Switch cabinet safety monitoring and control system based on artificial intelligence

Abstract

The invention relates to the technical field of intelligent power distribution, in particular to a switch cabinet safety monitoring and control system based on artificial intelligence, and the system comprises an operation situation sensing module which collects the multi-dimensional operation data of a switch cabinet in real time, and carries out the calculation to generate a dynamic correlation intensity matrix; the abnormal traceability and prediction module is used for carrying out time sequence analysis and mode recognition on the dynamic correlation intensity matrix, outputting traceability positioning information of potential faults and predicting risk evolution; the control strategy generation module is used for generating a targeted adaptive control instruction according to the traceability positioning information and risk evolution prediction; and the execution and feedback module is used for executing the self-adaptive control instruction to adjust the operation state of the switch cabinet, collecting state feedback data after the instruction is executed and sending the state feedback data back to the operation situation sensing module. According to the method, a complete closed-loop control chain from sensing, decision making to verification is realized, and the long-term adaptability and the system evolution capability of a control strategy are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent power distribution technology, and in particular to a switchgear safety monitoring and control system based on artificial intelligence. Background Technology

[0002] With the continuous development of distribution automation systems, switchgear, as one of the key primary devices, is receiving increasing attention for its safe operation in power systems. Existing switchgear monitoring methods mostly focus on static parameter detection, such as single-point threshold alarms for temperature, current, and voltage, lacking in-depth analysis of the coupling relationships between multi-dimensional operational data. Especially when facing nonlinear evolution characteristics or potential multi-factor linkage faults, insufficient sensing capabilities often lead to delayed early warnings and inaccurate location. At the same time, existing systems generally lack the ability to predict the evolution trend of fault modes, resulting in the inability to intervene in potential risks in advance, affecting the overall operational safety and maintenance efficiency of the power system.

[0003] Furthermore, most traditional control systems employ rule-based setpoint strategies, which cannot adaptively adjust to real-time changes in the operating status and are ill-equipped to handle dynamic risk evolution under complex operating conditions. In particular, after an abnormal state occurs, existing control strategies lack an execution feedback mechanism, making it impossible to assess the effectiveness of intervention or further optimize the strategy library and control model, thus restricting the closed-loop response and self-evolution capabilities of the entire system. Therefore, there is an urgent need for an intelligent monitoring and control method that integrates multi-source data analysis, risk prediction, adaptive decision-making, and feedback optimization to improve the operational safety and autonomous decision-making capabilities of switchgear under multi-dimensional operating conditions. Summary of the Invention

[0004] This invention provides a switchgear safety monitoring and control system based on artificial intelligence.

[0005] An artificial intelligence-based switchgear safety monitoring and control system includes the following modules: Operational situation awareness module: collects multi-dimensional operational data of the switchgear in real time and calculates and generates a dynamic correlation strength matrix, which is used to characterize the coupling relationship between various operational data. Anomaly tracing and prediction module: performs time series analysis and pattern recognition on the dynamic correlation strength matrix, and outputs source location information and risk evolution prediction for potential faults; Control strategy generation module: Based on the source tracing and location information and risk evolution prediction, it generates targeted adaptive control commands; Execution and Feedback Module: Executes the adaptive control commands to adjust the operating status of the switchgear, and collects the status feedback data after the command execution and sends it back to the operating status perception module to form a closed-loop optimization.

[0006] Optionally, the operational situation awareness module includes: Data acquisition unit: Simultaneously collects multi-dimensional operational data through temperature sensor groups, current sensors, voltage sensors, vibration sensors, and ultrasonic sensors deployed in the switch cabinet; Association Calculation Unit: Within a set sliding time window, it calculates the association strength index between any two different dimensions of runtime data time series using the mutual information method; Construction of dynamic correlation strength matrix: The correlation strength indices are combined into a symmetric matrix to construct a dynamic correlation strength matrix.

[0007] Optionally, the association calculation unit includes: Time series discretization: Within a set sliding time window, the time series of any two types of data in the multidimensional operation data of the switchgear are discretized respectively, and the corresponding continuous numerical intervals are divided into several discrete intervals. Joint frequency statistics: Based on the discrete intervals, perform combined statistics on all sample points in the two types of time series to determine the joint frequency of data points in each set of discrete interval combinations; Probability density construction: Based on the joint frequency of occurrence, the joint probability density function of the two types of data is calculated, and the marginal probability density functions of the two types of data are obtained by summing the joint probability along the univariate direction. Association strength calculation: Based on the degree of information deviation between the joint probability density function and the marginal probability density function, the statistical correlation between the two types of operational data is quantified, and finally the association strength index value between the two types of operational data is obtained, and a dynamic association strength matrix is ​​constructed.

