Intelligent fireproof electric control cabinet operation method

Abstract

The application discloses an intelligent fireproof electrical control cabinet operation method, through multi-type sensor distributed collection, a multi-source data set of strong binding of physical position and network topology label is established, accurate normalization, noise suppression and unified feature calibration of original data are realized, further, an abnormal resonance identification algorithm and a cause-effect inference are adopted, an equipment abnormal motion correlation mode is extracted, a dynamic equipment trust graph is constructed based on a Bayesian update and an attention mechanism, a graph neural network is used to intelligently predict a risk diffusion path and a high-risk node distribution. The system automatically triggers and issues a hierarchical alarm and a protection plan according to a risk cascade threshold, realizes active prevention and control such as physical isolation and load degradation, and periodically corrects risk relations and model parameters with the aid of an execution feedback mechanism, the application has high-dimensional space-time monitoring capability, can realize high-precision, whole-process dynamic optimization and self-adaptive evolution of power distribution network risk identification, and effectively improves group intelligent protection level and system stability.

Description

Technical Field

[0001] This invention relates to the field of equipment risk analysis and early warning technology, and in particular to an intelligent fireproof electrical control cabinet operation method. Background Technology

[0002] Currently, the analysis and early warning technology for group risks of equipment in intelligent power distribution systems has become an important development direction for ensuring the security and intelligent protection of large-scale power distribution networks. As the complexity of power distribution networks in scenarios such as industrial parks, commercial complexes, and intelligent buildings continues to increase, the detection of local hidden dangers in a single electrical control cabinet or independent equipment is no longer sufficient to meet the needs of collaborative safety management of multiple devices.

[0003] In recent years, the industry has widely adopted various types of sensors (such as current, voltage, temperature, vibration, and communication status) for real-time parameter acquisition in the operation monitoring and equipment hazard diagnosis of intelligent fire-resistant electrical control cabinets. Algorithms such as expert rules, statistical analysis, or machine learning are used to achieve anomaly detection and early warning for individual devices. In some solutions, to improve system interoperability, simple device correlation modeling has also been introduced, such as using partitioned risk aggregation based on topology or principal component analysis, or implementing directional coupling anomaly detection for known physically adjacent devices. This aims to achieve, to some extent, parallel processing of risk information from multiple devices and the discovery of local correlations.

[0004] However, current mainstream technologies generally focus on the following areas: First, they are based on the mining and early warning of abnormal features of single or local devices, lacking systematic and real-time modeling and tracking of the dynamic relationships between multiple devices in large-scale networks. Second, static mapping based on physical topology or rule knowledge is insufficient to reflect the actual need for device-level risk relationships to adaptively adjust with load fluctuations, environmental changes, and abnormal evolution processes. Third, group events such as risk diffusion and cascading evolution often rely on phased manual inspections or delayed event tracing, resulting in delayed responses and coarse perception when facing complex scenarios such as group failures and synchronous anomalies. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, the present invention provides an intelligent fireproof electrical control cabinet operation method.

[0006] The technical solution of this invention is implemented as follows: a method for operating an intelligent fireproof electrical control cabinet, comprising:

[0007] S1: Collect electrical parameters, thermal parameters, vibration parameters, communication signals and historical fault information from various types of sensors deployed in electrical control cabinets within complex power distribution networks. Record the physical location, cabinet structure and network topology label for each sensor sampling point to obtain multi-source and multi-dimensional monitoring raw data.

[0008] S2: Normalize the original data of the multi-source and multi-dimensional monitoring, and perform preprocessing operations such as time domain alignment, noise reduction, and unit standardization based on the physical location, cabinet structure and different environmental conditions of each sensor to generate a uniformly calibrated feature dataset.

[0009] S3: Input the uniformly calibrated feature dataset into the abnormal resonance identification algorithm module, and use time series cross-correlation analysis, causal inference method and segmented clustering algorithm to extract the abnormal resonance feature patterns between each cabinet and its associated equipment, and obtain the initial set of equipment clusters with synchronous evolution and cascade evolution.

[0010] S4: Based on the abnormal resonance feature pattern, combined with the power distribution network topology and the physical attributes of each device, the risk transmission weights and directions between all device nodes are calculated through Bayesian update and attention mechanisms to generate an adaptive device trust graph containing risk relationship edge weights and node labels.

[0011] S5: For the adaptive device trust map, learn and reason about the historical and current feature data within a specified sliding time window based on graph neural network or graph attention network, and dynamically predict the risk diffusion path, impact range and high-risk node clustering area among the electrical control cabinet equipment groups;

[0012] S6: Determine whether the risk weight, diffusion speed and spatial distribution of high-risk paths in the graph reasoning results exceed the set threshold. If the group risk cascading or evolution triggering conditions are met, determine and output the graded alarm signal and corresponding protection plan parameters based on the trust relationship link level.

[0013] S7: Send the graded alarm signals and protection plan parameters to the electrical control system. Based on the physical structure, communication capabilities and network distribution of different cabinets, execute control actions such as partial circuit breaking, network-wide load reduction, designated equipment isolation and priority monitoring, and record the execution feedback after the control action.

[0014] S8: Based on the executed control actions and the feature data collected subsequently, the model is periodically corrected for the equipment trust map and abnormal resonance identification parameters, dynamically correcting the risk relationship between equipment nodes, optimizing the risk capture sensitivity and prediction accuracy for the next cycle, and realizing adaptive evolution of the model.

[0015] The intelligent fireproof electrical control cabinet operation method provided in this application has the following beneficial effects:

[0016] (1) This invention adopts a multi-level distributed data acquisition and normalization process, supporting real-time acquisition of synchronous, lossless, and multi-type raw parameters from up to thousands of sensors. Combined with multi-level data standardization and spatiotemporal alignment processing, it significantly improves data consistency and model generalization capabilities for subsequent risk modeling. Compared with traditional step-by-step acquisition and single physical signal monitoring, this solution significantly improves the accuracy of concurrent acquisition across the entire network and historical data mapping, laying a solid foundation for group hazard analysis in high real-time and large-scale equipment scenarios.

[0017] (2) By combining time series cross-correlation analysis, causal inference and segmented clustering algorithms, it is possible to accurately and flexibly explore the synchronous evolution and cascade driving links between devices, breaking through the limitations of existing technologies in focusing on static or single device risks, and effectively solving the problem of difficulty in identifying dynamic risk transmission paths between devices across space, time and type.

[0018] (3) This invention constructs an adaptive device trust graph, which uses Bayesian updates and attention mechanisms to realize real-time conditional probability correction of risk transmission weights and directions, and combines device attributes to complete fine-tuning of risk relationships. Compared with traditional static modeling based on empirical thresholds or fixed topologies, the dynamic trust graph can automatically optimize with the evolution of actual monitoring data and abnormal events, reflecting the latest risk linkage relationships of the current network structure and device status. Attached Figure Description

[0019] Figure 1 This is a flowchart of an intelligent fireproof electrical control cabinet operation method according to the present invention;

[0020] Figure 2 This is a sub-flowchart of an intelligent fireproof electrical control cabinet operation method according to the present invention;

[0021] Figure 3 This is another sub-flowchart of the operation method of an intelligent fireproof electrical control cabinet according to the present invention. Detailed Implementation

[0022] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0023] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having,” etc., specify the presence of the stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0025] Please see Figures 1-3 As shown, an operating method for an intelligent fireproof electrical control cabinet includes:

[0026] S1: Collect electrical parameters, thermal parameters, vibration parameters, communication signals and historical fault information from various types of sensors deployed in electrical control cabinets within complex power distribution networks. Record the physical location, cabinet structure and network topology label for each sensor sampling point to obtain multi-source and multi-dimensional monitoring raw data.

[0027] S2: Normalize the original data of the multi-source and multi-dimensional monitoring, and perform preprocessing operations such as time domain alignment, noise reduction, and unit standardization based on the physical location, cabinet structure and different environmental conditions of each sensor to generate a uniformly calibrated feature dataset.

[0028] S3: Input the uniformly calibrated feature dataset into the abnormal resonance identification algorithm module, and use time series cross-correlation analysis, causal inference method and segmented clustering algorithm to extract the abnormal resonance feature patterns between each cabinet and its associated equipment, and obtain the initial set of equipment clusters with synchronous evolution and cascade evolution.

[0029] S4: Based on the abnormal resonance feature pattern, combined with the power distribution network topology and the physical attributes of each device, the risk transmission weights and directions between all device nodes are calculated through Bayesian update and attention mechanisms to generate an adaptive device trust graph containing risk relationship edge weights and node labels.

[0030] S5: For the adaptive device trust map, learn and reason about the historical and current feature data within a specified sliding time window based on graph neural network or graph attention network, and dynamically predict the risk diffusion path, impact range and high-risk node clustering area among the electrical control cabinet equipment groups;

[0031] S6: Determine whether the risk weight, diffusion speed and spatial distribution of high-risk paths in the graph reasoning results exceed the set threshold. If the group risk cascading or evolution triggering conditions are met, determine and output the graded alarm signal and corresponding protection plan parameters based on the trust relationship link level.

[0032] S7: Send the graded alarm signals and protection plan parameters to the electrical control system. Based on the physical structure, communication capabilities and network distribution of different cabinets, execute control actions such as partial circuit breaking, network-wide load reduction, designated equipment isolation and priority monitoring, and record the execution feedback after the control action.

[0033] S8: Based on the executed control actions and the feature data collected subsequently, the model is periodically corrected for the equipment trust map and abnormal resonance identification parameters, dynamically correcting the risk relationship between equipment nodes, optimizing the risk capture sensitivity and prediction accuracy for the next cycle, and realizing adaptive evolution of the model.

[0034] Step S1: Collect electrical parameters, thermal parameters, vibration parameters, communication signals, and historical fault information output by various types of sensors deployed in each electrical control cabinet within the complex power distribution network. Record the physical location, cabinet structure, and network topology label for each sensor sampling point to obtain multi-source, multi-dimensional monitoring raw data, specifically including:

[0035] S1.1: Identify the equipment nodes of all electrical control cabinets in the preset power distribution network topology, and generate a list of electrical control cabinet numbers to be collected based on the structured network topology model to construct a multi-device distributed collection object list;

[0036] The data collection targets are all the pre-set electrical control cabinets in the complex power distribution network. The initial configuration data of the power distribution facilities is imported through the structured network topology model, including the unique physical identifier of each control cabinet, geographical location parameters, the bus segment number to which it is connected, and the relationship between upstream and downstream nodes.

