems cannot automatically coordinate and respond. Alarm information is scattered and delayed, heavily relying on passive manual responses, leading to poor emergency response timeliness and increasing the risk of major safety accidents.

[0004] Therefore, how to break down the information barriers between various safety-related subsystems in existing refueling stations and build a unified, IoT-based digital safety operation platform with real-time perception, intelligent analysis, proactive early warning, and coordinated control capabilities is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a digital operation system and method for energy stations based on the Internet of Things, which improves the accuracy and lead time of complex risk early warning and achieves millisecond-level linkage control.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a digital operation method for energy stations based on the Internet of Things, comprising the following steps: The IoT sensing layer at the power station site collects environmental safety parameters, equipment status parameters, and personnel behavior data at a frequency set by a dynamic priority sampling and scheduling mechanism. The edge computing gateway then performs data cleaning, time alignment, and layered lightweight feature extraction to generate key indicator vectors. The key indicator vectors are mapped to the corresponding spatial grid units in the three-dimensional spatial grid model of the power station according to the dynamic spatial grid refinement mechanism. Then, the multimodal data vectors of each spatial grid unit in the current time window are assembled by weighted interpolation, thereby constructing a spatiotemporal data cube. Using a composite risk association identification model, single-modal baseline assessment and cross-modal spatiotemporal association pattern matching are performed in parallel on the multimodal data vectors of each spatial grid cell to obtain independent deviation score vectors and rule matching score vectors, which are then fused to obtain the comprehensive risk score of the spatial grid cell. The comprehensive risk score is compared with the dynamically adaptive first threshold and second threshold. When the score reaches or exceeds the first threshold but is lower than the second threshold, a warning response is executed. When the score reaches or exceeds the second threshold, a linkage alarm response is executed. An ordered instruction set is automatically generated and sent to the field execution equipment through the Internet of Things execution gateway. Complete data from each security incident is packaged into a structured security incident instance and stored in a historical database. The database is then periodically analyzed to generate parameter adjustment suggestions for the composite risk association identification model, which are then updated after confirmation.

[0007] The dynamic priority sampling and scheduling mechanism specifically includes: The dynamic sampling schedule table for each sensing terminal is obtained, and the current sampling rate is dynamically adjusted according to the attention score of the terminal data in the previous scheduling cycle. The attention score is calculated in real time based on the degree to which the instantaneous value of the terminal exceeds the historical dynamic envelope, the abnormal state of adjacent terminals, and the comprehensive risk score of the spatial grid unit to which it belongs in the previous cycle. When the attention score exceeds the score threshold, the sampling frequency is increased; otherwise, it is gradually reduced to the basic sampling frequency.

[0008] The hierarchical lightweight feature extraction includes: The first sublayer extracts temporal feature primitives from continuous analog quantities, including instantaneous statistical features, dynamic trend features, and spectral energy features. The second sublayer generates structured event features from non-video discrete events and image / video data, including electronic fence events calculated from UWB positioning data and semantic labels for helmet wearing status, mobile phone holding actions, and flame / smoke areas detected from video frames by embedded neural network models. The third sublayer generates cross-channel correlation features between different perception channels managed by the same edge computing gateway. When the feature primitives of multiple channels show a positive correlation exceeding a preset threshold, correlation feature labels are generated.

[0009] The dynamic spatial grid refinement mechanism is as follows: The system acquires the preset base resolution of the basic 3D spatial grid model and automatically subdivides or merges the grid locally based on the real-time risk distribution and the density of sensing terminals. When the comprehensive risk score of a grid cell exceeds 80% of the first warning threshold or the number of sensing terminals deployed in the grid exceeds the preset density limit, the grid is divided into multiple sub-grids in the horizontal direction. When the comprehensive risk score of a grid is close to zero for a long period of time and no sensing terminals are deployed, it is merged with adjacent grids of the same type. Grid changes are updated to the spatial grid index table in real time and re-evaluated every five minutes.

[0010] The weighted interpolation assembly includes: For each spatial grid cell, according to the predefined multimodal data vector template, the values ​​of each asynchronously arriving feature field are filled with the current time window by the most recent valid value plus time decay weight, linear interpolation, or the cumulative duration of the event. When the same feature field has multiple independent sources, a weighted average fusion is performed based on the historical accuracy of the sensor and the data age, and a consistency check is performed to remove outliers.

[0011] Among them, the single-modal baseline assessment adopts a dynamic baseline model, which maintains a time-series prediction baseline, multi-scale fluctuation envelope and mode state memory for each continuous feature field. For the current value of the input feature, the current operating mode is first matched to select the corresponding prediction mean and prediction standard deviation, the standardized residual is calculated, then the trend enhancement factor is introduced and multi-scale envelope verification is performed, and finally it is converted into an independent deviation score of 0 to 100 through the S-type mapping function.

[0012] Among them, cross-modal spatiotemporal correlation pattern matching is performed based on a composite risk rule base. Each rule includes the applicable spatial range, time window length, preconditions, condition pattern, and rule weight. The matching process includes: selecting an applicable rule subset based on the current grid functional block and environmental conditions, extracting the historical vector sequence of the corresponding time window, calculating the continuous clause matching degree for each condition clause, and then obtaining the overall rule matching score through weighted average or fuzzy logic operation.

[0013] Secondly, the present invention provides an IoT-based digital operation system for energy stations, applied to an IoT-based digital operation method for energy stations as provided in the first aspect, comprising: The IoT sensing layer includes environmental safety sensing terminal groups, device status sensing terminal groups, personnel and behavior sensing terminal groups, and edge computing gateways. It is used to collect multi-source heterogeneous data with a dynamic priority sampling and scheduling mechanism and perform hierarchical lightweight feature extraction at the edge to output key indicator vectors. The digital twin and data fusion layer is deployed on a station-level local server or in the cloud. It includes a device digital twin model library, a spatiotemporal data alignment engine, and a dynamic spatial grid refinement module, which is used to map key indicator vectors to a three-dimensional spatial grid model and assemble them into a spatiotemporal data cube. The multimodal risk identification and analysis layer includes a single-modal baseline evaluator, a cross-modal spatiotemporal association rule engine, and a risk score calculator, which are used to calculate the comprehensive risk score for the multimodal data vector of each spatial grid cell; The security policy execution and linkage control layer includes a policy instruction set generator and an IoT execution gateway, which are used to generate and issue ordered instruction sets to field execution devices based on the comparison results of the comprehensive risk score and the first threshold and the second threshold. The digital operations management presentation layer includes a holographic security posture dashboard, an incident handling closed-loop workbench, and a security policy optimizer, which are used to realize the visualization of security posture, closed-loop recording of incidents, and model self-optimization.

[0014] The strategy instruction set generator internally maintains a risk mode-instruction template mapping table and integrates an instruction timing orchestration and conflict detection module. This module divides instructions into parallel execution groups and sequential execution groups based on the dependencies between instructions, and performs mutual exclusion processing on instructions with resource conflicts according to the safety priority principle. The IoT execution gateway supports instruction confirmation and retry mechanisms, atomic control of instruction execution chains, and local persistent black box records to ensure that linkage control can still be completed when the network is interrupted.

[0015] The digital operation management presentation layer's closed-loop event handling workbench automatically generates standardized handling suggestion cards for each linked alarm event and displays a list of automatically executed instructions and their status in a pop-up window on the station control screen. After on-site handling is completed, on-duty personnel can upload the handling results, on-site photos, and debriefing conclusions to the workbench. The system automatically packages the entire chain data of the event into a security event instance and stores it in the historical database. The security policy optimizer periodically analyzes historical event instances and generates adjustment suggestions for the first threshold, the second threshold, and the rule weights and dynamic baseline parameters in the composite risk association identification model. After confirmation by management personnel, these suggestions are fed back to the corresponding level.