[0008] Optionally, the anomaly tracing and prediction module includes: Graph structure construction unit: maps the dynamic association strength matrix into a dynamic attribute graph, where each node in the dynamic attribute graph represents one dimension of operational data, and the edge weights between nodes are defined by the corresponding association strength index; Temporal pattern recognition unit: The sequence of the dynamic attribute graph in continuous time steps is learned by a spatiotemporal graph neural network to extract the evolution rules of nodes and edge weights in the graph; Source tracing and localization submodule: By calculating the graph attention weights of the spatiotemporal graph neural network, the key nodes that contribute the most to the current graph structure anomaly are identified, and the set of key nodes is the source tracing and localization information; The risk prediction submodule, connected to the time-series pattern recognition unit, is used to continue inputting the sequence of the dynamic attribute graph into the time-series prediction model to predict the evolution trajectory of the dynamic correlation strength matrix over a future period of time, and to define the part of the trajectory that exceeds the upper limit threshold of normal evolution as risk evolution prediction.

[0009] Optionally, the source tracing and location submodule includes: Abnormal node weight calculation: Based on the spatiotemporal learning results of graph neural networks, obtain the attention weight of the corresponding node for each running dimension; Key node screening: Nodes are sorted according to their attention weights, and key nodes whose attention weights exceed the contribution weight threshold are selected as the source localization results for abnormal states.

[0010] Optionally, the risk prediction submodule includes: Graph Modeling and Temporal Learning: Based on dynamic attribute graphs, we use graph neural networks to extract the structural relationships between nodes and combine them with recurrent networks to model the evolution of the graph over time. Association prediction and risk identification: Based on the learning results of the dynamic attribute graph, predict the future changes in the association strength between various dimensions, and identify potential risk relationships by comparing them with the preset normal evolution upper limit threshold, and construct a set of risk paths.

[0011] Optionally, the control strategy generation module includes: Strategy Knowledge Base: It contains multiple control strategy templates. Each template defines an anomaly pattern-control action mapping relationship. The anomaly pattern is jointly characterized by the combination of key node types in the source tracing and location information and the risk trend in the risk evolution prediction. Strategy matching engine mechanism: Receives the source location information and risk evolution prediction, and matches the most suitable control strategy template from the strategy knowledge base based on a hybrid decision-making algorithm of rule reasoning and case matching; Instruction generation unit: Instantiates the matched control strategy template into one or more adaptive control instructions that can be directly parsed by the underlying actuator.

[0012] Optionally, the hybrid decision algorithm includes: Rule-based reasoning and matching: Based on the preset abnormal patterns and control strategies mapping rules, determine whether the combination of node types in the current source tracing and location information and the risk trend in the risk evolution prediction match the matching conditions of a certain strategy template in the strategy knowledge base, and generate a rule-based reasoning score. Case similarity matching: Based on the results of rule reasoning, the feature similarity between the current input and the historical application cases corresponding to each template in the strategy knowledge base is further compared, and the case matching score is calculated based on the degree of matching between node combination and risk trend trajectory.

[0013] Optionally, the execution and feedback module includes: Command distribution and execution unit: Receives the adaptive control command, parses it and sends it to the corresponding underlying physical actuator to adjust the operating status of the switch cabinet; Feedback data acquisition and new correlation matrix construction unit: Within a preset effect observation window after the execution of the adaptive control command, multi-dimensional operation data of the switch cabinet is collected again, and a new dynamic correlation strength matrix is ​​generated by comparing the multi-dimensional operation data in the effect observation window with that before the command execution. Effectiveness Quantification Evaluation Unit: Generates a quantitative evaluation value of the intervention effect by comparing the degree of reduction of abnormal correlation modalities before and after instruction execution; Closed-loop optimization signal generation unit: The intervention effect evaluation value, the corresponding source location information, risk evolution prediction and the executed adaptive control command are encapsulated into a reinforcement learning sample, and the reinforcement learning sample is fed back to the operation situation perception module and the control strategy generation module.

[0014] Optionally, the underlying physical actuator may include a controllable circuit breaker, a smart load switch, a variable frequency fan, or a partial discharge suppressor.

[0015] The beneficial effects of this invention are: This invention achieves full-link dynamic monitoring and control of switchgear operation status through a multi-module linkage structure of operational situation awareness, anomaly tracing and prediction, strategy generation, and execution feedback. It employs mutual information calculation to construct a dynamic correlation strength matrix, considering not only the nonlinear coupling relationships between various operational data dimensions but also introducing a spatiotemporal graph neural network to model the evolution of the graph structure over multiple time steps, thus enhancing the ability to identify potential risks in advance. Furthermore, it combines graph attention mechanisms to extract key nodes for source tracing and location, and formulates control strategies based on risk evolution trends, demonstrating foresight, pertinence, and adaptability, thereby enhancing the operational safety and anomaly response efficiency of the switchgear under complex operating conditions.