[0037] A node identification algorithm (parameters: topology edge table, device attribute table) is adopted to traverse and resolve the identity of each cabinet node in the entire network based on the network topology relationship and physical connection attributes, so as to achieve unique identification of each electrical control cabinet node;

[0038] Furthermore, by using a structured data matching method (parameters: device model, interface type, regional grouping information), based on the comparison results between the topology and the actual project construction blueprint, virtual nodes that have no effective monitoring needs or are in maintenance / disuse status are filtered out, and only effective cabinet nodes that need to be included in real-time data collection are retained.

[0039] Furthermore, through a numbering generation algorithm (parameters: location numbering rules, hierarchical increment rules, redundancy check bits), a unique acquisition number is assigned to each validly identified electrical control cabinet, and an information list containing multiple fields such as node number, physical location, and cabinet type is automatically generated;

[0040] Furthermore, by applying a multi-device distributed collection object list construction strategy, the above-mentioned numbered list is grouped according to spatial distribution, partition load capacity, and network topology association to generate a multi-device collection object distribution table suitable for subsequent parallel collection scheduling, thereby achieving the connection between node-level indexing and batch management requirements;

[0041] Through the above-mentioned algorithms for device node identification, effective screening, number generation, and object list construction, all target electrical control cabinets in the preset power distribution network topology are standardized and structured into multi-device distributed acquisition objects, realizing the foundation for unified preparation and batch scheduling before the acquisition of multi-source raw monitoring data;

[0042] For example, in a three-layer distributed power distribution network consisting of 104 electrical control cabinets in an industrial park, the wire connections, busbar segment numbers, and cabinet physical locations are imported through a structured topology table. After node identification and filtering, 96 valid cabinets (excluding maintenance or backup nodes) are retained for monitoring. A numbering algorithm is used to assign numbers according to the format "area number-floor number-cabinet number-redundancy bit", such as A03-F2-C15-07, forming a 96-bit object list. Objects are automatically grouped according to building zones and main / branch topology relationships, outputting four main zone object lists. Actual deployment verification shows that the automatically generated object list is completely consistent with the physical configuration, ensuring a one-to-one correspondence between the parallel acquisition scheduling numbers and actual locations, providing a solid foundation for subsequent distributed parallel acquisition, sensor scheduling, and data traceability. In practical applications, zero false nodes, zero redundant objects, and 100% accuracy in network-wide object statistics are achieved.

[0043] S1.2: For each electrical control cabinet node, the bus control module schedules in parallel to drive various types of sensor nodes (including current, voltage, temperature, humidity, vibration, and communication status) deployed in the cabinet to enter the data acquisition working state, so as to realize the timed synchronous sampling of multi-source raw monitoring data.

[0044] Based on the multi-device distributed acquisition object list, the input conditions include the information of each electrical control cabinet node defined in the distribution table, as well as the sensor type configuration and parameter settings assigned to each node;

[0045] A parallel bus scheduling algorithm (parameters: bus address table, device priority, acquisition cycle configuration) is adopted to realize task-level parallel scheduling of all sensor nodes (including current sensors, voltage sensors, temperature sensors, humidity sensors, vibration sensors and communication status monitoring modules) distributed in each electrical control cabinet;

[0046] Furthermore, by using the multi-type sensor synchronous wake-up command sending module (parameters: control cabinet node number, sensor ID list, wake-up sequence), the acquisition channels of each type are switched to the acquisition working state, ensuring that all distributed sensors are turned on in tandem within the predetermined sampling period, thereby achieving consistency guarantee of sample-level time accuracy.

[0047] Furthermore, based on the multi-source timing synchronous sampling algorithm (parameters: sampling period T, synchronization fault tolerance threshold Δt, data packet sequence number), the data sampling trigger action of all activated sensors is executed to obtain multiple types of monitoring data such as electrical parameters, thermal parameters, vibration parameters, and communication signals. In addition, the original signal data stream synchronously collected across the entire network is recorded according to the timestamp for each round of sampling.

[0048] Furthermore, through the local data caching and transmission queue management module (parameters: sampling buffer size, queue priority scheduling, real-time feedback flag), the collected multi-source monitoring data is temporarily stored locally and forwarded efficiently to prevent sampling data loss and data packet timing disorder, and to ensure the integrity of the subsequent data processing chain.

[0049] Through the above parallel scheduling and timed synchronous sampling processing method, all target electrical control cabinets in the acquisition object distribution table constructed in the previous step are synchronously switched to online data acquisition state, and various types of sensors are driven to collect multi-source raw monitoring data in batches, realizing the distributed, multi-channel, high-concurrency acquisition technology goal for complex power distribution scenarios;

[0050] For example, in an electrical building section of an industrial park, 96 electrical control cabinets are configured with 4 current sensors, 2 voltage sensors, 2 temperature sensors, 1 humidity sensor, 1 high-speed MEMS vibration sensor, and 1 communication status monitoring module per cabinet, for a total of 960 acquisition channels. Under a distributed parallel bus control framework, the acquisition period T is set to 1 second, and the synchronization fault tolerance threshold Δt is set to 10 ms. Bus priority queues are allocated sequentially using an address-indexed scheduling algorithm. Sensor nodes are synchronously woken up via CAN bus communication, receiving sampling activation commands in batches from the central scheduling module, and all nodes enter the acquisition working state within 0.2 seconds. After each round of timed synchronous sampling, all channels upload data packets with a unified sampling time tag, generating 960 sets of multi-type raw signal streams per second. For occasional single-point loss of synchronization during sampling, the fault tolerance mechanism automatically records data with a capture delay of less than 5 ms, ensuring that ≥99.99% of data packets across the network are acquired in chronological order without loss. Through caching queues and priority forwarding, the initial import of collected data is completely unblocked. The actual measured sampling latency under maximum load is <15ms, node online rate is 100%, and sampling integrity rate reaches 99.999%. Ultimately, in the example scenario, high-frequency synchronous acquisition from nearly ten thousand sensors, distributed scheduling, and lossless concurrent data collection were achieved, providing a comprehensive multi-source raw data foundation for the high-precision intelligent analysis module.

[0051] S1.3: Acquire raw signal streams from multiple types of sensors, including electrical parameters, thermal parameters, vibration parameters, communication signals, and historical fault information, and attach a unique physical location tag to each data stream based on the sensor ID and sampling timestamp; S1.4: Use a cabinet structure recognition algorithm to perform structure mapping on the sensor output data, and accurately bind the raw monitoring data with the corresponding electrical control cabinet physical unit, sub-distribution power supply module, and bus segment code to generate equipment structure tags;

[0052] S1.5: Combining the distributed network communication protocol, the monitoring raw data packets, which have been assigned physical location and cabinet structure tags, are transmitted to the central data acquisition server via multicast, and each packet is identified by network topology number to ensure spatiotemporal consistency and data traceability;

[0053] S1.6: For all raw monitoring data streams stored in the central data acquisition server, multi-level indexing and integration are performed based on physical location tags, cabinet structure tags, and network topology tags to establish a multi-source, multi-dimensional raw monitoring data database with node-time-space mapping relationships, in order to support subsequent data normalization and intelligent analysis processes.

[0054] Step S2: Normalize the raw data from the multi-source, multi-dimensional monitoring. Based on the physical location of each sensor, cabinet structure, and different environmental conditions, perform preprocessing operations such as time-domain alignment, noise reduction, and unit standardization to generate a uniformly calibrated feature dataset. Specifically, this includes:

[0055] S2.1: The raw data of multi-source and multi-dimensional monitoring is classified and indexed based on the physical location of the sensor, the cabinet structure and the sampling timestamp. The time-series synchronization labeling algorithm is used to perform network-wide time-domain alignment processing on each group of sampling data to generate data blocks with unified reference time and spatial labels, laying the time-space foundation for subsequent noise suppression and calibration.

[0056] The raw monitoring data from various types of sensors in the distributed electrical control cabinet is used as input. The input data already includes the physical location of the sensors, cabinet structure labels, and precise sampling timestamps.

[0057] By adopting a data classification and index mapping method (parameters: physical location label, cabinet structure label, sampling timestamp), various types of collected data are classified and aggregated according to multi-dimensional attributes of space, structure and time, forming a data index chain with multi-level retrieval capabilities;

[0058] Furthermore, through the time synchronization marking algorithm (parameters: full network sampling time base, clock offset correction model, fault tolerance window Δt), all sampling timestamps are globally corrected, and the same batch of data distributed in different control cabinets and physical locations are synchronized and aligned based on the preset sampling step size, generating a unified time base sampling bucket across nodes and cabinets;

[0059] Furthermore, by adopting the linkage coding rule of spatial labels and structural labels, spatial location index and cabinet structure code are automatically added to each group of time-domain aligned data blocks, realizing the three-dimensional integrated data grouping and binding of time, space and structure;

[0060] Furthermore, a data integrity verification algorithm (parameters: data packet ID, sampling round, index hash) is applied to perform integrity and uniqueness checks on the data blocks that have been classified and spatiotemporally synchronized in sequence, to ensure that the correspondence of the sampling data of each channel in each data block, the synchronization window and spatiotemporal label are correct, and to mark the location of missing or out-of-synchronization data.

[0061] By using multi-level index fusion and group verification processing, data blocks of physical location, cabinet structure, and sampling time sequence ternary mapping are output in a unified mode, laying a high-precision data foundation for subsequent processing steps such as noise elimination, scale normalization, and feature vector construction.

[0062] Through the above processing, the multi-source raw data independently collected in the previous step are mapped into a unified spatiotemporal aligned data block, realizing a network-wide consistent time base standard and structured output data format for multi-sensor asynchronous sampling data, which significantly improves the accuracy and spatiotemporal traceability of subsequent data normalization and intelligent analysis.