[0016] A digital operation system and method for a fueling station based on the Internet of Things according to the present invention collects multi-source data through an IoT sensing layer deployed with a sensing terminal group for environment, equipment and personnel via a dynamic priority sampling scheduling mechanism, and an edge computing gateway performs layered lightweight feature extraction to generate key indicator vectors; a digital twin and data fusion layer maps the data to a 3D spatial grid model based on a dynamic spatial grid refinement mechanism to assemble a spatiotemporal data cube; a multi-modal risk identification and analysis layer adopts a composite risk association identification model to perform single-modal dynamic baseline assessment and cross-modal spatiotemporal association pattern matching in parallel to calculate a comprehensive risk score; a security policy execution layer compares the score with dynamic dual thresholds to execute an ordered instruction set of attention warning or linkage alarm; and a digital management presentation layer realizes event closed-loop recording and model self-optimization. The present invention realizes full-state one-map perception of the fueling station, improves the accuracy and advance time of composite risk warning, achieves millisecond-level linkage control, and forms a continuously evolving safe operation knowledge base. Brief Description of the Drawings

[0017] To more clearly describe the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings required in the description of the embodiments or the prior art will be briefly introduced below.

[0018] Figure 1 is a schematic diagram of steps of the digital operation method for a fueling station based on the Internet of Things according to the first embodiment of the present invention.

[0019] Figure 2 is a schematic flowchart of the digital operation method for a fueling station based on the Internet of Things provided by the present invention.

[0020] Figure 3 is a schematic structural principle diagram of the digital operation system for a fueling station based on the Internet of Things according to the second embodiment of the present invention. Detailed Description of the Embodiments

[0021] Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numerals in different drawings refer to the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application.

[0022] The terms used in the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a", "the" and "said" used in the present application and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0023] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0024] The first embodiment of this application is as follows: Please see Figures 1-3 This invention provides a digital operation method for energy stations based on the Internet of Things, comprising the following steps: S1. The IoT sensing layer at the power station site collects environmental safety parameters, equipment status parameters, and personnel behavior data at a frequency set by a dynamic priority sampling and scheduling mechanism. The edge computing gateway then performs data cleaning, time alignment, and layered lightweight feature extraction to generate key indicator vectors.

[0025] Specifically, the IoT sensing layer comprises three heterogeneous terminal groups: environmental safety sensing terminal group, equipment status sensing terminal group, and personnel and behavior sensing terminal group. Each terminal group includes various physical types of sensors or data acquisition devices. For example, in the environmental sensing group, there are catalytic combustion combustible gas sensors, infrared spectroscopy micro-leakage detectors, and thermocouple temperature sensors; in the equipment status sensing group, there are Hall current sensors, vibration accelerometers, and ultrasonic wave sensors; and in the personnel and behavior sensing group, there are Hisilicon embedded smart cameras, UWB positioning base stations, and nine-axis inertial measurement unit smart badges.

[0026] Each edge computing gateway maintains a dynamic sampling schedule table, which records three core parameters for each connected sensing terminal: the basic sampling frequency (preset by the security importance level of the device type and installation location), the current dynamic sampling multiplier (initially 1x), and the "attention score" of the terminal's data in the previous scheduling cycle. The "attention score" is calculated in real time by the lightweight anomaly prediction module built into the gateway, based on: whether the terminal's instantaneous value exceeds the 80th percentile of its own historical dynamic envelope; whether anomalies have occurred in neighboring terminals (spatial related parties); and the current comprehensive risk score of the spatial grid unit to which the terminal belongs (the initial value of this score is provided by feedback from a higher level in the previous frame; if there is no feedback, it is set to zero).

[0027] When the attention score of a terminal exceeds a set threshold, the edge computing gateway instructs the terminal to temporarily increase its sampling frequency to "base sampling frequency × current dynamic sampling rate × 2", while updating the dynamic sampling rate to twice its original value. A decay timer is set; if there is no sustained high attention, the rate will be halved every 10 sampling cycles until it returns to 1. Conversely, when the terminal's attention score remains zero for an extended period, the rate can be reduced to 0.5 times, while still ensuring that the minimum sampling frequency is not lower than the minimum resolution required by security specifications.

[0028] The raw data streams generated by each sensing terminal at the aforementioned dynamic frequency are fed into the edge computing gateway of their respective station area (such as charging area, refueling area, or storage tank area) via wired (RS485, CAN, Ethernet) or wireless (Zigbee, Wi-Fi, 5G) methods. Each edge computing gateway has a built-in multi-protocol parsing engine that can automatically identify and parse data frames of different formats such as Modbus RTU, Modbus TCP, OPC UA, MQTT, and RTSP (video stream), extracting device identifiers, timestamps, and sensor measurement values.

[0029] After completing protocol parsing, the edge computing gateway performs coarse spatiotemporal alignment. For non-video data, the gateway buffers arriving data packets according to a fixed time window (e.g., 100 milliseconds). Since the actual sampling times of different terminals are not strictly synchronized, the gateway adopts the "nearest neighbor timestamp matching" principle: at the end of each time window, for each terminal's data within that window, the nearest valid sample value to the end of the window is taken as the representative value for that terminal in that window; if a terminal has no data arriving within the window (possibly due to communication jitter), the representative value of the previous window is retained and a data missing flag is added. For video stream data, the gateway does not perform pixel-level window alignment, but instead caches its continuous frames as video segments and records the precise timestamp of each keyframe for subsequent fine spatiotemporal alignment by the upper layer.

[0030] After coarse alignment, uploading the raw data stream directly in its entirety would consume significant network bandwidth and burden cloud computing. Therefore, the edge computing gateway performs layered, lightweight feature extraction for different data types, uploading only the extracted key indicator vectors, anomaly feature fragments, and necessary compressed raw data packages to the digital twin and data fusion layer. Specifically, this includes the following three sub-layers: First sub-layer: Temporal feature primitive extraction. For each continuously changing physical quantity, the gateway maintains a sliding window of length N (N adapts to the signal change rate; for example, the current window takes 200 sampling points, and the gas concentration window takes 20 sampling points). At the end of each window's sliding step, the following set of lightweight feature primitives is calculated: Instantaneous statistical characteristics: maximum value, minimum value, mean value, and standard deviation within the window.

[0031] Dynamic trend characteristics: The first-order linear slope (representing the rate of increase or decrease) and the second-order difference mean (representing acceleration) of the data sequence within the window are fitted using the least squares method. Among them, the first-order slope is a key indicator for judging risks such as thermal runaway and accelerated leakage.

[0032] Spectral energy characteristics: For vibration or acoustic signals, the gateway runs a simplified Fast Fourier Transform (FFT) module to extract energy amplitudes in a specified frequency band (e.g., near the frequency of a specific fault characteristic of the device) as features, without retaining the full-band spectrum.

[0033] The total data volume of the aforementioned feature primitives is much smaller than that of the original sampling sequence. At the same time, the gateway calculates the "information entropy" or "outlier degree" of the data within the window. If a significant time-domain impact pulse or abnormal energy peak in the spectrum is found, the short-time original waveform segment (e.g., 50 sampling points before and after the impact) is lightly compressed (e.g., using run-length encoding or differential encoding) and temporarily stored as optional auxiliary data to be uploaded.

[0034] Second sub-layer: Structured event feature generation (for non-video discrete events and image / video).

[0035] For UWB positioning data, the gateway calculates the two-dimensional or three-dimensional coordinates of each smart badge in real time, and generates discrete structured event tags such as "entering the area", "leaving the area", "timeout of stillness in the area", and "fall detection trigger" based on the preset electronic fence area (such as the boundary of the storage tank area and the charging parking space line), while recording the occurrence time and duration of stay.

[0036] For video streams captured by smart cameras, the gateway does not directly upload continuous video. Instead, it loads a lightweight embedded neural network model locally (with parameters less than 2MB after model pruning and quantization). This model detects key semantic information in video frames in real time: human keypoint skeletons, helmet wearing status, phone holding actions, falling postures, and pixel areas of flames or smoke. The model outputs structured event labels and confidence scores, and compresses the corresponding video frames or short clips (1 second before and after the event) for the detected events using H.265. The original video stream is only buffered in a loop within the gateway for the most recent 30 minutes and is not uploaded.

[0037] Third sub-layer: Cross-channel correlation feature generation.