[0016] This invention not only executes adaptive control strategies to adjust the operating state of switchgear, but also designs a quantitative evaluation mechanism for intervention effects. By dynamically comparing the number of abnormal modes before and after control, an interpretable intervention scoring function is constructed. The scoring results are then combined with information such as execution instructions, risk trends, and key nodes to form reinforcement learning samples, which are fed back to the front-end module for iterative optimization of model parameters and strategy knowledge base. This achieves a complete closed-loop control chain from perception and decision-making to verification, effectively improving the long-term adaptability of the control strategy and the system's evolutionary capability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Fig. 1 This is a system block diagram of an embodiment of the present invention; Fig. 2 This is a logical framework diagram of an embodiment of the present invention. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Those skilled in the art may employ other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0020] like Figs. 1-2 As shown, an artificial intelligence-based switchgear safety monitoring and control system includes the following modules: Operational Situation Awareness Module: Collects multi-dimensional operational data of the switchgear in real time and calculates and generates a dynamic correlation strength matrix, which is used to characterize the coupling relationship between various operational data. The operational situation awareness module specifically includes: Data acquisition unit: Used to synchronously collect multi-dimensional operational data through various types of sensors deployed inside the switch cabinet, specifically including: Temperature sensor (used to monitor temperature rise at critical nodes and busbars); Current sensor (used to monitor the load current of the main circuit and branch circuits). Voltage sensor (used to monitor bus and load voltage); Vibration sensors (used to monitor abnormal vibrations caused by loose mechanical structures or connections); Ultrasonic sensors (used to monitor high-frequency characteristic signals such as partial discharge and arcing); Data collected by various sensors is uniformly encoded into multidimensional time series data, represented as follows: ; in, for A multidimensional set of runtime data at any given time. For the first Sampled values ​​of various operational status data (temperature, current, voltage, vibration, ultrasound). The total number of dimensions of the running data. The corresponding number of sensor types.

[0021] Correlation Calculation Unit: The correlation calculation unit is connected to the data acquisition unit and is used to perform calculations within a set sliding time window. Within this framework, the correlation strength index between time series of any two different dimensions of running data is calculated and the output is organized in matrix form. For any two different data dimensions , The degree of nonlinear dependence is quantified using mutual information, and the final output of the mutual information serves as an index of the correlation strength between the two, including: (1) Time series binning (discretization): binning continuous time series , In the sliding time window The intervals within the interval are discretized as follows: ; The range of values ​​is divided into categories using either a fixed width or a fixed number of categories. A discrete interval, represented as: ; ; in, The sliding time window length ranges from 30 to 60 seconds. Too short a window will result in insufficient statistics, while too long a window will mask short-term abnormal trends. The sliding time window is chosen instead of global calculation to respond promptly to changes in the operating state and improve the system's sensitivity and adaptability. The number of boxes is a fixed value. Too few bins result in insufficient resolution, while too many bins dilute the accuracy of probability estimation. Represents a discrete interval; (2) Calculate the joint probability density function: in the sliding window Internal Statistics Drop into each pair of boxes joint frequency The joint probability density function is obtained using frequency estimation, and is expressed as: ; in, They represent Discretized , No. Each interval This indicates that the samples fell into the box simultaneously. and The number of times, the range of values ​​is , This represents the total number of sampling points within the window, ranging from 100 to 6000 points, corresponding to a sampling rate of 10-200 Hz. Sufficient sampling points ensure the accuracy of probability estimation; a higher sampling frequency... The larger, Let be the joint probability density function, i.e. and Joint probability density value; (3) Calculate the marginal probability density function: The marginal probability is obtained by summing the joint probabilities, and is expressed as: ; ; in, Represents the marginal probability density function; (4) Calculate mutual information based on joint probability and marginal probability , represented as: ; in, Indicates the first With the Mutual information between data types, if The corresponding term is conventionally considered to be 0. The logarithmic base can be the natural logarithm. The output unit is nat. If it is 2, the unit is bit. The larger the value, the stronger the correlation between the two variables. 0 indicates independence.

[0022] (5) Mutual information The output has the following properties: ; The larger the value, the stronger the coupling relationship between the two types of operational data; It can capture both linear and nonlinear correlations; Therefore, the final correlation strength index is directly taken as .