[0063] For example, in the power distribution system of a four-story large commercial complex, 128 electrical control cabinets are deployed, each integrating 8 sensor nodes (4 for current, 2 for temperature, 1 for vibration, and 1 for communication), collecting data every second. The physical location is coded as Lx-Ry-Cz-Nn (floor-area-cabinet number-node number), the sampling timestamp is synchronized with the UTC standard, the maximum sampling step is 1 second, and the clock tolerance offset Δt is set to 5 ms. For the 128 cabinets × 8 channels of data stream, spatial grading (floor, area), structural grading (main circuit, branch circuit), and sampling time grading (one time base bucket per second) are used, generating a total of 128 × 8 × 3600 data blocks / hour. A timing synchronization marking algorithm is used to correct the sampling time difference, aligning the sampling instantaneous data of all network nodes to within ±2 ms. Furthermore, through a three-dimensional integrated index and data integrity verification, 1.6% of missing data caused by instantaneous network latency was detected and automatically supplemented, ultimately achieving 99.98% complete alignment of the sampled data in terms of space, structure, and time base. All data blocks can be uniquely traced to specific sensor nodes, cabinet locations, and physical structures. The implementation of this step provides a high-precision, hierarchically indexable data foundation for subsequent noise suppression, feature standardization, and intelligent risk analysis. During the testing period, data consistency, alignment accuracy, and spatial mapping traceability all reached industry-leading levels.

[0064] S2.2: Using the spatiotemporally aligned data block output in step S2.1 as input, a multidimensional wavelet denoising algorithm and adaptive median filtering technology are used to perform noise suppression and anomaly removal processing on the sensor signals of each physical unit under different cabinet structures and environmental conditions, thereby eliminating random noise, power frequency disturbances and isolated outliers in the original data of multi-source monitoring, and ensuring the fitting degree and robustness of feature data.

[0065] S2.3: For data blocks that have completed noise suppression, according to the sensor technical parameter table and cabinet physical structure association information, implement unit standardization algorithm and dynamic interval linear stretching, and perform dimensionless unified processing on various electrical parameters, thermal parameters and vibration parameters according to the set statistical interval and physical magnitude to obtain cross-device and cross-cabinet dimensional consistency characteristic data;

[0066] S2.4: Based on the obtained scale consistency feature data, normalized residual compensation is performed using environmental attribute factors (such as temperature and humidity, changes in surrounding load, etc.). The data offset caused by sensor distribution, cabinet structure differences, or changes in environmental conditions is supplemented by a multivariate environmental adaptation normalization model, further optimizing the overall accuracy of feature data.

[0067] S2.5: For the feature dataset that has undergone the above multi-level standardization and normalization, feature vector groups are constructed based on device network topology labels and historical operating status. Through dynamic feature aggregation algorithm, the output results of multiple links such as spatiotemporal alignment, noise cancellation, scale consistency, and environmental compensation are integrated into a unified calibrated feature dataset, which serves as the sole input for the subsequent abnormal resonance identification algorithm, ensuring that the upstream data processing is intelligent and standardized throughout the entire chain.

[0068] Step S3: Input the uniformly calibrated feature dataset into the abnormal resonance identification algorithm module, and use time series cross-correlation analysis, causal inference methods, and segmented clustering algorithms to extract abnormal resonance feature patterns between each cabinet and its associated equipment, thereby obtaining the initial set of equipment clusters with synchronous and cascaded evolution, such as... Figure 2 As shown, it specifically includes:

[0069] S3.1: Group the various time-series feature parameters (such as electrical parameters, thermal parameters, vibration parameters, etc.) in the uniformly calibrated feature dataset according to the electrical control cabinet and network topology labels. Based on the multi-dimensional time series structure, extract the basic time-series attributes of the input objects to obtain the time-series feature matrix of the cabinet and equipment to be analyzed.

[0070] S3.2: Based on the above time series feature matrix, the correlation between the main sequence and response sequence of each pair of device nodes is calculated using the time series cross-correlation analysis algorithm to quantify the intensity of synchronous change when abnormal fluctuations occur, thereby obtaining the cross-correlation coefficient matrix of the correlation between devices;

[0071] For the time series feature matrix data output by step S3.1, the input objects include multi-dimensional time series feature parameters grouped by physical location, cabinet structure, and network topology labels;

[0072] The time series cross-correlation analysis algorithm (parameters: input main sequence, response sequence, synchronization window length τ, time lag range Δl) is adopted. For each pair of electrical control cabinet equipment nodes, the cross-correlation function of the main sequence and response sequence is calculated in batches according to the set sliding window to realize the quantitative extraction of the synchronization change intensity between equipment nodes.

[0073] Furthermore, by setting covariance normalization constraints, zero mean and unit standard deviation preprocessing is applied to each group of main and response sequences to eliminate the differences in influence caused by different physical quantities and improve the comparability of cross-correlation analysis;

[0074] Furthermore, the cross-correlation coefficient R between the main sequence x(i) and the response sequence y(i) under a time lag l is calculated using the following formula. xy (l) Perform calculations:

[0075]

[0076] Where N is the observation window length and l is the number of lag steps. These are the sequence mean and σ, respectively. x σ y These are the standard deviations;

[0077] Furthermore, after accumulating all node combinations, according to the lag step size and correlation coefficient threshold screening strategy, node combinations below the correlation threshold are set to zero, thereby improving the sensitivity of abnormal fluctuation synchronization relationship screening.

[0078] Furthermore, the calculated cross-correlation coefficients of all node combinations are organized into a cross-correlation coefficient matrix M, with node pairs as row and column indices. corr It is used to quantitatively describe the intensity of synchronous changes between various device nodes, providing an input benchmark for subsequent causal discrimination and cluster analysis.

[0079] Through the above chain algorithm processing, the multi-cabinet and multi-physical parameter time-series feature matrix output in step S3.1 is transformed into a cross-correlation coefficient matrix between equipment nodes with synchronous correlation labels, which effectively quantifies the temporal correlation of risks in complex power distribution network equipment groups and achieves the key technical effect of early capture of abnormal changes within the group.

[0080] For example, in the power distribution system of a large manufacturing plant, the A-term current sensor outputs of each of the six main electrical control cabinets in zone L1 are selected. The synchronization window length τ is set to 300 seconds, and the lag step Δl is set to -10 to +10 seconds. The main sequence of each control cabinet is combined with the response sequences of the other five cabinets, and the correlation strength between each pair of nodes at each lag step is calculated according to the above cross-correlation coefficient formula. Before the actual calculation, the mean-standard deviation normalization is performed on all current sampling sequences. Within the sample, during a plant-wide load switch, the maximum cross-correlation coefficient between cabinets L1-1 and L1-2 reaches 0.86 (lag of 2 seconds), the maximum value between cabinets L1-1 and L1-3 is 0.32 (no significant lag), and the maximum cross-correlation value between cabinets L1-4 and L1-5 is less than 0.1. Based on the correlation coefficient threshold of 0.6, only the L1-1 / L1-2 combination is marked as significantly synchronized. Finally, a 6×6 cross-correlation coefficient matrix is ​​output, with the diagonal lines representing autocorrelation (all 1), and the remaining elements filled according to the maximum correlation coefficient output by the algorithm. When the system verifies synchronous anomaly alarms, because the cross-correlation coefficient matrix is ​​dynamically updated, it can accurately identify abnormal cabinet combinations that exhibit synchronous large fluctuations in the same time period, achieving sensitive capture and early alarm of cross-cabinet risk linkage changes;

[0081] S3.3: Using the cross-correlation coefficient matrix as input, causal inference methods (such as Granger causality test and controlled transfer entropy analysis) are used to further determine the temporal driving relationship of abnormal fluctuations among various devices, realize the quantitative identification of synchronous and cascading influence paths, and output the causal index of abnormal driving between device nodes.

[0082] The cross-correlation coefficient matrix M output in step S3.2 corr For the input object, the feature parameters have been normalized according to physical location, cabinet structure and time-series attributes, and the corresponding sensors are time-normalized feature quantities such as electrical parameters, thermal parameters, vibration parameters, etc.

[0083] Causal inference method is used (parameters: node combination pairs, time series data window T) win Maximum lag order L max This enables the quantitative identification of the driving relationship of abnormal fluctuations among various device nodes. Furthermore, using the Granger causality test algorithm, the master-slave sequence of each pair of device nodes is tested at a set maximum lag order L. max Within this framework, an autoregressive model and a controlled regression model are constructed sequentially between device node A and device node B, and the following causality discriminant statistic is calculated:

[0084]

[0085] Among them, RSS r The sum of squared residuals of the controlled regression model, RSS urLet L be the sum of squared residuals of the autoregressive model, N be the sample size, and L be the sum of squared residuals. max As the lag order, the Granger causality significance of a single pair of nodes is determined by comparing the calculated F-statistic with the significance threshold.

[0086] Furthermore, controlled transfer entropy analysis (parameters: node time series, condition variable set, embedding dimension m) is used to quantitatively assess the information transfer driving strength between nodes. The core formula is:

[0087]

[0088] Among them, TE A→B Let b be the information transfer entropy from node A to node B. t+1 Let B be the state of node B at time t+1. and The m-dimensional embedded historical states of node B and node A at time t;

[0089] Furthermore, by performing the above causality test and information transfer entropy analysis on all node pairs respectively, a set of anomaly-driven causality indicators between node pairs is generated, including multivariate indicators such as significance P-value, F-statistic, and transfer entropy value TE, so as to achieve quantitative output of the primary and secondary relationships of nodes, synchronous and cascading anomaly-driven paths;

[0090] Furthermore, based on node pairs with significant causality or high transfer entropy thresholds, the structured output device node anomaly-driven causality index matrix M is generated. causal As the result of the discrimination of synchronous and cascading association patterns, it provides anomaly-driven relationship input for subsequent clustering and trust graph generation;

[0091] By using a chain combination of causal inference methods, the cross-correlation coefficient matrix M corr Transformed into an anomaly-driven causality index M that includes the main driving direction, cascaded links, and causality strength identifier. causal It enables accurate identification of the driving path of dynamic abnormal fluctuations in multiple devices in complex power distribution networks;

[0092] For example, in the power distribution system of a high-rise office building, for eight electrical control cabinets distributed in the same power supply area, the abnormal current characteristic sequence of their respective main circuits is extracted, and a causal verification timing window T is set. win = 600 seconds, maximum lag order L max =8. Taking cabinets A and B as a group, the Granger causality test algorithm is applied to fit the autoregressive and controlled regression models of A on B and B on A, respectively, to obtain F. A→B=6.87, corresponding to a p-value of 0.0035 (less than the 0.01 significance threshold), indicating a significant causal relationship between A and B. Using the controlled transfer entropy algorithm with an embedding dimension of m=4, the joint probability density of the time series data of A and B is estimated to obtain TE. A→B =0.145, which is higher than the set high-drive threshold of 0.1, further confirming that there is a significant abnormal drive in the information transfer from A to B. Based on the causal discrimination results of all the above pairs of nodes, with AB and CD as the main pairs, the abnormal drive causality index matrix M is input. causal Finally, an 8×8 node causality matrix is ​​output, and the combination of nodes with significant values ​​in the matrix represents the anomaly driving link;

[0093] In the event of a sudden change in load across the entire plant, step S3.3 can output the causal driving path between all node pairs within one minute, accurately revealing cabinet A as the core driving node and B, C, etc. as controlled nodes, greatly improving the early identification capability of cascading group anomalies and the visualization effect of dynamic links. This causal index matrix provides a key data foundation for subsequent equipment cluster clustering and adaptive trust graph construction;

[0094] S3.4: Based on the cross-correlation coefficient matrix and the anomaly-driven causality index, a fast segmented clustering algorithm (such as spectral clustering and density peak clustering) is applied to classify the equipment nodes into clusters. The equipment nodes that evolve synchronously or have time cascade effects are divided into anomaly resonance feature pattern groups, forming a preliminary equipment cluster division result.