[0038] Preliminary cross-feature correlations are performed between different sensing channels managed by the same edge computing gateway to generate higher-order "correlation features". For example, the gateway simultaneously manages the current harmonic characteristic element of a charging pile and the vibration acceleration characteristic element installed on the charging pile's casing. When the window slope of the current harmonic distortion rate (THDi) and the energy amplitude of the vibration acceleration in a specific frequency band show a positive correlation exceeding 0.7 in two consecutive windows, the gateway generates a label of "Loosening of power devices inside the charging pile - abnormal discharge correlation feature" as one of the key indicators.

[0039] After the above-described hierarchical extraction, the edge computing gateway summarizes all terminals and cross-channel correlation features within each time window (e.g., 1 second) and assembles them into a key indicator vector. This vector contains the following fields: timestamp (end time of the window), spatial grid code (predefined grid ID corresponding to the area managed by the gateway), feature primitives for each physical quantity / event quantity, and a list of cross-channel correlation feature labels.

[0040] The gateway also maintains an upload priority queue. Based on the comprehensive risk score of each spatial grid cell (this score is fed back from the upper layer every 500 milliseconds, or estimated using a local simplified risk assessment model if there is no feedback), the key indicator vector containing high-risk grid data is placed at the head of the queue and allowed to be uploaded to the digital twin and data fusion layer via 5G or wired network with the shortest latency (<50 milliseconds). For key indicator vectors of low-risk grids, they are batch-packaged and transmitted using non-real-time reliable transmission. In addition, for previously stored abnormal raw waveform segments or key video segments, the gateway only uploads them along with the key indicator vectors when explicitly requested by the upper layer or when the local risk assessment reaches the "attention warning" level; otherwise, they are retained in local circular storage for later traceability.

[0041] The edge computing gateway ultimately outputs two parts: first, a periodic key indicator vector stream, which accounts for approximately 1% to 5% of the total original data, and is pushed to the upper layer in real-time or near real-time; second, supplementary evidence packages triggered on demand, including short-term original waveforms, key video frames, complete screenshots, etc., for upper-layer verification and model training. Simultaneously, the gateway also stores a complete local circular cache of the original data (capacity for the most recent 24 hours) to support rapid local playback and fault tracing, without relying on a cloud connection.

[0042] S2. Based on the dynamic spatial grid refinement mechanism, the key indicator vectors are mapped to the corresponding spatial grid units in the three-dimensional spatial grid model of the power station. Then, through weighted interpolation, a multimodal data vector of each spatial grid unit in the current time window is formed, thereby constructing a spatiotemporal data cube.

[0043] Specifically, firstly, in the digital twin and data fusion layer, a basic three-dimensional spatial grid model is pre-established for each refueling station. This model, based on the actual Building Information Model (BIM) and site plan of the refueling station, divides the entire station area into several regular, non-overlapping three-dimensional grid units. Each grid unit has a unique spatial grid code, which adopts a composite structure of "hierarchy-block-coordinate": the first level represents functional blocks (such as charging areas, refueling areas, storage tank areas, and station buildings); the second level represents sub-areas (such as the charging area, which can be further subdivided into row A pile positions and row B pile positions); and the third level is the specific coordinate grid, with a basic resolution of 0.5m × 0.5m × 0.5m (length × width × height). The height dimension is mainly layered according to the equipment installation height and the standing height of personnel.

[0044] This invention introduces a dynamic spatial grid refinement mechanism. Based on real-time risk distribution and sensing terminal density, this mechanism automatically subdivides or merges the grid locally.

[0045] The triggering conditions for grid subdivision are as follows: When the comprehensive risk score (output feedback from the risk identification layer of the previous cycle) of a certain grid cell within the current time window exceeds 80% of the first warning threshold, or when the number of sensing terminals deployed in the grid exceeds the preset density limit (e.g., more than 3 types of terminals), the system divides the grid cell into four equal sub-grids on the XY plane (the size becomes 0.25m × 0.25m), while the height direction remains unchanged. The sub-grids inherit the spatial grid code of the original grid and append a sub-region identifier at the end.

[0046] Mesh merging trigger conditions: For areas where the comprehensive risk score is close to zero over a long period (e.g., more than 24 hours) and there are no sensing terminals deployed (e.g., open lanes), the system merges it with adjacent meshes of the same type into a larger mesh (the size can be 1 meter × 1 meter) to reduce the computational burden.

[0047] Dynamic maintenance: The refinement and merging of the spatial grid is not static, but is reassessed every 5 minutes based on the latest risk score distribution and terminal activity, resulting in an adaptive spatial resolution. All changes are updated in real time to the spatial grid index table for mapping and querying.

[0048] Each key indicator vector output in step S1 contains the device ID or spatial location information of the source sensing terminal. The digital twin and data fusion layer first parses the "device identifier" or "installation coordinates" field in this vector and associates it with a pre-set terminal-grid mapping table. This table records the spatial grid code to which each sensing terminal belongs. The specific mapping rules are as follows: Point-based sensing terminals: The finest spatial grid cell to which a terminal belongs is determined directly based on its installation coordinates or the center point of its coverage area.

[0049] For area-based or region-based sensing data: For this type of data, the system calculates the percentage of overlap between the polygonal boundary of the detection area and each spatial grid cell. If a detected event covers three adjacent grid cells with overlap areas of 60%, 30%, and 10%, respectively, the structured label and its confidence level for the event will be proportionally allocated to these three grid cells, with each grid cell receiving a weighted copy of the label.

[0050] Cross-channel association features: Association features generated by the edge computing gateway are associated with multiple terminals. Based on the common operating area of ​​these terminals, the system typically takes all grid cells within the convex hull (minimum convex polygon cover) of their respective grids and copies the association feature to each grid cell, while attaching a distance weight.

[0051] For all different data sources mapped to the same grid cell, the system also records their data age, which is the time difference between the data generation time and the current processing time. If a grid cell does not receive any new key indicator vectors within a preset time window, the state of the cell is marked as "data missing," and the fusion result of the previous window is used, but with an additional attenuation coefficient.

[0052] After spatial mapping, each spatial grid cell receives key indicator features from different sources and at different times within a given time window (because the sampling frequency and edge upload latency of each terminal are inconsistent). To form a unified "snapshot," the system performs time window alignment and weighted interpolation assembly. The specific process is as follows: Time window definition: The system uses a fixed-length time window (e.g., 1 second) as the basic step size, and each window is identified by its end timestamp (e.g., T0, T1, T2, ...). For each window, the system collects all key indicator features mapped to the current grid cell within that window (i.e., the time period (T-1, T)).

[0053] Multimodal data vector template: Each spatial grid cell predefines a fixed-length multimodal data vector structure, which contains various security feature fields that the grid may involve.

[0054] Weighted interpolation assembly of asynchronous data: Since not every feature field receives a new measurement value in every time window, this invention uses the method of "most recent valid value + time decay weight" to fill the data, and introduces linear interpolation for key dynamic features.

[0055] For slowly changing physical quantities (such as ambient combustible gas concentration or ambient temperature): take the most recent valid measurement before the current window and multiply it by a time decay factor (e.g., the weight is multiplied by 0.95 for each window). If there is no new data for several consecutive windows, the validity flag of this field gradually decreases until it falls below a threshold and is marked as "invalid".

[0056] For rapidly changing physical quantities (such as equipment temperature slope and current harmonics): Since these characteristics are directly related to dynamic risks, the system will find the two most recent valid measurements before and after the current window, perform linear interpolation on the timestamp, and estimate the approximate value at the end of the current window.

[0057] For discrete event labels: These labels are typically transient. If the system detects such an event label within the current window, it marks it as "active"; otherwise, it marks it as "inactive." If the event label needs to represent duration (such as "duration"), the cumulative duration from the start of the event to the end of the current window is recorded.

[0058] Multi-source confidence fusion: When a feature field receives values ​​from multiple independent sources within the same window, the system uses a weighted average for fusion. The weights are determined by the sensor's historical accuracy and the current data age. A fused confidence score is also generated and appended to the vector.