[0023] In the monitoring of switchgear operation status, multiple sensors (temperature, current, voltage, vibration, and ultrasound) simultaneously collect time-series data of different dimensions. Since there are often potential coupling relationships between these operational data, a method capable of simultaneously capturing linear and nonlinear correlations is necessary to identify the strength of this coupling. This step employs the mutual information method as a means of calculating the correlation strength index between multidimensional operational data. First, the time series of any two different dimensions are binned and discretized, mapping continuous data into a finite number of intervals for subsequent probability estimation. Then, the joint frequency of these two series in various interval combinations is statistically analyzed to construct a joint probability density function. Next, the marginal probability density functions of each variable are obtained by marginally summing the joint probabilities. After obtaining the joint and marginal probabilities, the information coupling strength between the pair of variables is calculated using the standard mutual information formula. This mutual information value is defined as the correlation strength index between the two types of operational data, constituting an element of the dynamic correlation strength matrix. Mutual information can capture any correlation pattern, whether linear, nonlinear, monotonic, or non-monotonic, and is suitable for complex disturbance patterns in the operating state of electrical equipment. It does not require the data distribution to follow a Gaussian or other prior model, so it is suitable for handling strong disturbances, abrupt changes, and non-stationary data sequences that occur during the operation of switchgear. Mutual information supports sliding window evolution analysis. By dynamically updating the mutual information matrix through a sliding time window, it is possible to track the temporal evolution of correlation relationships, which helps to identify the risk evolution trend caused by abrupt changes in coupling relationships in advance.

[0024] Construction of dynamic correlation strength matrix: This involves combining the above correlation strength indicators... Combining these into a symmetric matrix form, the dynamic correlation strength matrix is ​​represented as follows: ; in, This is a dynamic correlation strength matrix. It is the first With the The strength of association between data types: diagonal elements of a matrix are 0, indicating no association between the data types; sliding window. The time period used for calculation was controlled.

[0025] Anomaly tracing and prediction module: Performs time series analysis and pattern recognition on the dynamic correlation strength matrix, and outputs source location information and risk evolution prediction for potential faults; The anomaly tracing and prediction module specifically includes: Graph structure building unit: dynamic correlation strength matrix Mapped to a dynamic property graph ; in, The set of nodes in a graph, each node This represents one dimension of operational data. , For time The set of edges at time, , The edge weight matrix is ​​derived from the dynamic correlation strength matrix, i.e. Each edge The weight is , indicating the first With the The data in each dimension are related at time. The strength of the association; Temporal pattern recognition unit: This unit identifies graph sequences over consecutive time steps. The input is fed into a spatiotemporal graph neural network for modeling and learning, extracting the spatiotemporal dependency patterns of node states and edge weight evolution in the graph sequence, with each node at time... The input features are represented as follows: ; in, For nodes At any moment The feature vector is constructed from the corresponding original data, including mean, standard deviation, rate of change, maximum value, etc. within the window, and includes 4-32 dimensions. It extracts representative information of the data running in the current dimension as the basis for node representation. The dimension should not be too high to avoid overfitting and computational overhead.

[0026] Spatiotemporal graph neural networks can simultaneously capture the spatial dependencies of graph structures and the dynamic changes of time series, and include two modules: (1) Spatial modeling: at each time step Graph convolution is used to aggregate features of graph structures. Each node not only considers its own features, but also integrates the feature information of its connected nodes. During the fusion, the features are weighted according to the edge weights (i.e., the dynamic association strength) to reflect the degree of coupling in each dimension. The output is the updated representation of each node at this time step. (2) Time modeling: The representation of each node at all time steps is sequentially input into the time modeling network. The Transformer structure is used to capture the evolution trend of the node state over time and output the comprehensive dynamic feature representation of each node within the current time window.

[0027] Source tracing and localization submodule: Based on the graph attention mechanism output by the spatiotemporal graph neural network, it constructs a node contribution weight vector. This represents the degree of influence of each node under the current abnormal state. The node contribution weight is calculated by the graph attention network and is expressed as: ; in, For attention weight vectors, The feature transformation weight matrix, with hidden layer dimensions ranging from 8 to 64, is used to map input features to the attention space. Larger hidden layer dimensions capture finer features but increase model complexity. This represents a vector concatenation operation. Larger values ​​indicate a larger node The higher the contribution to the current abnormal situation, This is an activation function with a negative slope, where the negative slope parameter ranges from 0.01 to 0.2. It maintains its responsiveness to small negative inputs and helps capture edges that are weakly correlated but exhibit strong abrupt changes. It is a standardization function that maps all attention scores to a probability distribution that sums to 1; Based on all nodes Sort by weight, selecting those with a weight higher than the contribution weight threshold. A set of several nodes is defined as the anomaly source localization result at the current moment, i.e., the source localization node set, and is represented as: ; in, For a moment The set of traceability and location nodes, The contribution weight threshold, ranging from 0.1 to 0.2, can effectively filter out non-critical nodes and retain only nodes that significantly contribute to changes in the overall graph structure as root causes of anomalies.