[0095] The cross-correlation coefficient matrix M output from the preceding step S3.3 corr With the abnormal driving causality index matrix M causal For the input data, the input objects already include the intensity of synchronous changes between devices, master-slave causal relationships, multi-dimensional time-series characteristic parameters, physical location and cabinet structure labels;

[0096] A fast segmented clustering algorithm is used (parameters: node feature matrix, similarity threshold θ). s Causality weight w c Minimum cluster size n min ), in conjunction with M corr and M causal Two types of indicators are used to calculate the comprehensive similarity index S between pairs of device nodes. ij This reflects the resonant clustering tendency of the synchronization and cascading relationships between nodes;

[0097] By constructing a comprehensive similarity matrix, a spectral clustering algorithm (parameters: Laplace matrix L, number of clusters K) is used to determine the cluster affiliation of all network device nodes. Specifically, the standardized Laplace matrix is ​​first formed according to the following formula:

[0098] L = ID -1 / 2 WD -1 / 2

[0099] Where W is the overall similarity matrix, D is the degree matrix, and I is the identity matrix;

[0100] Furthermore, the eigenvectors corresponding to the K smallest eigenvalues ​​of L are calculated to form a three-dimensional feature space. The K-Means algorithm is then used (parameters: cluster center K, iteration upper limit T). max The node projections are grouped and partitioned to obtain preliminary clustering results of the device nodes;

[0101] Furthermore, for node combinations with significant synchronous evolution or those exhibiting master-slave causal driving chains, a density peak clustering algorithm (parameter: local density threshold ρ) is employed. T Distance threshold δ T This further refines the heterogeneity within large clusters, improving the resolution of anomalous resonance feature patterns.

[0102] By marking the resonance type attributes (such as synchronous resonance and hysteresis) of each node pair within the cluster in the clustering results, an abnormal resonance feature pattern group data structure with the node set as the core is generated.

[0103] By combining multiple stages of rapid segmentation clustering, spectral clustering and density peak clustering, the input multidimensional synchronization and causality indicators are transformed into preliminary classification results of abnormal resonance feature patterns at the equipment group level, thereby realizing the structured identification of synchronous evolution and cascading abnormal trends among multiple devices in complex power distribution networks.

[0104] For example, in the low-voltage power distribution network of a high-rise commercial complex, 12 electrical control cabinets distributed across two floors were selected, and M-type control was implemented to monitor the current and temperature characteristics of each main circuit. corr With M causal Dual-indicator fusion, setting a similarity threshold θ s =0.65, number of clusters K=4, minimum cluster size n min =2. After spectral clustering, four clusters were formed; further subdivision was performed using density peak clustering, identifying one group in each spatial layer that appeared in the same time period and had a high correlation with current synchronization (S). ij ≥0.8) and the existence of a temperature cascade causal relationship (TE) A→B Anomaly nodes with a value ≥0.12 were identified. The final clustering formed five anomalous resonance feature pattern groups, covering nine devices exhibiting anomalous fluctuations. Structured outputs, including node group attribute labels, resonance type, and response delay, were generated. System operation tests showed that this strategy improved the cross-cabinet cascading anomaly detection rate by more than 15% compared to simple correlation clustering, effectively supporting the intelligent identification of spatially coordinated and temporally cascading risks and subsequent trust graph construction.

[0105] S3.5: Perform statistical characteristic screening and labeling on the abnormal resonance feature pattern groups obtained by clustering, and label and output the feature labels (such as resonance type, response delay, and influence intensity) of the device node clusters in a structured manner to obtain the initial set of synchronous evolution and cascade evolution device clusters for downstream trust graph generation.

[0106] Step S4: Based on the abnormal resonance feature pattern, combined with the power distribution network topology and the physical attributes of each device, a Bayesian update and attention mechanism is used to calculate the risk transmission weights and directions between all device nodes, generating an adaptive device trust graph containing risk relationship edge weights and node labels. Figure 3 As shown, it specifically includes:

[0107] S4.1: Multi-dimensional data fusion is performed on the abnormal resonance feature pattern output from the abnormal resonance feature extraction submodule, as well as the power distribution network topology parameters and equipment physical attribute parameters to construct a multi-dimensional feature vector group, so as to realize the unified expression of the interaction information of each equipment node under abnormal conditions and obtain a high-dimensional risk correlation input matrix.

[0108] The input data includes the abnormal resonance feature pattern output by the preceding abnormal resonance feature extraction submodule, the multi-node interaction time series feature parameter set, the structured topology parameters of the power distribution network, and the physical attribute indicators of each device node.

[0109] By using anomaly resonance feature pattern structured data, the abnormal label, response latency, impact intensity, resonance type, and the device cluster identity to which the node belongs are extracted for each device node within a specific time window, so as to realize the dynamic marking of key nodes under abnormal conditions.

[0110] Furthermore, by integrating information such as topology labels, physical connections, upstream and downstream hierarchical structures, and spatial geographic coordinates of distribution network nodes through the network structured parameter interface, a sparse matrix of inter-node connection relationships and a spatial adjacency weight matrix are constructed.

[0111] Furthermore, the physical attribute parameters of each device (such as device model, current capacity, fault tolerance threshold, historical fault statistics, remaining life assessment, etc.) are normalized and quantified to generate attribute vectors that support the fusion of abnormal data.

[0112] Furthermore, a feature concatenation and multi-dimensional vector fusion algorithm (parameters: feature standardization interval, missing attribute completion strategy, concatenation order) is adopted to combine the anomaly label vector, topological connection matrix and physical attribute normalization vector into a node-level high-dimensional feature input vector, and this vector is used as a unit to concatenate with the features of adjacent nodes to form edge weight candidate features.

[0113] Furthermore, through feature encoding and synthesis rules, the feature concatenation results of all node-node pairs are integrated into a high-dimensional risk association input matrix. Each row of the matrix represents a risk interaction encoding vector of a pair of nodes, covering multi-dimensional dynamic characteristics such as anomaly label, response intensity, spatial adjacency, and physical fault tolerance.

[0114] Through the above multi-stage data processing and fusion operations, the original indicators scattered across multiple dimensions such as anomaly detection, physical attributes, and network topology are uniformly transformed into a structured, highly available, high-dimensional risk correlation input matrix. This provides standardized data support for subsequent conditional probability modeling, Bayesian parameter iteration, and attention mechanism weighting, enabling a multi-dimensional unified expression of interactive information between electrical control cabinet equipment nodes under abnormal conditions.

[0115] For example, in a power distribution network of a large industrial park, for 20 main circuit electrical control cabinets distributed in area A1, the abnormal resonance characteristic patterns of each device in the most recent 24-hour period are collected. These include the maximum abnormal fluctuation amplitude (range 0-1), average response delay (range 0-30 seconds), the cluster label (e.g., Cluster-2), and the main drive type (e.g., synchronous or cascaded). The power distribution network topology is output by an automatic network scanning tool. The network is a hierarchical tree structure, with each node having upstream and downstream connections and three-dimensional geographic coordinates, and a spatial adjacency radius limited to within 20 meters. The physical attributes of each device node (e.g., model number: TBB-1200, current capacity: 800A, historical failure rate: 0.005, remaining life assessment 80%) have been standardized and normalized. Using a feature splicing strategy, the abnormal label, physical attribute encoding, and network topology attributes of each device node are sequentially concatenated to form a high-dimensional node feature vector of length 32. The system then performs feature concatenation on all node pairs that may have direct physical connections or anomalous resonance-driven relationships within the network, resulting in a high-dimensional feature vector set of 400 candidate node pairs. All node pair feature vectors are organized in rows by node-node pair, outputting a 400×32 high-dimensional risk association input matrix. The system periodically updates this input matrix for all newly acquired data within 48 hours, thereby initializing the network-level risk transmission conditional probability, subsequent Bayesian iterative correction, and attention weighting. In real-world testing scenarios, this fusion matrix input algorithm achieves consistent modeling of multiple characteristics such as node resonance attributes, spatial distribution, and physical margin, establishing an efficient foundation for automated mapping of trust relationships in downstream equipment and accurate identification of dynamic risk weights. This fusion scheme supports integrated description of single-node-multidimensional and cross-node-multi-attribute relationships, enabling dynamic modeling and data standardization of large-scale equipment node relationships in complex power distribution networks, significantly improving the precision of dynamic risk detection for equipment groups and the quality of model input.

[0116] S4.2: Based on the high-dimensional risk association input matrix, the conditional probability inference algorithm is applied to initialize the probability of potential risk transmission relationships between each device node. Based on the abnormal resonance characteristics and prior network and physical attributes, an initial risk transmission probability matrix is ​​generated.

[0117] The input condition is the high-dimensional risk correlation input matrix output in step S4.1. The matrix integrates the abnormal resonance feature patterns, time-series statistical parameters, power distribution network topology labels, and various physical attribute normalization indicators between each pair of electrical control cabinet equipment nodes.