[0059] Outlier Detection and Correction: During assembly, the system also performs a rapid consistency check. For example, if the "ambient temperature" of a grid is simultaneously obtained from three sensors, two of which are around 25°C, and the third suddenly reports 85°C, the system determines 85°C as an outlier, discards it, and records the anomaly event of that sensor. If the values ​​from all sources contradict each other and exceed a reasonable range, the feature field is marked as "unreliable," and a system alarm is triggered.

[0060] After assembling each spatial grid cell and each time window as described above, the system organizes the multimodal data vectors of all grid cells according to the "time dimension - spatial grid dimension," forming a continuous, three-dimensional spatiotemporal data cube. Wherein: Timeline: The timestamp sequence ends with a window that increases in fixed steps (1 second).

[0061] Space axis: A list of all active space grid codes.

[0062] Feature axes: The feature fields in the multimodal data vector of each grid cell.

[0063] The spatiotemporal data cube is not generated entirely in memory at once, but is updated in real time using a "sliding window" method: each time a new time window data slice (i.e., the vector set of all grids at the current moment) is generated, it is immediately pushed to the next layer (multimodal risk identification and analysis layer), while old slices that exceed the historical retention period (e.g., 30 days) are discarded or dumped to cold storage.

[0064] Finally, this spatiotemporal data cube, as the output of step S2, is continuously supplied to the multimodal risk identification and analysis layer in step S3 in the form of a data stream. Simultaneously, the latest slice of the cube is also sent to the digital operation management presentation layer for real-time rendering of the entire site's security posture on a 3D map.

[0065] S3. Using the composite risk association identification model, single-modal baseline assessment and cross-modal spatiotemporal association pattern matching are performed in parallel on the multimodal data vectors of each spatial grid cell to obtain independent deviation score vectors and rule matching score vectors, which are then fused to obtain the comprehensive risk score of the spatial grid cell.

[0066] Specifically, the composite risk association identification model. This model is not a simple threshold comparator or a single neural network, but a hierarchical, interpretable, and self-learning hybrid architecture. Its internal structure mainly consists of three sequentially cooperating modules: a single-modal baseline estimator, a cross-modal spatiotemporal association rule engine, and a risk score calculator. The first two modules process the input current data vector in parallel, outputting an "independence deviation score vector" and a "rule matching score vector," respectively. Finally, the risk score calculator merges these two to generate a comprehensive risk score.

[0067] The model's input is a multimodal data vector, denoted V_T, representing a spatial grid cell at the end of time window T. This vector is a fixed-length ordered list of features, each feature field having a clear physical meaning and data type (continuous value or discrete label). The model's output is a comprehensive risk integral between 0 and 100, along with the corresponding risk level ("Normal", "Warning", "Linked Alarm").

[0068] Internally, the single-modal baseline evaluator independently analyzes each continuous feature field (such as temperature, concentration, and current harmonics) in V_T, determining its deviation from its normal behavior pattern and outputting a vector composed of independent deviation scores for each feature. The cross-modal spatiotemporal association rule engine treats multiple feature fields in V_T as a whole, matching them against rule patterns in a predefined composite risk rule base to determine if the current data vector triggers a known composite risk type, outputting a set of rule matching scores (one score per rule). Finally, the risk score calculator merges these two scores according to dynamic weights, considering historical risk trends, to calculate the final comprehensive risk score.

[0069] The goal of unimodal baseline assessment is to compute an independent deviation score for each continuous feature field in a multimodal data vector. This score reflects the degree of anomalousness of the current measurement relative to the feature's historical normal range of fluctuation. To avoid false alarms or missed alarms caused by using fixed thresholds, this invention introduces a dynamic baseline model that adaptively learns the normal range of variation for each feature over different time periods.

[0070] For each continuous feature field in each spatial grid cell, the system establishes an initial dynamic baseline model during the historical data accumulation phase (the first 72 hours of normal operation data). This model consists of three core parts: Time series prediction baseline: A lightweight time series prediction algorithm is used to predict the expected value μ_pred of the feature in the current time window and the prediction uncertainty interval (standard deviation σ_pred) based on historical values ​​from several past time windows. The prediction model is updated periodically (e.g., retrained every hour) to adapt to long-term trends such as equipment aging and seasonal environmental changes.

[0071] Multi-scale fluctuation envelope: In addition to point prediction, the model also records the normal fluctuation range of the feature at multiple time scales: short time scale (the magnitude of change in the past minute), medium time scale (the maximum and minimum values ​​in the past hour), and long time scale (the statistical distribution in the past 24 hours).

[0072] Mode state memory: The model also maintains a "mode state" dictionary related to this feature. The system automatically determines the current mode by recognizing the current and voltage characteristics of the charging pile, and maintains different dynamic baseline parameters for different modes.

[0073] For the current input feature value x, the single-modal baseline estimator performs the following steps to calculate its deviation score (values ​​range from 0 to 100, where 0 indicates complete normality and 100 indicates extreme abnormality): Step 1: Pattern Matching and Baseline Selection. Based on auxiliary fields related to pattern judgment in the current data vector (such as the current value of the charging pile), determine the current operating mode of the feature. Then, select the predicted mean μ_pred and predicted standard deviation σ_pred for the corresponding mode from the dynamic baseline model.

[0074] Step 2: Calculate the absolute deviation. Calculate the standardized residual between the current value x and the predicted mean μ_pred: d = |x - μ_pred| / σ_pred. This value represents the standard deviation factor by which the current value deviates from the expected value.

[0075] Step 3: Trend Enhancement Factor. To prevent long-term anomalies that drift slowly but do not exceed the standard deviation threshold from being ignored, the evaluator also calculates the first-order differencing trend of the current feature over the past 6 time windows. If multiple consecutive windows show the same direction of change (continuous increase or continuous decrease), a trend coefficient k (e.g., the slope of change over the most recent 6 windows) is calculated. The final trend enhancement factor is trend_factor = 1 + α * |k|, where α is a preset coefficient (e.g., 0.5). This factor further enhances the deviation score.

[0076] Step 4: Multi-scale envelope verification. Compare the current value x with the multi-scale fluctuation envelope: if x exceeds the limit of the short-time-scale envelope, add an extra penalty score (e.g., 20 points); if it exceeds the medium-time-scale envelope, add 10 points; if it exceeds the long-time-scale envelope, add 5 points. These penalty scores are accumulated into the base deviation score.

[0077] Step 5: Score Mapping and Limiting. The original deviation index calculated above is converted into a score of 0-100 using an S-shaped mapping function, with an upper limit of 100. The mapping function ensures that the score is lower (<20) when d is less than 1.5 times the standard deviation, and closer to 100 when d exceeds 4 times the standard deviation.

[0078] Step 6: Output. Each feature field ultimately receives an independent deviation score, and all scores are combined into a vector S_single. Furthermore, for discrete features (such as "whether smoke was detected"), the deviation score is directly defined as follows: if the label is "yes" and the confidence level is higher than the threshold, the score is 80 (configurable); otherwise, it is 0.

[0079] Cross-modal spatiotemporal correlation pattern matching does not consider each feature in isolation, but rather treats multiple feature fields in the current multimodal data vector as a whole pattern and matches them with rule patterns in a pre-defined composite risk rule base. These rules represent composite risks indicated when multiple single anomalies occur simultaneously or in a specific temporal sequence, such as the "precursor to charging thermal runaway" and "cross-risk of leakage and ignition source" mentioned earlier.

[0080] The composite risk rule base is a dynamically expandable knowledge base, and each rule has the following structured definition: Unique identifier for the rule.

[0081] Applicable spatial scope: This rule applies to which functional blocks (such as charging-only areas) or grid cell types.

[0082] Time window length: The time span that pattern matching needs to consider.

[0083] Prerequisite: Some rules only take effect under specific global conditions.

[0084] Conditional Pattern: This is a combination of logical clauses, each describing a condition that a feature field should satisfy. Conditions can be numerical ranges, discrete states, temporal relationships, or spatial relationships.