[0028] Risk prediction submodule: This module generates dynamic attribute graph sequences. Continue inputting into the time series forecasting model to predict future time periods. The sequence of dynamic correlation strength matrices within Any pair of dimensions in the prediction result correlation strength value Compared with the set upper limit threshold of normal evolution The comparisons include: (1) Based on the constructed dynamic attribute graph , to obtain the graph sequence input ; in, This refers to the length of the historical time window, ranging from 20 to 200 time steps. Too short a time step makes it difficult to capture trends, while too long a time step leads to latency and computational burden. The step size is predicted, and the value ranges from 1 to 20.

[0029] (2) at each moment , node features Adjacency Matrix The input is fed into a graph convolutional network to extract the structure-aware representation of each node: ; in, This is the node embedding representation of the graph convolution output. The output dimension of the graph convolution has a value range of 8-64. Smaller values ​​have weak expressive power, while larger values ​​affect real-time performance. (3) Output the graph convolution at all time points As a time series input, it is fed into the GRU network for dynamic modeling, represented as: ; in, It is the time of each node. The state representation of GRU cells allows them to capture the nonlinear temporal evolution trajectory of node states, supporting both short-term and long-term memory. This is the hidden state from the previous time step.

[0030] (4) Use the GRU state of the last time step Using the current input, predict future time periods. The graph structure (i.e., the correlation matrix) for each future time step The predicted edge weight matrix is ​​generated using a linear mapping layer, and is expressed as: ; in, The predicted association strength matrix represents the different dimensions at time [time value missing]. The predicted coupling relationship.

[0031] (5) For each matrix obtained from the prediction Take any one of the elements Compared with the set upper limit threshold of normal evolution Comparison, represented as: The prediction suggests a risk of coupling anomalies. The correlations that exceed the upper limit of normal evolution in the prediction results are filtered out to form a set of potential risk paths in the future.

[0032] in, The value ranges from 0 to 1, representing the predicted future association strength. The upper limit threshold for normal evolution is 0.7-0.85. Too low a value will lead to false alarms, and too high a value will lead to false negatives.

[0033] Control strategy generation module: Generates targeted adaptive control commands based on source tracing and location information and risk evolution prediction; The control strategy generation module specifically includes: Strategy Knowledge Base: A control strategy knowledge base is pre-stored, denoted as... Each strategy template Defined as a set of exception mode-control action mapping relationships, expressed as: ; in, This serves as a strategy knowledge base, including 10-100 templates to ensure coverage of various abnormal combination scenarios. The number of templates should not be too many to avoid reducing matching efficiency, nor too few to prevent insufficient coverage. For the first A strategy template, This represents the total number of strategy templates, ranging from 10 to 100. This indicates abnormal pattern characteristics, which originate from combinations of key node types (such as temperature + vibration). This indicates the type of risk evolution trend (e.g., rapid rise, slow fluctuation), and describes the changing pattern of risk indicators over time. This indicates the corresponding control action template, including the control target, control quantity, adjustment method, etc.

[0034] The strategy matching engine mechanism automatically selects the most suitable control strategy template from a preset control strategy knowledge base based on the abnormal characteristics identified in the current operating state, including: (1) Matching score calculation: When the key node type combination output from the current source tracing and location and the risk evolution trend output from the risk prediction module are received, the strategy matching engine uses a hybrid decision-making algorithm based on rule reasoning and case matching to calculate the score of the knowledge base. All policy templates in the dataset are matched together, specifically including: 1) Rule-based reasoning matching score Scoring is based on preset mapping rules and is represented as follows: ; in, Indicates the type of risk trend at the current moment. Indicates the current combination of key nodes. A higher value indicates a stronger rule match. It is a rule-matching function based on a preset mapping table. By constructing a preset anomaly pattern-control strategy rule mapping table, it explicitly maps different key node combinations and risk trend types to specific strategy templates. During matching, the currently identified key node type and risk trend are compared item by item with each template definition in the mapping table. The comparison results are divided into three categories based on the degree of matching: If the combination of key nodes and risk trends are completely consistent with a certain template definition, the score is the highest value of 1; If the key node part includes or has similar risk trends, but not exactly the same, then a middle score of 0.5 is given; If there is no identifiable structural correspondence, it is considered a mismatch, and the score is the lowest value of 0.