[0118] A conditional probability inference algorithm (parameters include: node feature vector set, initial feature prior P0, anomaly resonance index, network connection weight, physical fault tolerance, etc.) is used to probabilistically model the potential risk transmission relationship between pairs of device nodes. Furthermore, through joint distribution modeling of high-dimensional feature variables, the conditional probability P of a risk transmission relationship between any node i (risk source) and node j (response object) is calculated. ij Its core probability formula is:

[0119]

[0120] Among them, X i Let F be the anomaly label feature vector of node i. ij T is an index of abnormal resonance intensity. ij For network topology adjacency weights, A ij For physical property matching degree;

[0121] Furthermore, by using Gaussian mixture models or kernel density estimation to perform joint probability density modeling on high-dimensional feature variables, a weighted focus on the conditional probabilities of node pairs with strong anomalous resonance, topological compactness, and insufficient physical margin in the feature space is achieved. Furthermore, based on prior indicators such as anomalous resonance type (synchronous / cascaded), spatial distance between nodes, and historical failure rate, prior weights are set for the risk transmission probability of different node pairs, increasing the probability of node pairs with high dynamic correlation and known historical driving relationships, while decreasing the probability of low-impact or distant node pairs.

[0122] Furthermore, by combining statistical results within the time window of the abnormal resonance event, the maximum likelihood estimation method is used to optimize the risk transmission probability of node pairs, improving the fitting accuracy of the initial matrix to the actual operating state and risk propagation path. Further, the conditional probability outputs of all archived device node pairs are normalized to generate an initial risk transmission probability matrix of size N×N. The matrix rows and columns correspond to the risk source nodes and response nodes, respectively, and the matrix elements are the risk transmission probability values ​​from node i to node j.

[0123] Through the above probability inference and normalization process, the high-dimensional risk association input matrix is ​​systematically transformed into an initial risk transmission probability matrix with physical causality, data correlation and network structure attributes, realizing the explicit quantification of the potential for hidden danger propagation among multiple devices in complex power distribution networks, laying the foundation for subsequent Bayesian dynamic correction and adaptive trust graph evolution.

[0124] For example, in a test scenario of a large-scale intelligent building power distribution system, the system is configured with 16 main electrical control cabinets. The frequency of historical abnormal events for each cabinet ranges from 3 to 65, and the spatial adjacency weight T is... ij Value range: 0.2-1.0, Physical attribute matching degree A ij The ratio of the capacities of the two devices is standardized to between 0 and 1. For all node combinations, the system first integrates the anomalous resonance intensity F. ij The quantitative output from the previous step (such as F) ij =0.82 represents abnormally active interaction between two nodes), and then input each feature variable into the conditional probability inference formula. For example, for node 5 to node 8, input X5 with an abnormal label of 1, F 5,8 =0.82, T 5,8 =0.95, A 5,8 =0.76, output P after statistical modeling 5,8 =0.67. After normalization, the main diagonal value of the initial risk propagation probability matrix of the entire 16×16 is close to 0. The probability of strong resonant adjacency combinations (F>0.7, T>0.7, A>0.7) is higher than 0.6, and the probability of distant or physically weakly correlated node pairs is lower than 0.2. In six rounds of network anomaly linkage event detection over three consecutive months, the initial probability matrix covered 92.4% of the actual risk propagation links, providing a highly reliable data foundation for subsequent Bayesian iterative optimization and dynamic adaptive adjustment of trust edge weights.

[0125] S4.3: For the initial risk transmission probability matrix, a Bayesian update mechanism is adopted. Combining newly received monitoring data, historical resonance evolution priors and node attribute parameters, the risk transmission weights between nodes are continuously corrected to achieve dynamic optimization of conditional probabilities and output a dynamic risk transmission weight matrix.

[0126] For the initial risk propagation probability matrix of the input The real-time newly collected equipment monitoring data, historical abnormal resonance evolution priors, and physical property parameters of each node are used as input conditions.

[0127] Bayesian update method is used (parameter: prior probability matrix) Latest observation data stream D new Historical Anomalous Events Prior E prior Node attribute vector A attr This enables dynamic conditional probability correction of risk transmission relationships among nodes;

[0128] By calculating the posterior probability of risk propagation for each pair of device nodes (i,j) The following Bayesian update formula is used for conditional probability recursion:

[0129]

[0130] in, Let P(D) be the probability of risk propagation at a previous time step. new |R ij Let ) be the likelihood probability of the new monitoring data under the assumption that there is risk transmission from node i to node j, and the denominator is normalized;

[0131] Furthermore, through the observed monitoring data D new Within a specific time window, the statistical characteristics of anomalous resonance events, such as anomalous amplitude, duration, and dominant frequency, are used to extract an anomalous responsibility index, which is then combined with the prior probability E of historical anomalous events. prior Joint modeling is used to enhance the sensitivity of conditional probabilities to long-term evolutionary trends;

[0132] Furthermore, regarding the device node attribute vector A attr The parameters related to physical fault tolerance, historical failure rate, and remaining lifetime are dynamically weighted, and an attribute-adaptive weighted Bayesian network is adopted to explicitly introduce the effect of node physical attributes on the direction and weight of risk propagation chain into the posterior probability inference process.

[0133] Furthermore, the Bayesian update process described above is iterated periodically, and the risk weights between nodes are dynamically adjusted as monitoring data and historical priors accumulate. Under multiple rounds of data input, a gradually converging dynamic risk transfer weight matrix is ​​formed.

[0134] Through the aforementioned Bayesian update and dynamic weighting techniques, the static initial risk propagation probability matrix is ​​transformed... Based on the actual network operation status, multi-source monitoring data streams and historical anomaly evolution information, it is continuously transformed into a dynamic risk transmission weight matrix that reflects the actual risk relationship between equipment nodes, thereby achieving the effect of real-time conditional probability optimization and quantitative modeling of the risk evolution correlation of electrical control cabinet groups.

[0135] For example, in a typical large industrial park power distribution network scenario, the system monitors 24 main electrical control cabinets within its monitoring area, synchronously updating multi-source data such as current, temperature, and vibration every 10 minutes. Initial risk transmission probability matrix. Based on last month's feature analysis and equipment physical attribute settings, the probability distribution between nodes is between 0.12 and 0.68. The system monitored abnormal synchronous current jumps between cabinets A and B between 2:00 AM and 3:00 AM. The latest observation data is D. new The data shows that the abnormal amplitude of both nodes is higher than 0.82, lasting for 8 minutes. Meanwhile, historical data prior E... prior Analysis revealed 7 abnormal master-slave events between cabinet A and cabinet B within the past three months, with a priori frequency significantly higher than the average for the same period. Cabinet A's attribute vector A... attr This indicates a historical failure rate of 0.028, moderate physical fault tolerance parameters, and a remaining lifespan of 86% for cabinet B, demonstrating high adaptability. According to the Bayesian update formula, the observed likelihood P(D) new |R AB =1)=0.79, P(D new |R AB =0) = 0.15, substitute into the previous probability Output after iterative calculation This step is performed sequentially on all 552 node pairs in the network. After three rounds of real-time data accumulation and continuous correction based on historical events, the final dynamic risk transmission weight matrix is ​​obtained. The weights of medium- and high-relevance edges range from 0.65 to 0.91, effectively highlighting easily propagated fault paths. Field backtracking analysis shows that this matrix can accurately reflect fault linkage trends 30-60 minutes in advance, supporting refined high-risk link identification and subsequent intelligent linkage protection decisions.

[0136] S4.4: Using an attention mechanism, the feature vector group after fusing the dynamic risk transmission weight matrix and device node attributes is deeply weighted to enhance the sensitivity to key high-risk nodes and their associated links, and obtain a weighted structured risk edge weight feature set.

[0137] S4.5: Using a weighted structured risk edge weight feature set as input, and based on the three elements of electrical control cabinet equipment nodes, risk relationship edge weights, and node risk labels, an adaptive equipment trust graph containing dynamic edge weights and node labels is generated through a graph-based data modeling method. This enables full-process digital modeling and dynamic representation of risk transmission behavior between equipment in complex power distribution networks.

[0138] Step S5: For the adaptive device trust map, based on a graph neural network or graph attention network, learn and infer historical and current feature data within a specified sliding time window to dynamically predict the risk diffusion path, impact range, and high-risk node clustering areas among various electrical control cabinet equipment groups. Specifically, this includes:

[0139] S5.1: Based on the adaptive device trust graph generated in the previous step, obtain historical feature data and current feature data within a specified sliding time window, and use them as data input for training and inference of the graph neural network model to ensure time continuity and state diversity, and form a feature vector group based on node-edge attributes;

[0140] S5.2: The node attribute vectors and edge weight relationships of the adaptive device trust graph are aggregated using a graph convolutional neural network (GCN) or a graph attention network (GAT) to calculate the spatial aggregation distribution of risk information among devices and realize the high-order feature extraction of the risk status of the device group.

[0141] Using the node attribute feature vectors and edge weight relationships of the adaptive device trust graph as input conditions, the historical and current feature data within a specified sliding time window are aggregated.

[0142] The Graph Convolutional Neural Network (GCN) algorithm is used (parameter: node feature matrix). Normalized adjacency matrix of a graph Convolution weights W (l) The activation function σ(·) is used to realize the transmission of spatial structure information and feature synthesis between device nodes, and to obtain the high-order spatial features of each layer of nodes.

[0143] The core GCN processing procedure is as follows:

[0144] The spatial aggregation of node features is achieved through convolution operations as shown in the following diagram:

[0145]

[0146] Among them, H (l) Let W be the feature matrix of the nodes in the l-th layer. (l) Here are the trainable weights for this layer, σ(·) is the non-linear activation function, and H... (0) =X is the initial input feature;

[0147] Furthermore, a multi-layer stacking method is adopted to continuously propagate node information and update the spatial aggregation features of nodes at each layer, thereby improving the model's ability to capture the risk interaction characteristics of high-order neighborhoods.

[0148] As an optional technical solution, the Graph Attention Network (GAT) algorithm (parameters: node feature set X, adaptive attention weight coefficient α) is used. ij (Number of heads K) enables the heterogeneity weighting of device nodes and their adjacent edges during spatial aggregation, automatically highlighting the information aggregation contribution of high-risk or critical impact links.

[0149] The main steps of GAT include:

[0150] For each node i, the attention weight of its neighbor node j is calculated using the following formula:

[0151]

[0152] Where W is the learned linear transformation weight, Here, || represents the attention weight vector, and || represents feature concatenation. Let i be the set of neighbors of node i. The input features for node i;

[0153] After calculating the attention weights, the neighbor features of each node i are weighted and aggregated to output the higher-order node features after spatial aggregation:

[0154]

[0155] Furthermore, a multi-head attention mechanism is used to aggregate multiple sets of weights in parallel to improve the model's representation breadth of different spatial relationships. The final node spatial features are output after integrating the results of each head.