[0085] Rule weight: Indicates the degree to which the rule contributes to the final comprehensive risk score when it is triggered.

[0086] Rule confidence: The statistical validity of the rule obtained from historical verification.

[0087] For the multimodal data vector V_T of the current grid cell, the cross-modal spatiotemporal association rule engine performs the following matching process: Step 1: Rule Filtering. Based on the functional block type of the current grid and global environmental conditions (such as temperature and time), filter out a subset of rules applicable to the current scenario from the rule base. For example, rules for the tank area are not applicable to the charging area grid.

[0088] Step 2: Time Window Data Extraction. For each selected rule, the engine extracts the historical vector sequence (including the current window) of that grid within the most recent L time windows (L is defined by the rule) from the spatiotemporal data cube, starting from its required time window length. This provides information on temporal evolution.

[0089] Step 3: Conditional Pattern Matching. For each conditional clause in the rule, the engine checks whether the extracted time series data satisfies the clause. The matching of conditional clauses is not a strict Boolean judgment, but rather calculates a clause match degree (a continuous value between 0 and 1). For example: For numerical range conditions (such as "temperature slope > 0.6℃ / s"): if the actual value is 0.8, the matching degree is 1; if it is 0.5, the matching degree is 0; if it is between 0.5 and 0.58, the linear interpolation yields 0.2. This fuzzy matching enhances the robustness of the rule.

[0090] Step 4: Calculate the overall rule matching score. The overall matching score of a rule, match_score, is obtained by the weighted geometric mean or weighted arithmetic mean of the matching degrees of all its condition clauses, taking into account whether the rule's preconditions are met (if not, output 0 directly). Additionally, if there are logical AND / OR relationships between the condition clauses in the rule, they are processed according to the corresponding Boolean logic continuous fuzzy operations (e.g., using the minimum value to replace AND, and the maximum value to replace OR). The final score is converted into the final rule matching score (0~100) through a non-linear gain function and multiplied by the rule's confidence factor.

[0091] Step 5: Output. The engine performs the above calculations on all the selected rules and outputs a rule matching score vector S_rule, where each element corresponds to the matching score of a rule.

[0092] After obtaining the single-modal independent deviation score vector S_single and the cross-modal rule matching score vector S_rule, the risk integral calculator is responsible for fusing them into the comprehensive risk integral R of the current grid cell. The calculation process is as follows: Step 1: Feature Deviation Aggregation. All feature deviation scores in S_single are weighted and summed according to preset feature importance weights (e.g., temperature slope has a higher weight than ambient humidity) to obtain the single-modal aggregated score P_single (range 0~100). If a feature is currently in a data missing state, its weight is temporarily set to zero.

[0093] Step 2: Rule matching score aggregation. Take the maximum value of all rule matching scores in S_rule to obtain the composite rule aggregation score P_rule (range 0~100). Weighted summation can also be used, but the maximum value better highlights high-risk rules.

[0094] Step 3: Dynamic Weight Fusion. The basic fusion formula is: R = β * P_rule + (1-β) * P_single. Here, β is an adaptive coefficient with an initial value of 0.6, indicating that under normal circumstances, the weight of the composite rule is slightly higher than that of the single-modal deviation. The system dynamically adjusts β based on recent event statistics: if, in a recent period, the effectiveness of warnings triggered by composite rules has been verified to be significantly higher than that of single-modal warnings, then β is increased; otherwise, it is decreased. This adjustment is performed weekly to ensure that the fusion strategy is always optimal.

[0095] Step 4: Temporal Smoothing and Suppression. To prevent drastic jumps in the risk integral due to instantaneous noise, the calculator performs exponential smoothing on the current window's R and the historical risk integral R_prev from the previous window: R_smoothed = γ * R + (1-γ) * R_prev, where γ is typically 0.7. Simultaneously, if the risk integral of the current grid has consistently been below a low threshold (e.g., 5 points) over the past 10 windows, a downsuppression coefficient (e.g., multiplied by 0.9) is applied to the current integral to reduce unnecessary fluctuations in low-risk regions.

[0096] Step 5: Level Determination. Finally, the smoothed R_smoothed is compared with two preset thresholds: If R_smoothed < 30, it is judged as "normal" level and no alarm is generated.

[0097] If 30 ≤ R_smoothed < 70, it is judged as "attention warning" level.

[0098] If R_smoothed ≥ 70, it is determined to be at the "linked alarm" level.

[0099] The final output of step S3 includes: the comprehensive risk integral and risk level of each spatial grid cell in the current time window, along with information for interpretability: which single-modal features or composite rules mainly contributed to the integral. This information is passed to step S4 for linkage control, and also fed back to the dynamic spatial grid refinement mechanism of step S2 (for adjusting grid resolution) and the rule evolution module of step S3 itself (for model updating).

[0100] S4. Compare the comprehensive risk score with the dynamically adaptive first threshold and second threshold. When the score reaches or exceeds the first threshold but is lower than the second threshold, execute the attention warning response. When the score reaches or exceeds the second threshold, execute the linkage alarm response. Automatically generate an ordered instruction set and send it to the field execution device through the Internet of Things execution gateway.

[0101] Specifically, based on accident inversion and safety expert assessment of a large amount of historical operating data from refueling stations, the system presets two core thresholds: First threshold (early warning threshold): set at 30 points. When the comprehensive risk score reaches or exceeds 30 points but is below 70 points, it indicates that the grid unit has exhibited suspicious patterns or precursory characteristics that deviate from normal conditions, but has not yet posed a direct and imminent threat to personnel, equipment, or the environment. This level of triggering should attract the attention of operations management personnel, and non-emergency intervention measures such as inspections, patrols, or remote verification should be recommended.

[0102] The second threshold (alarm threshold) is set at 70 points. When the comprehensive risk score reaches or exceeds 70 points, it indicates that the risk pattern of the grid cell has highly matched a known complex hazardous scenario, or the degree of single-modal anomaly has approached the equipment limit or safety specification red line, and there is a high probability that it will evolve into a serious accident such as fire, explosion, or personal injury. This level requires the system to immediately and automatically execute a series of active protection and personnel evacuation commands without manual confirmation.

[0103] To adapt to different equipment types, operational intensity, environmental characteristics, and risk preferences at various refueling stations, the system incorporates a threshold adaptive adjustment module. This module dynamically optimizes the threshold based on the following principles: False alarm to false negative balance: The system compiles weekly statistics on alerts and alarms from the past seven days, calculating the proportion of "false alarm events" (events the system identifies as alerts or linked alarms, but which are confirmed by manual review to pose no actual risk). If the false alarm rate exceeds a preset upper limit (e.g., 5%) for two consecutive weeks, the first threshold is automatically adjusted upwards to reduce sensitivity to normal fluctuations. Conversely, if a "false alarm event" occurs (an actual accident or near-miss occurs, but the risk score within one minute prior to the accident never reached 30 points), the system immediately adjusts the first threshold downwards (e.g., reduces it by 3 points) and notifies safety management personnel for review.

[0104] Time-based and operating-condition adaptive: The system automatically switches between different threshold levels based on the busy time of the refueling station and the external environment.

[0105] The results of threshold adjustments are recorded in the system configuration library in real time, and each adjustment is accompanied by a reason for adjustment, which is available for management personnel to approve or roll back, ensuring that humans have ultimate control.

[0106] Based on the comparison results of the comprehensive risk score with the first threshold and the second threshold, the system will trigger three different levels of response strategies: when the score is below the first threshold, no external output will be generated (only data will be recorded); when the score is between the first threshold and the second threshold, an "attention warning" response will be executed; when the score reaches or exceeds the second threshold, an "association alarm" response will be executed.

[0107] When the overall risk score is ≥ 30 and < 70, the system determines it to be at the "Attention Warning" level. At this point, the risk has not yet reached the urgency level requiring automatic power cutoff or personnel evacuation, but operations personnel need to intervene as soon as possible. Response actions include: Generate a structured warning message: The policy instruction set generator compiles a warning message for this grid cell, which includes: trigger time, grid code, comprehensive risk score, and the main contributors that cause the score to increase.