[0035] 2) Case similarity matching score The score is calculated based on the similarity of key nodes and trends in historical cases, and is expressed as follows: ; in, , The Jaccard similarity method is used to measure the similarity between the current abnormal state and the historical strategy template in terms of node combination. The set of key nodes currently identified is compared with the set of key nodes defined in the template, and the ratio of the size of their intersection and union is calculated as the structural similarity score, which reflects the degree of overlap between the two abnormal patterns in the node dimension. The higher the value, the more similar the structure. This method does not depend on the order and does not distinguish between primary and secondary features, making it suitable for expressing the similarity of multi-dimensional feature combinations. As a weighted average function that integrates structural and trend similarity, structural similarity and trend similarity are assigned weights respectively. The two are summed according to the weight ratio to obtain the final case matching score. This allows for flexible adjustment of the matching logic and improves the overall adaptability. (2) Comprehensive matching score and strategy selection: Based on the obtained rule reasoning matching score and case similarity matching score, the final matching score is calculated, and the strategy template with the highest score is selected, as follows: ; ; in, Indicates the first The total matching score for each template These are weighting coefficients used to balance rule-based reasoning and case matching. , The strategy template with the highest score.

[0036] Instruction generation unit: generates the matched policy template Instantiated as a set of control instructions that can be issued and executed, it is represented as: ; Each control command This includes the controlled object (such as a fan, circuit breaker, alarm, etc.) and the controlled quantity. (Switch status, speed, power, etc.) and control intensity factor , This represents the number of control commands in this round. Control strength Calculated based on the risk trend indicator, it is expressed as: ; in, No. The control strength of each control command The basic control intensity is 30%-80% of the equipment's rated control value. This is an adjustment coefficient, and its value range is... A higher value indicates a more sensitive system, suitable for scenarios with high requirements for responding to mutation risks; a lower value indicates a more stable and conservative control. This is the risk trend change rate, reflecting the severity of the trend. Its value ranges from 0.01 to 1.0, and the larger the value, the faster the risk increases.

[0037] Execution and Feedback Module: Executes adaptive control commands to adjust the operating status of the switchgear, and collects status feedback data after command execution and sends it back to the operating status awareness module to form a closed-loop optimization; The execution and feedback module specifically includes: Instruction distribution and execution unit: a set of adaptive control instructions output by the control strategy generation module. Control commands will be sent to the physical actuator set. , represented as: ; Each of them This refers to a low-level execution component, including a controllable circuit breaker, intelligent load switch, variable frequency fan, or partial discharge suppressor. Control commands are issued and translated via a protocol, and the execution action occurs within a time period. Internal implementation; Feedback data acquisition and new correlation matrix construction unit: a preset observation window after the control command is executed. Within this process, new multidimensional operational data sequences are collected, and the same mutual information method as before execution is used to calculate the dynamic correlation strength matrix for the new time period. Specifically, this includes: 1. Feedback Data Collection: (1) Set observation window: After the control command is executed, set a preset effect observation time window. The value range is 300-900s, covering a sufficient number of state change processes; (2) Multidimensional sensor signal acquisition: The sources of the acquired feedback data include electrical signals, thermal status, partial discharge signals and operating status; (3) Data formatting and unified timing: All acquisition channels are connected to a unified clock synchronization mechanism to ensure that the data has a unified time base. All sensor data are subjected to interpolation alignment and sliding window standardization processing, and the output time series matrix is ​​represented as: ; in, Number of signal channels The number of sampling times. Indicates the first The channel is in The value at time; 2. Construction of a new association matrix: (1) Sliding window slicing: Dividing the time series into multiple windows to obtain A data block, represented as: ; in, This represents the total number of sliding windows.

[0038] (2) Mutual information calculation: For each sliding window data Calculate each pair of channels Mutual information is represented as: ; in, Indicates channel and The joint probability density, , Represents the marginal probability density function. In the first Signal channel in a sliding window and The corresponding mutual information value.

[0039] (3) Smooth integration: The mutual information in all windows is weighted and averaged to obtain the elements of the final association matrix, forming a new dynamic association matrix, represented as: ; ; in, Indicates signal channel and The average mutual information across all sliding windows represents the final dynamic association strength. This is the new dynamic correlation matrix.

[0040] This step employs the mutual information method to quantify the coupling relationships between multi-dimensional operational data of the switchgear. First, the time series of operational data for each dimension are divided within a sliding window to ensure temporal continuity. Then, for any two different types of data series, their marginal probability distributions and joint probability distributions are estimated. Next, based on the mutual information formula, the shared information content is calculated to obtain a numerical correlation strength index. This index can capture non-linear correlations between variables and is suitable for anomaly identification needs in complex industrial scenarios. Finally, all calculation results are summarized to form a dynamic correlation strength matrix.