[0156] Through the above spatial aggregation steps, the high-dimensional data such as the attribute vectors, risk labels, and edge weight structures of the initial input nodes are aggregated and weighted by heterogeneity in the topological space through multiple rounds of aggregation, thereby realizing the hierarchical and joint spatial encoding expression of the risk information of the equipment group.

[0157] By using spatial aggregation processing of GCN / GAT, the risk features of each node in the adaptive device trust graph are fused together in multiple layers to output a high-order feature vector that expresses the spatial risk status of each device node, providing an accurate, high-dimensional, and structured data foundation for subsequent risk evolution path and diffusion trend modeling.

[0158] For example, in a smart power distribution network in a campus containing 32 electrical control cabinets, a sliding time window of 48 hours is selected. The dimension of the input features collected for each device node is 40, the risk edge weight matrix is ​​generated from the previous stage, and each node has an average of 6 spatial adjacencies. A 3-layer GCN network is adopted, with the output dimension of each layer set to 64, and the first layer convolution weight W (1) Initialized to a standard Gaussian distribution, the weights of layers 2 and 3 are initialized using a He uniform distribution, and the activation function is ReLU. Input feature matrix. Normalized adjacency matrix Spatial aggregation feature extraction is achieved through the following steps:

[0159] (1) Input X is obtained by the first-level GCN transformation. The output size is 32×64;

[0160] (2) H (1) Input to the second-layer GCN, process it in the same way, and output H.(2) The dimension is always 32×64;

[0161] (3) After the third layer of overlay processing, the final spatial aggregation representation H is output. (3) ;

[0162] For the same scene, the GAT method is used, with the number of attention heads initialized to 8. Each head outputs 16-dimensional features, which are then concatenated to form a 32×128 spatial feature vector.

[0163] Network inference results show that after spatial aggregation, the clustering characteristics of historical high-risk nodes and high-risk associated links are significantly improved, the variance of node risk status representation is reduced by 30%, and the accuracy of high-risk cluster area identification is improved to over 97%.

[0164] The final output high-order feature vector serves as the input for subsequent time-series modeling and risk diffusion path determination, realizing information fusion and dimensionality reduction representation of the spatial correlation of risks in equipment groups, and improving the accuracy and efficiency of dynamic detection of group hazards in complex power distribution networks.

[0165] S5.3: After spatial aggregation is completed using graph neural network inference, the risk evolution trend of each node is modeled by time series learning based on the historical and current feature data sequences within the sliding time window, so as to obtain the dynamic change path of risk diffusion between equipment nodes.

[0166] For the high-order node spatial feature vector group obtained after spatial aggregation processing of graph neural network and its corresponding historical and current feature data sequences within the sliding time window, a time series modeling method is used to dynamically analyze the risk evolution trend of each device node.

[0167] A Long Short-Term Memory (LSTM) neural network structure is adopted (parameters: the input dimension is the dimension of the node spatial feature vector, the number of hidden units and the time step are set according to the sliding window length, and the activation function is tanh). The spatial feature sequence of each device node within the sliding window is processed step by step to realize the temporal dependency modeling of the node risk state.

[0168] Furthermore, by extending the model with bidirectional LSTM (parameters: forward and backward time series, hidden layers can be shared or set independently), the model's ability to capture the causal relationships between the past and future in the risk evolution process is enhanced, and the hidden state output of each time step is extracted to characterize the dynamic changes in the risk trend of the nodes.

[0169] Furthermore, a temporal attention mechanism (parameters: time step weight allocation rules, weighting coefficient optimization direction) is used to weight and aggregate the hidden states of each time step in the LSTM output, highlighting the contribution of recent high-risk characteristics and trend changes to the overall risk diffusion path modeling;

[0170] Furthermore, for the time-series output of all device nodes, a risk state collaboration matrix between nodes is constructed, and the Dynamic Time Warping (DTW) algorithm (parameters: optimal pairing distance metric, constraint radius) is used to measure the similarity of the risk evolution trajectory of each node and extract the critical path node combination of risk diffusion.

[0171] Through the above-mentioned time-series learning and dynamic clustering methods, a set of dynamic change paths reflecting the risk diffusion between device nodes is generated, and node risk trend sequences and diffusion link structure data are output, realizing the modeling of multi-device risk evolution chains across time and space.

[0172] By using temporal modeling and aggregation reasoning, the high-order feature vector sequence after spatial aggregation is transformed into a dynamic path for node risk diffusion, thereby enabling accurate capture and visualization of the evolution link of multi-device collaborative risk trends in the power distribution system.

[0173] For example, in a large-scale distributed power distribution network scenario with 32 electrical control cabinets in a factory, the spatial features of each node are processed over a 48-hour sliding window, with spatial feature sequences input every 10 minutes, totaling 288 frames. A single-layer LSTM is used, with an input dimension of 64, 128 hidden units, and a time step of 288. The parameters are automatically fitted by the Adam optimizer. A bidirectional LSTM is used to expand the same data, and the preceding and following temporal hidden states are concatenated for output, improving the detection capability of preceding and following links driven by sudden risks. The temporal attention mechanism uses softmax to allocate weights, strengthening the influence of the latest risk event on the aggregation trend. Analyzing the hidden state matrix output by the LSTM, the risk trends of nodes A, B, and C increase synchronously between 20:00 and 00:00, with the smallest DTW distance, and the corresponding path is dynamically marked as a high-risk diffusion path. All node combinations are evaluated, and a risk diffusion path matrix and dynamic collaborative relationships between nodes are output. The system detects 7 main diffusion links. In actual verification, this method can achieve link-level monitoring of sudden cascading failures 30 minutes in advance, and the accuracy rate of identifying high-risk paths across the entire network is 96%, laying a high-precision foundation for subsequent high-risk cluster area determination and graded alarm triggering;

[0174] S5.4: Determine the risk impact range of the risk diffusion path results output by the model, calculate the potential high-risk node cluster area based on the node risk weight and adjacency topology, and clarify the spatial concentration and propagation boundary of group risk;

[0175] S5.5: Based on the risk diffusion path and the identification results of high-risk node clusters, generate a dynamic risk prediction report between equipment groups, output a detailed risk association link, a list of key nodes to focus on, and a predicted risk level, providing data input for subsequent graded alarms and prevention and control decisions.

[0176] Step S6: Determine whether the risk weight, diffusion rate, and spatial distribution of high-risk paths in the graph reasoning results exceed a set threshold. If the conditions for group risk cascading or evolution triggering are met, then determine and output graded alarm signals and corresponding protection plan parameters based on the trust relationship link level, specifically including:

[0177] S6.1: Perform risk weight quantification calculation on high-risk path nodes of adaptive device trust graph and graph neural network inference, and use weighted boundary value accumulation and node risk label aggregation algorithm to obtain comprehensive risk weight index of high-risk path in order to quantify the impact of potential path.

[0178] S6.2: Based on the temporal node state changes of high-risk paths, perform dynamic calculation of diffusion speed, and use the risk weight increment method within the moving window to compare with the synchronous state transition of nodes to achieve quantitative assessment of the hazard diffusion rate and output risk propagation speed characteristic parameters.

[0179] Using the historical and current temporal feature data sequences of high-risk path nodes obtained by graph neural network inference as input objects, and relying on the temporal risk status changes of each node after high-order spatial feature aggregation, it is used for the quantitative assessment of the subsequent hidden danger risk diffusion rate.

[0180] A moving-time window sliding algorithm (parameters: window length W, step size Δt) is used to continuously increment the risk weight sequence of high-risk path nodes within each window. Specifically, the risk weight increment at the i-th time step is calculated using the following formula:

[0181] ΔR i =R i -R i-1

[0182] Among them, R i This represents the risk weight of a high-risk path node at time step i. This yields the incremental sequence of risk weights for each node within the sliding window.

[0183] Furthermore, a synchronous state transition comparison algorithm (parameter: risk state discrimination threshold θ) is used to mark the risk state transition times of each node within the same window, comparing the order and synchronization degree of high-risk state switching between nodes. Let the risk state transition times of nodes A and B within the window be t respectively. A With t B Define the synchronization state transition difference ΔT as:

[0184] ΔT AB =|t A -t B |

[0185] Where, if ΔT ABIf θ < , then nodes A and B are considered to be in a synchronous state transition;

[0186] Furthermore, combining the aforementioned incremental sequence and synchronous state transition interval values, a diffusion rate quantification model is employed to quantitatively estimate the diffusion rate of potential hazards along high-risk paths based on the cumulative rate of change of risk weights and the spatial / topological distance between synchronous transition nodes. This process is specifically implemented using the following formula:

[0187]

[0188] Among them, V diff For the characteristic parameter of risk diffusion rate, W represents the average state transition interval for all synchronous transfer nodes, and W is the length of the sliding window.

[0189] Furthermore, the above-mentioned quantitative assessment of the spread rate is performed on all high-risk paths, the spread rate characteristics of each path are collected, and those paths whose spread rate exceeds the system's preset threshold (such as the upper quantile or empirical threshold) are marked as rapidly spreading potential links.

[0190] By combining the risk weight increment method with the synchronous state transition comparison, the temporal characteristics of high-order nodes are transformed into path-level risk diffusion rate indicators, enabling quantitative modeling of the rate of hidden danger diffusion and screening of key links, providing solid quantitative support for downstream group risk cascading alarm decision-making.

[0191] For example, in the scenario of high-risk diffusion path analysis in a large industrial park power distribution network, for the high-risk path connecting nodes A, B, and C, assuming a sampling period of 5 minutes and a sliding window length W of 12 (i.e., 1 hour), the risk weight sequence of node A is {0.1, 0.15, 0.18, 0.22, ...}, and nodes B and C are {0.09, 0.14, 0.17, 0.20, ...} and {0.08, 0.12, 0.16, 0.19, ...}, respectively. For node A, the weight increment ΔR at each step is calculated according to the above formula. i For example, from step three to step four, ΔR4 = 0.22 - 0.18 = 0.04. The transition times for each node to reach a high-risk state (discrimination threshold θ = 0.18) are t. A =15 minutes, t B =20 minutes, t C =25 minutes, for node AB ΔT AB =5 minutes, BC node ΔT BC =5 minutes, arithmetic mean Minutes. Cumulative weight increment. ΔR i =0.4. Substitute into the formula,

[0192]

[0193] The diffusion rate was compared with the system's set stratification threshold of 0.005. Since 0.0067 > 0.005, it was determined to be an ultra-high-speed diffusion link. Actual evaluation results show that this method improves the sensitivity of capturing the diffusion rate of rapidly cascading hazards to over 95%, and maintains a 99% accuracy rate in identifying slow paths, effectively supporting the quantitative implementation of a stratified strategy for group risk alarms.