[0108] Push notifications to operations personnel's smart terminals: Warning messages are sent via mobile push service (integrated into the digital operations management platform) to the mobile apps or smartwatches of the following three types of personnel: The on-duty safety officer of the functional area where the grid is located (automatically identified through the scheduling system); Station control room duty supervisor; External maintenance engineers for the relevant equipment (only for equipment-related early warnings).

[0109] The push notification method supports sound alerts and vibration, but by default, it does not use methods such as sirens or broadcasts that may cause panic on site.

[0110] Remote verification assistance: The system automatically retrieves real-time video stream clips associated with the grid cell (15 seconds before and after the alert) and key sensor curves from the most recent 5 minutes, pushing them to the recipient in the form of a visual card. The recipient can directly "confirm receipt," "mark as false alarm," or "request escalation to a linked alarm" on their mobile terminal. If no one confirms or handles the alert within 15 minutes, the system automatically escalates the alert to a linked alarm and notifies higher-level management personnel.

[0111] Enhanced local data acquisition: To facilitate subsequent analysis, the system instructs the edge computing gateway to which the grid cell belongs to temporarily increase the data upload frequency of the sensing terminals related to the grid to the highest level (e.g., from 1Hz to 10Hz), and at the same time requires the gateway to temporarily store raw waveform data for a longer period of time for later analysis.

[0112] When the comprehensive risk score is ≥ 70, the system determines it to be at the "linked alarm" level. At this point, the risk is highly urgent, requiring the system to automatically execute a series of proactive safety control commands without waiting for manual confirmation. The response action is a set of ordered, multi-device coordinated instructions, the specific content of which depends on the grid type that triggered the alarm and the dominant risk mode.

[0113] The policy instruction set generator internally maintains a "risk pattern - instruction template" mapping table. For example: For the charging zone thermal runaway precursor risk mode (mainly triggered by rule R001), the instruction template includes: Disconnect the AC input power to the charging station and its adjacent charging stations (executed via a smart circuit breaker). Turn on the explosion-proof exhaust fan (if any) in the charging station area to maximum speed; Vibration and voice prompts are sent to all personnel wearing smart badges within a 5-meter radius of the grid via the UWB positioning system: "Please leave the area near charging station XX immediately." Automatically turn on the emergency lighting in the area and turn off the regular lighting in the area; Real-time video streams, sensor data, and alarm summaries are locked and pushed to the station control room's large screen, the supervisor's mobile terminal, and the regional fire control center.

[0114] When generating a specific instruction set, the dependencies and timing requirements between instructions are dynamically optimized: Parallel Execution Group: Mark instructions without resource conflicts (such as cutting off power, turning on ventilation, and pushing notifications) as "execute in parallel immediately" to respond as quickly as possible.

[0115] Sequential execution groups: Instructions with dependencies are ordered. For example, power must be cut off before emergency lighting can be turned on (to prevent electrical sparks from switching lighting under load). The module sets a minimum time interval (e.g., 50 milliseconds) for each group of instructions and issues them sequentially.

[0116] Conflict detection and mutual exclusion handling: If two adjacent grids in the same power station trigger linked alarms simultaneously, but one command requires shutting down the exhaust (e.g., for a specific risk) while the other requires opening the exhaust, the conflict detection module will prioritize the grid with the higher risk level to determine the exhaust status, or adopt a "safety first" principle (opening the exhaust is usually safer). The conflict event is also recorded for post-event analysis.

[0117] The ordered instruction set output by the policy instruction set generator is sent to the IoT execution gateway. This gateway is deployed locally at the station level and connects to each execution device via redundant fieldbuses (such as dual-channel Modbus TCP or MQTT over 5G). To ensure the reliability of the instructions under extreme conditions, the gateway employs the following mechanism: Command Confirmation and Retry: After each control command is issued, the gateway waits for the "Action Completion Readback" signal from the executing device. If no readback is received within 200 milliseconds, it will automatically retry twice, with an interval of 100 milliseconds between each attempt. If it still fails, the device will be marked as a "communication failure," and this information will be immediately fed back to the station control screen. At the same time, an attempt will be made to resend the command through a backup communication link (such as another bus).

[0118] Atomicity of the instruction execution chain: For a set of sequentially executed instructions, if one of the instructions fails, the gateway will not continue to execute subsequent instructions that depend on that device. Instead, it will immediately issue an emergency alarm for "partial linkage failure" to the administrator, providing the ID of the failed device and the cause of the failure. At the same time, the system will attempt to execute a preset "degradation linkage scheme" (such as manually disconnecting the upstream main switch) as a backup.

[0119] Local persistence and black box: All issued commands, readback status, retry records, and timestamps are stored by the gateway in local solid-state storage (with a capacity of at least 7 days' worth of records), forming an immutable execution black box. Even if the network connection with the cloud is interrupted, the local linkage control process still executes completely, and the logs are automatically synchronized to the upper layer after the network is restored.

[0120] After a linked alarm is triggered, the system does not completely replace human decision-making, but rather achieves a collaborative process of "automatic execution + manual confirmation." Specifically: Forced pop-up window on the station control screen: When an alarm is triggered, the station control room screen automatically switches to a 3D digital twin view of the high-risk grid and marks the risk area with a flashing red border. The pop-up window displays: alarm time, grid location, risk mode type, comprehensive risk score, list of automatically executed instructions and their execution status.

[0121] Pre-set response suggestion card: The workbench automatically generates a standardized response suggestion card for this risk mode, including: required personal protective equipment, key equipment to be inspected, emergency contact number, and the location of the nearest safe assembly point. On-duty personnel can click "One-click Call" to directly contact all personnel wearing smart work badges on site.

[0122] Handling Result Recording: After the on-site handling is completed, the on-duty personnel upload the handling result to the workbench, and can attach on-site photos and voice notes. The system automatically packages the entire chain of data of the event (sensing data, model judgment basis, automatic execution actions, and manual feedback) into a "security event instance" and stores it in the historical database for subsequent model optimization.

[0123] Forced Lockout: For certain high-risk alarm linkages (such as power cutoff to fuel dispensers), the system will require management personnel to enter a release command and verify their identity (such as by scanning a code or a secondary password confirmation) after handling the situation before power can be restored. This avoids blindly restoring power before the risk has been completely eliminated.

[0124] S5. Package the complete data of each security incident into a structured security incident instance and store it in the historical database. Regularly analyze and generate parameter adjustment suggestions for the composite risk association identification model, and provide feedback and updates after confirmation.

[0125] Specifically, after each "linked alarm" or important "attention alert" marked by the user is handled, the on-duty personnel enter the actual on-site handling process, fault cause analysis, and final results through the event handling closed-loop workbench of the digital operation management presentation layer. The system automatically packages the complete data of this event—from the original sensed signal, model judgment basis, automatic execution actions, to manual feedback conclusions—into a structured "security event instance." This instance is stored in the historical database to form a traceable digital archive; on the other hand, it is periodically analyzed by the security policy optimizer to identify biases in the model judgment (such as missed alarms and false alarms) and deficiencies in the linkage strategy, generating adjustment suggestions for thresholds, weights, or linkage instruction sets in the composite risk rules. After confirmation by management personnel, these suggestions are fed back to the multimodal risk identification and analysis layer to update the model parameters.

[0126] The second embodiment of this application is as follows: Please see Figure 3 This invention provides an IoT-based digital operation system for energy stations, applied to an IoT-based digital operation method for energy stations as provided in the first embodiment, comprising: The IoT sensing layer includes environmental safety sensing terminal groups, device status sensing terminal groups, personnel and behavior sensing terminal groups, and edge computing gateways. It is used to collect multi-source heterogeneous data with a dynamic priority sampling and scheduling mechanism and perform hierarchical lightweight feature extraction at the edge to output key indicator vectors. The digital twin and data fusion layer is deployed on a local server at the station level or in the cloud. It includes a device digital twin model library, a spatiotemporal data alignment engine, and a dynamic spatial grid refinement module, which is used to map key indicator vectors to a three-dimensional spatial grid model and assemble them into a spatiotemporal data cube. The multimodal risk identification and analysis layer includes a single-modal baseline evaluator, a cross-modal spatiotemporal association rule engine, and a risk score calculator, which are used to calculate the comprehensive risk score for the multimodal data vector of each spatial grid cell; The security policy execution and linkage control layer includes a policy instruction set generator and an IoT execution gateway, which are used to generate and issue ordered instruction sets to field execution devices based on the comparison results of the comprehensive risk score and the first threshold and the second threshold. The digital operations management presentation layer includes a holographic security posture dashboard, an incident handling closed-loop workbench, and a security policy optimizer, which are used to realize the visualization of security posture, closed-loop recording of incidents, and model self-optimization.