[0041] Effectiveness Quantification Evaluation Unit: Before control execution, a set of abnormal modal pairs (i.e., node pairs with strong associations exceeding a preset abnormal association strength threshold) is identified based on the dynamic association strength matrix. This is the pre-execution abnormal set, represented as: ; Within the effect observation time window, based on the new dynamic correlation strength matrix, new abnormal modal pairs are identified, i.e., the set of post-execution anomalies, represented as: ; Compare the number of elements in both. If the number of abnormal modalities decreases after implementation, the intervention is considered effective. The intervention effectiveness scoring function is defined as follows: ; in, For nodes The correlation strength before control execution, with a value range of 0-1, represents the data from two dimensions. The strength of the coupling between them For nodes The correlation strength after control execution has a value range of 0-1. For node pairs The threshold for abnormal association strength ranges from 0.3 to 0.7; lower values ​​indicate greater sensitivity, while higher values ​​indicate greater conservatism. To control the set of node pairs that were previously judged as abnormal. To control the set of node pairs that are subsequently deemed abnormal. For the intervention effect score, the closer the value is to 1, the more significantly the abnormal association modality is weakened. If the value is negative or close to 0, it means that the control is ineffective or failed.

[0042] The goal of control commands is to reduce or eliminate anomalous modes. Therefore, by comparing the correlation strength matrix within two time windows before and after control execution, and statistically analyzing whether the number of anomalous modes has decreased, the control effect can be quantified. If the number of anomalous modes decreases significantly after execution, it indicates that the control measures have played an effective role in the system; if there is no change or the number increases, it indicates that the control has failed or that the strategy needs to be adjusted.

[0043] Closed-loop optimization signal generation unit: Encapsulates the following four pieces of information into a reinforcement learning sample, represented as: ; in, To learn reinforcement samples, To locate key node combinations for tracing the source. As the trend of risk evolution, For the set of control instructions to be executed, To score the actual effectiveness of the intervention; Reinforcement learning samples It is used in the situational awareness module for model adaptive update and in the control strategy generation module for optimizing strategy template scoring mechanism and updating matching weights.

[0044] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0045] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A switchgear safety monitoring and control system based on artificial intelligence, characterized in that, Includes the following modules: Operational situation awareness module: collects multi-dimensional operational data of the switchgear in real time and calculates and generates a dynamic correlation strength matrix, which is used to characterize the coupling relationship between various operational data. Anomaly tracing and prediction module: performs time series analysis and pattern recognition on the dynamic correlation strength matrix, and outputs source location information and risk evolution prediction for potential faults; Control strategy generation module: Based on the source tracing and location information and risk evolution prediction, it generates targeted adaptive control commands; Execution and Feedback Module: Executes the adaptive control commands to adjust the operating status of the switchgear, and collects the status feedback data after the command execution and sends it back to the operating status perception module to form a closed-loop optimization.

2. The switchgear safety monitoring and control system based on artificial intelligence according to claim 1, characterized in that, The operational situation awareness module includes: Data acquisition unit: Simultaneously collects multi-dimensional operational data through temperature sensor groups, current sensors, voltage sensors, vibration sensors, and ultrasonic sensors deployed in the switch cabinet; Association Calculation Unit: Within a set sliding time window, it calculates the association strength index between any two different dimensions of runtime data time series using the mutual information method; Construction of dynamic correlation strength matrix: The correlation strength indices are combined into a symmetric matrix to construct a dynamic correlation strength matrix.

3. The switchgear safety monitoring and control system based on artificial intelligence according to claim 2, characterized in that, The association calculation unit includes: Time series discretization: Within a set sliding time window, the time series of any two types of data in the multidimensional operation data of the switchgear are discretized respectively, and the corresponding continuous numerical intervals are divided into several discrete intervals. Joint frequency statistics: Based on the discrete intervals, perform combined statistics on all sample points in the two types of time series to determine the joint frequency of data points in each set of discrete interval combinations; Probability density construction: Based on the joint frequency of occurrence, the joint probability density function of the two types of data is calculated, and the marginal probability density functions of the two types of data are obtained by summing the joint probability along the univariate direction. Association strength calculation: Based on the degree of information deviation between the joint probability density function and the marginal probability density function, the statistical correlation between the two types of operational data is quantified, and finally the association strength index value between the two types of operational data is obtained, and a dynamic association strength matrix is ​​constructed.