[0194] S6.3: Using the physical location labels of each node of the high-risk path and the power distribution network topology information, perform spatial distribution clustering analysis, apply spatial topology aggregation algorithm and thermal distribution mapping to accurately identify the distribution range of the high-risk path on the physical structure of the actual electrical control cabinet, and output spatial distribution characteristic parameters;

[0195] S6.4: Compare the comprehensive risk weight, diffusion speed characteristic parameters and spatial distribution characteristic parameters of high-risk paths with the system's preset hierarchical threshold model, and use a multi-condition composite decision-making mechanism to determine whether the trigger threshold of each level of alarm and protection plan is exceeded, so as to achieve accurate determination of the conditions for group risk cascading or evolution.

[0196] S6.5: Based on the judgment result, according to the trust relationship link level and the corresponding high-risk node, automatically match the graded alarm signal code, and link the protection plan parameter selection module to output the response command including alarm level, target equipment list and protection strategy parameters, and issue multi-level prevention and control suggestions and execution parameters to the electrical control system.

[0197] Step S7: The graded alarm signals and protection plan parameters are sent to the electrical control system. Based on the physical structure, communication capabilities, and network distribution of different cabinets, control actions such as partial circuit breaking, network-wide load reduction, designated equipment isolation, and priority monitoring are executed. Execution feedback is recorded after each control action, specifically including:

[0198] S7.1: The graded alarm signals and protection plan parameters output by the graded risk assessment module are parsed and processed to extract the safety protection level, control strategy type and execution priority queue set for specific electrical control cabinets and their nodes, and generate a graded control instruction set as the input basis for subsequent control actions.

[0199] S7.2: Based on the hierarchical control instruction set, the physical structure parameters, internal circuit distribution information and communication capability tags of each electrical control cabinet are dynamically matched and processed to select the optimal control path and execution terminal. Through industrial Ethernet, CAN bus or wireless communication protocol, the hierarchical control instruction set is accurately routed to the target cabinet or its subordinate sub-control unit to achieve adaptive distribution.

[0200] S7.3: The hierarchical control instruction set issued to the target electrical control cabinet drives the actuators such as power circuit breakers, thermal relays, and intelligent contactors to perform specific electrical protection control actions such as partial circuit breaking, segmented load reduction, area isolation, equipment shutdown and start-up, and priority monitoring according to the strategy, ensuring that the physical links of the control actions completely correspond to the protection requirements;

[0201] Using S7.3 as the implementation target, the input condition is a hierarchical control instruction set that has been sent to the target electrical control cabinet and its subordinate sub-control units via industrial Ethernet, CAN bus or wireless communication protocol, including the control strategy type, safety protection level and priority queue set for each node;

[0202] The instruction frame decoding and task assignment method (parameters: instruction set format, node identification code, task assignment rules) is adopted to parse each control instruction frame and map it to the corresponding actuator control interface inside the cabinet according to the set assignment rules.

[0203] Furthermore, by using the actuator driving algorithm (parameters: action type code, execution permission flag, loop address information), the physical actuators such as power circuit breakers, thermal relays, and intelligent contactors inside the cabinet are driven to enter standby, action, or self-test states respectively, so as to achieve precise execution of hierarchical control actions.

[0204] A control action type decision mechanism (parameters: partial circuit breaking, segmented load reduction, area isolation, equipment shutdown / start, priority monitoring, etc.) is adopted. Combined with the risk level of each node and the physical structure of the cabinet, the corresponding execution mode is automatically mapped: such as performing partial circuit breaking operation on high-risk circuits, performing segmented load reduction on overloaded sections, implementing physical isolation and equipment shutdown / start for specific areas or equipment, and assigning priority monitoring tasks to boundary nodes.

[0205] Using the motion status feedback acquisition module (parameters: execution feedback interface, motion latch timing, and abnormal self-check code), the status changes, latching flags, and abnormal self-check results of each physical actuator during the motion process are collected in real time to form fine-grained motion feedback data.

[0206] A consistency verification algorithm for control actions and protection requirements is adopted (parameters: target control strategy, protection level criterion, actuator type). The algorithm compares the consistency of the actual executed actions and their feedback information with the issued task instructions and set protection targets. For inconsistent nodes, abnormal alarm flags are automatically generated and the consistency verification results are output.

[0207] Through the above chain derivation based on multi-level task assignment-driving-action execution-feedback collection-consistency verification, the logical content of the hierarchical control instruction set is efficiently transformed into the protection control actions and complete feedback at the physical level of the cabinet, realizing a closed loop of the entire execution process and ensuring that the physical links of the control actions completely correspond to the protection requirements.

[0208] For example, in a centralized power distribution scenario in a high-rise building, the target electrical control cabinet is numbered CB-12. The issued hierarchical control command set includes: partial circuit breaking of circuit 1 (high-risk node), segmented load reduction of circuits 2-4 (overloaded sub-sections), and priority monitoring of circuit 5 (boundary node). Control commands are transmitted via the CAN bus. After full-frame parsing and matching, the circuit breaking command is mapped to the power circuit breaker QF1, the load reduction command is mapped to thermal relays RT2-RT4, and the priority monitoring command is sent to the intelligent contactor IC5 in enhanced sampling mode. After the actions are executed, each actuator returns an action completion signal, an action latch code, and an abnormal self-check code. A consistency verification module compares the action feedback with the original command, determining that the QF1 circuit breaking action is completely consistent with the target, and the RT2-RT4 load reduction process conforms to the set power curve. However, RT3 returns an abnormal self-check code (self-check error code E103), immediately generating a local abnormality warning flag. Finally, the execution log, action receipt, and abnormal node list for this control action are output. The verification results show that, under the set threshold and instruction queue, the physical operation response time of the control cabinet is less than 100ms, the action correspondence rate reaches 100%, and the anomaly detection success rate is over 98%.

[0209] S7.4: During the execution of various electrical protection control actions, the action status signals, power consumption change data, network communication feedback and abnormal alarm information of key operation nodes are collected and structured in real time to form control action receipts and execution logs, and process-level holographic feedback is realized.

[0210] S7.5: Based on the aggregated execution logs, control action receipts, and network feedback information, automatically verify the execution success rate and actual response time of the hierarchical control instruction set, compare the execution results with the original protection plan parameters, extract abnormal execution nodes and potential failure links, and generate structured execution feedback data packets to provide quantitative basis for the dynamic correction of the subsequent equipment trust map and the adaptive optimization of the risk model.

[0211] Step S8: Based on the executed control actions and the subsequently continuously collected feature data, periodically perform model correction on the device trust map and abnormal resonance identification parameters, dynamically correct the risk relationships between device nodes, optimize the risk capture sensitivity and prediction accuracy for the next cycle, and achieve adaptive model evolution, specifically including:

[0212] S8.1: Standardize the collection and archiving of hierarchical control actions (including partial circuit breaking, network-wide load reduction, and isolation of designated equipment) executed by the intelligent control system and the resulting equipment execution feedback data to obtain action execution characteristic data that reflects the effect of risk response and achieve data basis consistency for subsequent model correction.

[0213] S8.2: Based on the action execution feature data and the continuously collected sensor normalized feature data, analyze the temporal impact of intelligent control actions on abnormal resonance feature parameters, and use the causal effect analysis method to quantify the changing trend of abnormal resonance feature parameters of the equipment before and after the control action, so as to extract the causal response features of risk propagation under the action of control actions.

[0214] S8.3: Using causal response features as input, the risk transmission weights and direction parameters in the device trust graph are dynamically corrected through adaptive optimization algorithms (such as Bayesian update and self-attention mechanism), and the device node and edge weight representations are updated to achieve periodic optimization and local adaptive reconstruction of the risk relationship network topology.

[0215] Using action execution feature data and normalized sensor feature data as input, causal response features of abnormal resonance parameter changes caused by control actions are extracted as the basic dataset for adaptive correction in this stage.

[0216] A Bayesian update method (parameters: initial risk transmission probability matrix, posterior causal response distribution, historical node attributes) is employed to dynamically adjust the risk transmission weights between nodes in the device trust graph. For each risk relationship edge, the formula is used:

[0217]

[0218] in, Let P(E) be the posterior risk weight from node i to node j. ij |D) represents the causal response probability based on the observed data D. The prior risk weights for the previous period are used, and the denominator is normalized.

[0219] Furthermore, a self-attention mechanism (parameters: node embedding vector, local edge attributes, and anomaly response activation coefficients) is used to deeply weight the modified risk weights and directions, giving high-variability nodes and links with significant resonance higher dynamic sensitivity. The self-attention score is based on the following:

[0220] a ij =softmax(LeakyReLU(W a [x i ||x j ]))

[0221] Among them, a ij Let x be the attention score from node i to j. i x j Let W be the feature vector of two nodes. a For learning weights, || indicates concatenation;

[0222] Furthermore, by using a feature fusion scheme (parameters: historical risk activation records, adaptive normalization coefficients), combined with historical abnormal links and the latest feature factors, the synchronous update of node attributes and edge weights is achieved, and a periodically corrected device trust graph is output.

[0223] By using a dynamic topology reconstruction method, based on changes in risk directionality and edge weight threshold strategies, low-activity links are automatically eliminated and newly generated high-risk paths are added, thereby achieving periodic optimization of the risk relationship network and adaptive adjustment of its local structure.

[0224] Through the above adaptive optimization link, an updated device trust graph is output, enabling a refined dynamic representation of the risk transmission relationship, direction, and intensity indicators between devices.

