Electrical cabinet state monitoring and intelligent inspection system based on multi-source data fusion

Abstract

The invention relates to an electrical cabinet state monitoring and intelligent inspection system based on multi-source data fusion, and aims at solving the problems that a traditional electrical cabinet monitoring method is low in efficiency, poor in real-time performance, limited in coverage range and the like. A monitoring platform with high reliability and high real-time performance is constructed by deploying a multi-source sensor network, adopting an edge and cloud cooperative computing model and combining a secure communication protocol design. According to the system, through comprehensive application of a dynamic path planning algorithm, a multi-modal data fusion diagnosis technology and an abnormity diagnosis and early warning model, the inspection efficiency and the fault diagnosis precision are effectively improved. The system not only shows significant advantages in the aspects of data fusion precision, dynamic path planning efficiency and robustness of an anomaly diagnosis model, but also can provide accurate and real-time monitoring and response in a complex industrial environment, and provides reliable guarantee for safe operation of an electrical cabinet.

Description

Technical Field

[0001] The present invention relates to the field of electrical equipment monitoring, and particularly to an electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion and intelligent algorithms. Background Art

[0002] With the development of smart grids and new distribution systems, electrical cabinets, as key nodes for carrying the operation of primary equipment, signal transmission, and system control, have an important impact on the safety and stability of the distribution system. However, most traditional electrical cabinet status management methods rely on manual regular inspections and limited single-point sensor monitoring, suffering from problems such as lagging information collection, isolated data, and untimely fault response, making it difficult to adapt to the increasingly complex distribution network environment and equipment operation requirements.

[0003] Although some environmental parameter monitoring and temperature rise measurement methods have been introduced in the prior art, there is a lack of the ability to perform multi-parameter fusion modeling on the cabinet environment and electrical status, making it difficult to achieve comprehensive judgment and early warning of complex anomalies. In addition, most inspection strategies still adopt fixed cycles or static paths and cannot dynamically optimize task allocation according to the equipment risk level and on-site status, resulting in waste of maintenance resources and insufficient response timeliness. At the same time, in terms of the monitoring architecture, most systems still adopt a centralized computing method with a central server as the core. Facing a large amount of heterogeneous sensing data and high-frequency status update requirements, there are problems of bandwidth bottlenecks and inference delays, making it difficult to meet the real-time decision-making requirements in critical scenarios.

[0004] The present invention proposes an electrical cabinet status monitoring and intelligent inspection system that integrates multi-parameter sensor network design, edge computing, graph neural network analysis, and adaptive inspection control mechanisms. Aiming at the dynamic changes in the operating environment of electrical cabinets and the complex requirements of equipment status identification, a full-process intelligent technology path from data collection, feature fusion, status identification to response control is constructed, significantly improving the system's comprehensive perception ability of multi-source data, risk prediction ability, and terminal response efficiency. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent monitoring system for the operating environment of electrical cabinets and an intelligent inspection system for terminal status, which can combine multi-modal sensor data and artificial intelligence technology to achieve precise monitoring, status perception, and intelligent decision-making of the internal and external operating environments of electrical cabinets.