4. The switchgear safety monitoring and control system based on artificial intelligence according to claim 2, characterized in that, The anomaly tracing and prediction module includes: Graph structure construction unit: maps the dynamic association strength matrix into a dynamic attribute graph, where each node in the dynamic attribute graph represents one dimension of operational data, and the edge weights between nodes are defined by the corresponding association strength index; Temporal pattern recognition unit: The sequence of the dynamic attribute graph in continuous time steps is learned by a spatiotemporal graph neural network to extract the evolution rules of nodes and edge weights in the graph; Source tracing and localization submodule: By calculating the graph attention weights of the spatiotemporal graph neural network, the key nodes that contribute the most to the current graph structure anomaly are identified, and the set of key nodes is the source tracing and localization information; The risk prediction submodule, connected to the time-series pattern recognition unit, is used to continue inputting the sequence of the dynamic attribute graph into the time-series prediction model to predict the evolution trajectory of the dynamic correlation strength matrix over a future period of time, and to define the part of the trajectory that exceeds the upper limit threshold of normal evolution as risk evolution prediction.

5. The switchgear safety monitoring and control system based on artificial intelligence according to claim 4, characterized in that, The source tracing and location submodule includes: Abnormal node weight calculation: Based on the spatiotemporal learning results of graph neural networks, obtain the attention weight of the corresponding node for each running dimension; Key node screening: Nodes are sorted according to their attention weights, and key nodes whose attention weights exceed the contribution weight threshold are selected as the source localization results for abnormal states.

6. The switchgear safety monitoring and control system based on artificial intelligence according to claim 4, characterized in that, The risk prediction submodule includes: Graph Modeling and Temporal Learning: Based on dynamic attribute graphs, we use graph neural networks to extract the structural relationships between nodes and combine them with recurrent networks to model the evolution of the graph over time. Association prediction and risk identification: Based on the learning results of the dynamic attribute graph, predict the future changes in the association strength between various dimensions, and identify potential risk relationships by comparing them with the preset normal evolution upper limit threshold, and construct a set of risk paths.

7. The switchgear safety monitoring and control system based on artificial intelligence according to claim 4, characterized in that, The control strategy generation module includes: Strategy Knowledge Base: It contains multiple control strategy templates. Each template defines an anomaly pattern-control action mapping relationship. The anomaly pattern is jointly characterized by the combination of key node types in the source tracing and location information and the risk trend in the risk evolution prediction. Strategy matching engine mechanism: Receives the source location information and risk evolution prediction, and matches the most suitable control strategy template from the strategy knowledge base based on a hybrid decision-making algorithm of rule reasoning and case matching; Instruction generation unit: Instantiates the matched control strategy template into one or more adaptive control instructions that can be directly parsed by the underlying actuator.

8. The switchgear safety monitoring and control system based on artificial intelligence according to claim 7, characterized in that, The hybrid decision-making algorithm includes: Rule-based reasoning and matching: Based on the preset abnormal patterns and control strategies mapping rules, determine whether the combination of node types in the current source tracing and location information and the risk trend in the risk evolution prediction match the matching conditions of a certain strategy template in the strategy knowledge base, and generate a rule-based reasoning score. Case similarity matching: Based on the results of rule reasoning, the feature similarity between the current input and the historical application cases corresponding to each template in the strategy knowledge base is further compared, and the case matching score is calculated based on the degree of matching between node combination and risk trend trajectory.

9. The switchgear safety monitoring and control system based on artificial intelligence according to claim 7, characterized in that, The execution and feedback module includes: Command distribution and execution unit: Receives the adaptive control command, parses it and sends it to the corresponding underlying physical actuator to adjust the operating status of the switch cabinet; Feedback data acquisition and new correlation matrix construction unit: Within a preset effect observation window after the execution of the adaptive control command, multi-dimensional operation data of the switch cabinet is collected again, and a new dynamic correlation strength matrix is ​​generated by comparing the multi-dimensional operation data in the effect observation window with that before the command execution. Effectiveness Quantification Evaluation Unit: Generates a quantitative evaluation value of the intervention effect by comparing the degree of reduction of abnormal correlation modalities before and after instruction execution; Closed-loop optimization signal generation unit: The intervention effect evaluation value, the corresponding source location information, risk evolution prediction and the executed adaptive control command are encapsulated into a reinforcement learning sample, and the reinforcement learning sample is fed back to the operation situation perception module and the control strategy generation module.

10. The switchgear safety monitoring and control system based on artificial intelligence according to claim 9, characterized in that, The underlying physical actuators include controllable circuit breakers, intelligent load switches, variable frequency fans, or partial discharge suppressors.

Images