[0127] In this embodiment, the IoT sensing layer is deployed at the energy station site to collect multi-dimensional raw safety data. This layer specifically includes: Environmental safety sensing terminal group: including but not limited to combustible gas concentration sensors, flame and smoke detectors, ambient temperature and humidity sensors, and trace leak detectors based on laser spectroscopy technology deployed in refueling areas, charging areas, storage tank areas, and power distribution rooms.

[0128] Equipment status sensing terminal group: integrated or installed on core equipment such as charging piles, fuel dispensers, hydrogen dispensers, oil storage tanks, and hydrogen storage tanks, used to collect real-time operating parameters, such as the temperature of the internal power module of the charging pile, the output current and voltage harmonics, oil leakage monitoring of the fuel nozzle hose, tank pressure and liquid level, and micro-leakage acoustic signals of the hydrogen nozzle hose.

[0129] Personnel and behavior sensing terminal group: including high-definition smart cameras (embedded with behavior recognition algorithms) deployed in key areas, smart name tags or wristbands worn by on-site staff (with UWB high-precision positioning, fall detection, and emergency call functions), and perimeter radar and video fusion devices for identifying external personnel entering the danger zone.

[0130] Edge computing gateway: At least one edge computing gateway is set up in each refueling station area (such as charging area and refueling area). The gateway connects to all the above-mentioned sensing terminals through wired or wireless means, receives raw data and performs preliminary cleaning, alignment and abnormal feature extraction based on local lightweight model. Only key features, alarms and compressed waveform data are uploaded to the cloud or station-level central server to reduce network load and cloud computing pressure.

[0131] Digital Twin and Data Fusion Layer: Deployed on a station-level local server or in the cloud, this layer is used to build a unified digital image of the power station. This layer includes: Equipment Digital Twin Model Library: High-fidelity digital twin models are pre-built for each type of equipment (e.g., specific models of charging piles and fuel dispensers). These models not only include the parameter ranges for normal operation but, more importantly, the characteristic parameter variation curves under typical fault modes (e.g., aging DC contactors can cause a slower current rise time during the charging pile's startup phase). Real-time parameter streams collected on-site are continuously compared with their corresponding digital twin models.

[0132] Spatiotemporal data alignment engine: Aligns data from different sensing terminals and different sampling frequencies (e.g., 1Hz for gas sensors and 25Hz for cameras) according to a unified spatial grid encoding and timestamp for the charging station, forming a multimodal spatiotemporal data cube. The dataset of each spatial grid cell (e.g., dividing a charging station parking space into an independent grid) contains environmental parameters, equipment parameters, and personnel behavior parameters for that location during the same period.

[0133] Multimodal Risk Identification and Analysis Layer: This layer receives a spatiotemporal data cube aligned by the data fusion layer and analyzes it using a pre-trained "composite risk association identification model." The structure and operational logic of this model are as follows: Single-modal risk baseline: The model first independently evaluates each type of single-source data (such as the temperature curve of a charging pile or the concentration of combustible gas in a certain area) to determine whether it exceeds the traditional static threshold or dynamic baseline (generated in real time by a time series prediction model).

[0134] Cross-modal spatiotemporal correlation rule engine: When a single-modal assessment fails to reach the independent alarm threshold, or when multiple single-modal indicators are within normal boundaries but exhibit a specific spatiotemporal correlation pattern, this engine triggers deep inference. The engine has several preset composite risk rules, such as: Rule 1 (Precursor to Charging Thermal Runaway): The rule pattern is "The rate of temperature rise inside the charging pile exceeds 1.5 times the normal charging curve for 3 consecutive minutes" + "No personnel activity within a 2-meter grid around the charging pile during the same period" + "Abnormal decrease in the current of the charging pile's cooling fan". This pattern indicates a risk level where there is an internal equipment fault but it has not yet posed a direct threat to the external environment.

[0135] Rule Two (Risk of Intersecting Leakage and Ignition Sources): The rule pattern is "A sustained but slight increase in the concentration of combustible gas in the underground trench area of ​​a certain refueling island (less than 20% of the lower explosive limit)" + "A camera within 5 meters above the grid detected an open vehicle hood (potential ignition source)" + "Smart ground stakes detected continuous micro-vibration nearby (suspected flow of leaked fluid)". This pattern points to a high-risk compound risk of a concealed leak encountering a potential ignition source.

[0136] Risk Score Calculator: For each spatiotemporal grid, a comprehensive risk score is dynamically calculated based on the degree of deviation of the single mode and the weight of the matched composite rules. When the score exceeds the first threshold, a "Pay Attention Warning" is generated; when it exceeds the second threshold, a "Linked Alarm" is generated.

[0137] Model self-learning module: Every real security incident (regardless of whether it causes loss) and its entire process of perception data and manual handling records will be fed back to the model library as positive and negative samples to update the dynamic baseline and adjust the weight coefficients of the composite rules, so as to realize the continuous evolution of the model.

[0138] Security Policy Execution and Linkage Control Layer: Based on the output of the risk identification layer, this layer automatically generates and executes security policies. This layer includes: Strategy Instruction Set Generator: Automatically compiles and generates a set of ordered IoT control instructions for different types of risk scores and alarm levels. For example, for a warning of "precursor signs of charging thermal runaway," the instruction set might be: ① Push a maintenance reminder with the device ID and temperature trend graph to the maintenance personnel's mobile APP; ② Increase the real-time data sampling frequency of the charging pile. For a linked alarm of "cross-risk of leakage and ignition source," the instruction set might be: ① Immediately cut off the power to the submersible pump and fuel dispenser of the affected refueling island; ② Start the explosion-proof ventilation system in the area to maximum power; ③ Broadcast the danger zone clearance instruction to the smart name tags worn by all personnel within 5 meters via the UWB positioning system; ④ Automatically turn on the emergency lighting and evacuation indicator lights in the area; ⑤ Send the associated real-time video stream and alarm summary to the station control room screen and the duty supervisor's mobile terminal.

[0139] IoT execution gateway: Receives instruction sets and uses communication protocols compatible with each subsystem (such as Modbus, OPCUA, MQTT, etc.) to send control instructions to corresponding edge execution devices, such as smart circuit breakers, electric valves, audible and visual alarms, broadcasting systems, etc., to ensure reliable execution of instructions and status feedback.

[0140] Digital Operations Management Presentation Layer: Provides a human-computer interaction interface to achieve visualization and closed-loop management of security posture. This layer includes: Holographic Safety Situation Dashboard: Using the 3D digital map of the power station as a base, it dynamically overlays environmental risk cloud maps, equipment health status heat maps, real-time personnel locations, and electronic fence information. Changes in the risk score of any area are visually presented through color gradients and dynamic halos.

[0141] The incident handling closed-loop workbench displays a list of all "Attention Warnings" and "Linked Alarms" events. Each event is associated with: 30-second multimodal data segments (sensor curves, video recordings) before and after the trigger time, the set of automatically executed linkage instructions and their execution results, and a pre-set standardized handling suggestion form. After completing on-site handling, on-duty personnel can upload the handling results, on-site photos, and fill in a debriefing conclusion on the workbench, forming a fully digital closed loop.

[0142] Security Policy Optimizer: Regularly scans the historical event database, analyzes the triggering reasons, system response time, and manual handling efficiency of each "linked alarm", and automatically generates optimization suggestions for threshold parameters and linkage instruction set sequence in existing composite risk rules. After confirmation by management personnel, it can be updated to the multimodal risk identification and analysis layer with one click.