[0225] For example, in a smart control scenario for a high-rise industrial park power distribution network, the action execution characteristic data includes the circuit current reduction (63A→3A) after circuit breaker QF1 performs a circuit breaker operation, the load reduction feedback of thermal relay RT3, and the signal strength change of IC5 monitoring data (layout node CB-12-5). Comparing before and after control, the causal response characteristics between CB-12 and its neighbor CB-13 change significantly. Before the Bayesian update, the risk weight of CB-12→CB-13 was 0.72. After the circuit breaker operation, based on the observed data D, its posterior probability is corrected to 0.43. The self-attention mechanism is used to calculate the embedded features of the risk link from CB-12 / C-5 to CB-13. After aggregation, a_{CB-12,CB-13} = 0.72. Further feature fusion identifies a new risk link CB-15→CB-12. The weight before correction is 0.1, and the subsequent activation enhances it to 0.57. The final revised trust graph showed significant convergence compared to the previous cycle, with high-weight paths in the risk topology concentrated in the CB-12—CB-15—CB-13 subnets. In actual deployment, the revised trust graph enabled accurate capture of downstream high-risk nodes, improving the accuracy of risk diffusion tracking by more than 8%, and reducing the dynamic risk report response time to less than 45ms.

[0226] S8.4: Apply the updated device trust map with the applied map parameters, combine the latest abnormal resonance features and historical risk event labels, use the sliding window mechanism to periodically adjust the sensitivity of the group risk capture model, optimize the prediction accuracy of risk diffusion paths and high-risk node clusters, and output a dynamically corrected risk capture sensitivity model.

[0227] S8.5: Periodically evaluate the calibration quality and prediction results of the equipment trust spectrum, abnormal resonance parameters, and capture sensitivity model. Based on the evaluation feedback, automatically adjust the model hyperparameters to achieve stable convergence and beam robustness improvement of the risk modeling and prediction module, ultimately supporting the iterative upgrade of the risk adaptive evolution capability of the entire fireproof electrical control cabinet.

[0228] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0229] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for operating an intelligent fireproof electrical control cabinet, characterized in that, Includes the following steps: S1: Collect multi-source monitoring data from various types of sensors in electrical control cabinets within complex power distribution networks, record the physical location, cabinet structure, and network topology label of each sensor sampling point, and obtain multi-source, multi-dimensional monitoring raw data, specifically including: Equipment node identification is performed on all electrical control cabinets in the preset power distribution network topology. Based on the structured network topology model, a list of electrical control cabinet numbers to be collected is generated, and a multi-device distributed collection object list is constructed. For each electrical control cabinet node, the bus control module schedules in parallel to drive the various types of sensor nodes deployed in the cabinet to enter the data acquisition working state. Acquire raw signal streams from multiple types of sensors, including electrical parameters, thermal parameters, vibration parameters, communication signals, and historical fault information, and attach a unique physical location tag to each data stream based on the sensor ID and sampling timestamp; The cabinet structure recognition algorithm is used to perform structure mapping on the sensor output data, and the original monitoring data is accurately bound to the corresponding physical unit, sub-distribution power supply module and bus section code of the electrical control cabinet to generate equipment structure tags; Combining distributed network communication protocols, the monitoring raw data packets, which have been assigned physical location and cabinet structure tags, are transmitted to the central data acquisition server via multicast, and each packet is identified by a network topology number. All raw monitoring data streams stored in the central data acquisition server are indexed and integrated at multiple levels based on physical location tags, cabinet structure tags, and network topology tags to establish a multi-source, multi-dimensional raw monitoring data database with node-time-space mapping relationships. S2: Normalize the raw data from the multi-source and multi-dimensional monitoring, and perform preprocessing operations based on the physical location of each sensor, cabinet structure, and different environmental conditions to generate a feature dataset; S3: Input the feature dataset into the abnormal resonance identification algorithm module to extract the abnormal resonance feature patterns between each cabinet and its associated equipment, and obtain the initial set of equipment clusters with synchronous evolution and cascade evolution, specifically including: The various time-series feature parameters in the uniformly calibrated feature dataset are grouped according to the electrical control cabinet and network topology labels. Based on the multi-dimensional time series structure, the basic time-series attributes of the input objects are extracted to obtain the time-series feature matrix of the cabinet and equipment to be analyzed. Based on the aforementioned time-series feature matrix, the correlation between the main sequence and response sequence of each pair of device nodes is calculated using a time-series cross-correlation analysis algorithm to obtain the cross-correlation coefficient matrix between devices. Using the cross-correlation coefficient matrix as input, the causal inference method is used to further determine the temporal driving relationship of abnormal fluctuations among devices, and output the causal index of abnormal driving among device nodes. Based on the cross-correlation coefficient matrix and the anomaly-driven causality index, a fast segmented clustering algorithm is applied to cluster and classify the device nodes. Device nodes that evolve synchronously or have time cascade effects are divided into anomaly resonance feature pattern groups, forming a preliminary device cluster division result. Statistical characteristic screening and labeling processing are performed on the abnormal resonance feature pattern groups obtained by clustering. The feature labels of the device node clusters are labeled and structured output to obtain the initial set of synchronous evolution and cascade evolution device clusters. S4: Based on the abnormal resonance feature pattern, combined with the power distribution network topology and the physical attributes of each device, calculate the risk transmission weight and direction between all device nodes to generate an adaptive device trust graph. S5: For the adaptive device trust map, learn and reason about the historical and current feature data within a specified sliding time window to dynamically predict the risk diffusion path, impact range and high-risk node clustering area among the electrical control cabinet equipment groups; S6: Determine whether the risk weight, diffusion speed and spatial distribution of high-risk paths in the graph reasoning results exceed the set threshold. If the group risk cascading or evolution triggering conditions are met, determine and output the graded alarm signal and corresponding protection plan parameters based on the trust relationship link level. S7: Send the graded alarm signals and the protection plan parameters to the electrical control system, and execute control actions according to the physical structure, communication capabilities and network distribution of different cabinets, and record the execution feedback.

2. The method for operating an intelligent fireproof electrical control cabinet according to claim 1, characterized in that, Following step S7, the following is also included: S8: Based on the executed control actions and the feature data collected subsequently, the model is periodically corrected for the device trust map and abnormal resonance identification parameters to dynamically correct the risk relationship between device nodes.

3. The method for operating an intelligent fireproof electrical control cabinet according to claim 1, characterized in that, The various types of sensors include current sensors, voltage sensors, temperature sensors, humidity sensors, vibration sensors, and communication status monitoring modules.

4. The method for operating an intelligent fireproof electrical control cabinet according to claim 1, characterized in that, Step S2 specifically includes: The raw data from multi-source and multi-dimensional monitoring are categorized and indexed based on the physical location of the sensors, the cabinet structure, and the sampling timestamp. A time-series synchronization labeling algorithm is used to perform network-wide time-domain alignment processing on each set of sampled data to generate spatiotemporally aligned data blocks with unified reference time and spatial labels. Using the spatiotemporally aligned data block as input, a multidimensional wavelet denoising algorithm and adaptive median filtering technology are employed to perform noise suppression and anomaly removal processing on the sensor signals of each physical unit under different cabinet structures and environmental conditions. For data blocks that have completed noise suppression, according to the sensor technical parameter table and the cabinet physical structure association information, the unit standardization algorithm and dynamic interval linear stretching are implemented to uniformly process various electrical parameters, thermal parameters and vibration parameters according to the set statistical intervals and physical magnitudes, so as to obtain dimensional consistency characteristic data across equipment and cabinets; Based on the scale consistency feature data, normalized residual compensation is performed using environmental attribute factors. For the feature dataset that has undergone multi-level standardization and normalization, feature vector groups are constructed based on device network topology labels and historical operating status. Through a dynamic feature aggregation algorithm, the output results of multiple stages such as spatiotemporal alignment, noise cancellation, scale consistency, and environmental compensation are integrated into a uniformly calibrated feature dataset.

5. The method for operating an intelligent fireproof electrical control cabinet according to claim 1, characterized in that, The cross-correlation analysis and the causal inference are used to identify the synchronous evolution of abnormal events and the master-slave cascade relationship between device nodes. The segmented clustering includes at least spectral clustering or density peak clustering for device cluster division.

6. The method for operating an intelligent fireproof electrical control cabinet according to claim 1, characterized in that, Step S4 specifically includes: Multidimensional data fusion is performed on the abnormal resonance feature patterns output from the abnormal resonance feature extraction submodule, as well as the power distribution network topology parameters and equipment physical attribute parameters, to construct a multidimensional feature vector group and obtain a high-dimensional risk correlation input matrix. Based on the high-dimensional risk association input matrix, the conditional probability inference algorithm is applied to initialize the potential risk transmission relationship between each device node. Based on the abnormal resonance characteristics and prior network and physical attributes, an initial risk transmission probability matrix is ​​generated. For the initial risk transmission probability matrix, a Bayesian update mechanism is adopted, which combines newly received monitoring data, historical resonance evolution priors and node attribute parameters to correct the risk transmission weights between nodes and output a dynamic risk transmission weight matrix. By using an attention mechanism, the feature vector group after fusing the dynamic risk transmission weight matrix with the device node attributes is deeply weighted to obtain a weighted structured risk edge weight feature set. Using the weighted structured risk edge weight feature set as input, and based on the three elements of electrical control cabinet equipment nodes, risk relationship edge weights, and node risk labels, an adaptive equipment trust graph containing dynamic edge weights and node labels is generated through a graph-based data modeling method.

7. The method for operating an intelligent fireproof electrical control cabinet according to claim 6, characterized in that, Each row vector in the risk association input matrix includes at least the anomaly label, response intensity, spatial adjacency, and physical attribute normalization index of the target node; the initial risk transmission probability is dynamically corrected by Bayesian iteration combined with new monitoring data and prior historical anomaly events.

8. The method for operating an intelligent fireproof electrical control cabinet according to claim 1, characterized in that, Step S5 specifically includes: Based on the adaptive device trust graph generated in step S4, historical feature data and current feature data within a specified sliding time window are obtained to form a feature vector group based on node-edge attributes. The node attribute vectors and edge weight relationships of the adaptive device trust graph are aggregated using a graph convolutional neural network or a graph attention network to calculate the spatial aggregated distribution of risk information among devices. After spatial aggregation is completed using graph neural network inference, the risk evolution trend of each node is modeled by time series learning based on the historical and current feature data sequences within the sliding time window to obtain the risk diffusion path between device nodes. The risk impact range of the risk diffusion path results output by the model is determined, and the high-risk node cluster area is calculated based on the node risk weight and the adjacency topology relationship. Based on the risk diffusion path and the identification results of the high-risk node cluster area, a dynamic risk prediction report is generated among the equipment groups, and the risk association link details, key attention node list and predicted risk level are output.

Images