[0006] To achieve the above purpose, the present invention provides an electrical cabinet status monitoring and intelligent inspection system, including: an intelligent monitoring subsystem for the operating environment of the electrical cabinet and an intelligent inspection subsystem for terminal status; the intelligent monitoring subsystem for the operating environment of the electrical cabinet includes: the electrical cabinet body; Multi-parameter sensing network components, deployed inside and outside the electrical cabinet body, include: an environmental parameter monitoring sub-module, which detects temperature, humidity, condensation, gas concentration, smoke, and water immersion parameters inside the electrical cabinet; an electrical parameter monitoring sub-module, which monitors partial discharge, contact temperature, three-phase voltage, current, harmonics, and three-phase unbalance degree of the operating equipment in the electrical cabinet; Edge computing nodes, deployed in the electrical cabinet body or nearby locations, receive data collected by the environmental parameter monitoring module and the electrical parameter monitoring module, and perform data preprocessing, feature extraction, fusion analysis, and status recognition tasks; A data fusion mechanism, integrated within the edge computing node, includes: a Convolutional Neural Network (CNN), which extracts local features of the data; a Transformer architecture, which performs correlation modeling of multi-modal data to achieve fusion of different sensor data in the time domain and space domain; A dynamic regulation module, including: a temperature and humidity time series analysis sub-module based on Long Short-Term Memory Network (LSTM), which learns the historical change patterns of temperature and humidity inside the electrical cabinet and predicts future short-term trends; a dehumidifier linkage control unit, based on the prediction results of the LSTM model, automatically adjusts the start and stop states of the dehumidifier to achieve dynamic control of the humidity inside the cabinet; a closed-loop ventilation control sub-module, combining the real-time monitoring results of load current and contact temperature, dynamically adjusts the operation mode of the fan to achieve refined cooling of the heat accumulation area, and the control strategy is constructed based on weighted logic rules and PID feedback control algorithms; The terminal status intelligent inspection subsystem, including: a hierarchical inspection architecture, including: A perception layer, which uses fixed sensors and mobile robot sensors to collect environmental parameters of the electrical cabinet body and operation status data of electrical equipment; a data transmission layer, which transmits the data collected by the perception layer to the upper-layer decision-making and analysis module, and the transmission protocols include Modbus, I2C, UDP, CAN, and uses local area network or 5G network for data transmission; a decision-making layer, which uses edge computing nodes and cloud platforms to perform real-time analysis and fault diagnosis on sensor data, supports local and global data fusion, and performs status evaluation and prediction analysis based on deep learning and expert knowledge bases; the decision-making layer includes: a fault diagnosis and early warning model, including: a probability inference sub-module based on a Bayesian network, which compares real-time monitoring data with historical fault data to predict the fault probability of each device status; a Convolutional Neural Network (CNN) is used to extract the spatial features of images to identify the aging of switch contacts, and an SVM is used to diagnose the trend of insulation deterioration; For the SVM model, it is first necessary to prepare the historical failure data of the device, extract key parameters (such as device temperature, vibration intensity, etc.) through feature engineering, and then use scikit-learn to train the model. The training data includes the failure labels of the device; for CNN, use TensorFlow or PyTorch for image processing, extract the spatial features of the image through a convolutional neural network, and determine whether the device components are aging. For the Bayesian network, build a probabilistic inference model through libraries such as Pyro or pgmpy, combine real-time data with historical data, and infer the failure probability of the device. Subsequently, fault diagnosis output is performed. The output of each diagnosis is a failure probability value, which will be used as the input for subsequent task generation and determine the priority of the task. The failure probability value is the core input that determines the behavior of the robot.

[0007] The robot can learn how to select the optimal action according to the current state. The failure probability value is In a multi-robot system, the failure probability value directly affects the priority and urgency of the task. The higher the failure probability, the higher the urgency of the task, and greater rewards should be given to encourage the robot to process these tasks first. In the Q-value update, the immediate reward Rt is dynamically adjusted according to the failure probability In the Q-value update, the immediate reward Rt is dynamically adjusted according to the failure probability dynamically adjusted.

[0008] Among them, is the basic reward for task completion; α is a weight factor that controls the impact of the failure probability on the reward; is the priority factor of the task; considering Under the Q-value update rule of, the Q-value formula is: ; among them, it includes the immediate reward part: the immediate reward here not only includes the basic reward of the task, but also dynamically adjusts the reward according to the failure probability and the priority of the task. The higher the failure probability , the greater the reward value of the task; the future reward still affects through the discount factor γ, but the failure probability value will affect the current decision through the immediate reward.

[0009] At this time, the PPO algorithm optimizes the policy of the robot in each state through the policy gradient method, enabling the robot to select the optimal task according to the current state. The PPO algorithm is updated based on the difference between the current policy and the old policy to ensure that the robot can optimize its task selection policy after multiple trainings. PPO introduces a clipping operation to prevent the policy from being updated too much and ensure the stability of each update, so that the robot can gradually perform optimal task allocation in a multi-task environment.