[0143] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0144] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0145] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A digital operation method for energy stations based on the Internet of Things, characterized in that, Includes the following steps: The IoT sensing layer at the power station site collects environmental safety parameters, equipment status parameters, and personnel behavior data at a frequency set by a dynamic priority sampling and scheduling mechanism. The edge computing gateway then performs data cleaning, time alignment, and layered lightweight feature extraction to generate key indicator vectors. The key indicator vectors are mapped to the corresponding spatial grid units in the three-dimensional spatial grid model of the power station according to the dynamic spatial grid refinement mechanism. Then, the multimodal data vectors of each spatial grid unit in the current time window are assembled by weighted interpolation, thereby constructing a spatiotemporal data cube. Using a composite risk association identification model, single-modal baseline assessment and cross-modal spatiotemporal association pattern matching are performed in parallel on the multimodal data vectors of each spatial grid cell to obtain independent deviation score vectors and rule matching score vectors, which are then fused to obtain the comprehensive risk score of the spatial grid cell. The comprehensive risk score is compared with the dynamically adaptive first threshold and second threshold. When the score reaches or exceeds the first threshold but is lower than the second threshold, a warning response is executed. When the score reaches or exceeds the second threshold, a linkage alarm response is executed. An ordered instruction set is automatically generated and sent to the field execution equipment through the Internet of Things execution gateway. Complete data from each security incident is packaged into a structured security incident instance and stored in a historical database. The database is then periodically analyzed to generate parameter adjustment suggestions for the composite risk association identification model, which are then updated after confirmation.

2. The IoT-based digital operation method for energy stations as described in claim 1, characterized in that, The dynamic priority sampling scheduling mechanism specifically includes: The dynamic sampling schedule table for each sensing terminal is obtained, and the current sampling rate is dynamically adjusted according to the attention score of the terminal data in the previous scheduling cycle. The attention score is calculated in real time based on the degree to which the instantaneous value of the terminal exceeds the historical dynamic envelope, the abnormal state of adjacent terminals, and the comprehensive risk score of the spatial grid unit to which it belongs in the previous cycle. When the attention score exceeds the score threshold, the sampling frequency is increased; otherwise, it is gradually reduced to the basic sampling frequency.

3. The method for digital operation of energy stations based on the Internet of Things as described in claim 1, characterized in that, Hierarchical lightweight feature extraction includes: The first sublayer extracts temporal feature primitives from continuous analog quantities, including instantaneous statistical features, dynamic trend features, and spectral energy features. The second sublayer generates structured event features from non-video discrete events and image / video data, including electronic fence events calculated from UWB positioning data and semantic labels for helmet wearing status, mobile phone holding actions, and flame / smoke areas detected from video frames by embedded neural network models. The third sublayer generates cross-channel correlation features between different perception channels managed by the same edge computing gateway. When the feature primitives of multiple channels show a positive correlation exceeding a preset threshold, correlation feature labels are generated.

4. The method for digital operation of energy stations based on the Internet of Things as described in claim 1, characterized in that, The dynamic spatial grid refinement mechanism is as follows: The system obtains the preset base resolution of the basic three-dimensional spatial mesh model and automatically subdivides or merges the mesh locally based on the real-time risk distribution and the density of sensing terminals. When the comprehensive risk score of a certain mesh cell exceeds 80% of the first warning threshold or the number of sensing terminals deployed in the mesh exceeds the preset density limit, the mesh is divided into multiple sub-mesh in the horizontal direction. When the long-term comprehensive risk score of a grid is close to zero and no terminal is deployed, it is merged with adjacent grids of the same type; grid changes are updated in real time to the spatial grid index table and reassessed every five minutes.

5. The method for digital operation of energy stations based on the Internet of Things as described in claim 1, characterized in that, Weighted interpolation assembly includes: For each spatial grid cell, according to the predefined multimodal data vector template, the values ​​of each asynchronously arriving feature field are filled with the current time window by the most recent valid value plus time decay weight, linear interpolation, or the cumulative duration of the event. When the same feature field has multiple independent sources, a weighted average fusion is performed based on the historical accuracy of the sensor and the data age, and a consistency check is performed to remove outliers.

6. The method for digital operation of energy stations based on the Internet of Things as described in claim 1, characterized in that, The single-modal baseline assessment uses a dynamic baseline model, which maintains a time-series prediction baseline, multi-scale fluctuation envelope, and mode state memory for each continuous feature field. For the current value of the input feature, the current operating mode is first matched to select the corresponding prediction mean and prediction standard deviation, the standardized residual is calculated, then a trend enhancement factor is introduced and multi-scale envelope verification is performed, and finally it is converted into an independent deviation score of 0 to 100 through a sigmoid mapping function.

7. The IoT-based digital operation method for energy stations as described in claim 1, characterized in that, Cross-modal spatiotemporal correlation pattern matching is performed based on a composite risk rule base. Each rule includes the applicable spatial range, time window length, preconditions, condition pattern, and rule weight. The matching process includes: selecting an applicable subset of rules based on the current grid functional blocks and environmental conditions, extracting the historical vector sequence of the corresponding time window, calculating the continuous clause matching degree for each condition clause, and then obtaining the overall rule matching score through weighted average or fuzzy logic operation.

8. A digital operation system for energy refueling stations based on the Internet of Things (IoT), applied to the digital operation method for energy refueling stations based on the IoT as described in claim 1, characterized in that, include: The IoT sensing layer includes environmental safety sensing terminal groups, device status sensing terminal groups, personnel and behavior sensing terminal groups, and edge computing gateways. It is used to collect multi-source heterogeneous data with a dynamic priority sampling and scheduling mechanism and perform hierarchical lightweight feature extraction at the edge to output key indicator vectors. The digital twin and data fusion layer is deployed on a station-level local server or in the cloud. It includes a device digital twin model library, a spatiotemporal data alignment engine, and a dynamic spatial grid refinement module, which is used to map key indicator vectors to a three-dimensional spatial grid model and assemble them into a spatiotemporal data cube. The multimodal risk identification and analysis layer includes a single-modal baseline evaluator, a cross-modal spatiotemporal association rule engine, and a risk score calculator, which are used to calculate the comprehensive risk score for the multimodal data vector of each spatial grid cell; The security policy execution and linkage control layer includes a policy instruction set generator and an IoT execution gateway, which are used to generate and issue ordered instruction sets to field execution devices based on the comparison results of the comprehensive risk score and the first threshold and the second threshold. The digital operations management presentation layer includes a holographic security posture dashboard, an incident handling closed-loop workbench, and a security policy optimizer, which are used to realize the visualization of security posture, closed-loop recording of incidents, and model self-optimization.

9. The IoT-based digital operation system for energy stations as described in claim 8, characterized in that, The policy instruction set generator internally maintains a risk mode-instruction template mapping table and integrates an instruction timing arrangement and conflict detection module. This module divides instructions into parallel execution groups and sequential execution groups according to the dependencies between instructions, and performs mutual exclusion processing on instructions with resource conflicts according to the safety priority principle. The IoT execution gateway supports instruction confirmation and retry mechanisms, atomic control of instruction execution chains, and local persistent black box records to ensure that linkage control can still be completed when the network is interrupted.

10. The IoT-based digital operation system for energy stations as described in claim 8, characterized in that, For each linked alarm event, the closed-loop event handling workbench of the digital operation management presentation layer automatically generates a standardized handling suggestion card and displays the list of automatically executed instructions and their status in a pop-up window on the station control screen. After the on-site handling is completed, the on-duty personnel can upload the handling results, on-site photos, and debriefing conclusions on the workbench. The system automatically packages the entire chain data of the event into a security event instance and stores it in the historical database. The security policy optimizer periodically analyzes historical event instances and generates adjustment suggestions for the first threshold, the second threshold, and the rule weights and dynamic baseline parameters in the composite risk association identification model. After confirmation by the management personnel, the suggestions are fed back to the corresponding level.

Images