[0010] The Q - value helps the robot select the optimal task by evaluating the long - term reward of the task. For tasks with a higher failure probability, the system will increase their selection priority through the update of the Q - value. The PPO algorithm optimizes the task - selection strategy through policy gradients to ensure that the robot selects tasks with high Q - values, improving the task - execution efficiency.

[0011] The closed - loop operation and maintenance process includes: a work - order generation and scheduling sub - module, which automatically generates maintenance work orders according to real - time monitoring and fault - diagnosis results, and dispatches tasks according to the severity of equipment failures and the skills of maintenance personnel; a maintenance - task execution sub - module, which guides maintenance personnel to the designated location for on - site maintenance based on GIS positioning and the scheduling information of work - order tasks, and uses mobile terminals to update and feedback real - time data with the system; a feedback and optimization sub - module, which uses real - time operation and maintenance data to feedback and optimize the fault - analysis model.

[0012] The environmental - parameter monitoring sub - module detects temperature and humidity using MEMS sensors, detects condensation using distributed optical - fiber temperature - measurement components, detects gas concentration using electrochemical gas sensors, and detects smoke and water - immersion events using a combination of optical sensors and conductive electrodes.

[0013] For the electrical - parameter monitoring sub - module, partial discharge is detected using ultrasonic sensors and high - frequency current transformers, the contact temperature is detected using non - intrusive wireless temperature - measurement technology, and the sensor signals are transmitted based on the LoRa communication protocol; harmonic and three - phase - imbalance monitoring is achieved using high - frequency current transformers and combined with the FFT algorithm for spectrum analysis.

[0014] The basic form of the PID controller is: where is the temperature deviation, , , are the proportional, integral, and differential gains respectively.

[0015] The convolutional neural network (CNN) specifically includes the following operation steps: S01: Use convolutional operations to extract local features of the image; S02: The convolutional kernel convolves the input image to obtain a feature map: ; where I is the input image and F is the feature map. S03: Use an activation function to increase the non - linear expression ability; S04: Predict the output y; The predicted output y uses a loss function , and the loss function is the cross - entropy loss: , where is the probability value predicted by the CNN model.

[0016] The SVM diagnosis steps include: S01: Find an optimal hyperplane to maximize the margin; S02: Through the constraint condition: ; The optimization objective is to maximize the margin , that is, to minimize the objective function: ; S03: Introduce Lagrange multipliers , and perform classification prediction on the obtained support vector solution: where are the weight coefficients of the support vectors, and the prediction result f(x) is used to determine whether the equipment has insulation deterioration.

[0017] The feedback optimization sub-module includes knowledge base iteration, specifically including: a data acquisition interface that receives status data and text description information from work order execution results and inspection feedback, and performs semantic parsing and structured processing; a label calibration sub-module that evaluates and updates the confidence level of fault labels based on the deviation between manual feedback and automatic judgment results, and uses a time window backtracking mechanism to correct label drift; a knowledge graph update sub-module that uses semantic relation extraction methods to extract new fault events, trigger conditions, and handling solutions from maintenance logs and integrates them into the node and edge weight structure of the graph structure-based knowledge graph; a rule reconstruction sub-module; the knowledge base iteration outputs the updated fault causality, response strategy, and model parameters to the diagnostic model in the decision-making layer.

[0018] The convolutional neural network (CNN) includes Chebyshev polynomial approximation. The graph convolutional layer constructed by the Chebyshev polynomial, based on the graph structure of sensor data in the production process, combines the Chebyshev polynomial to approximately calculate high-order convolutions, optimizes the feature fusion of each node of the sensors in the reactor through local information transmission, and obtains the key indicators of the reaction process in real time. The processing steps include: S01: The sensor data fusion module processes the sensor data, creates a graph convolutional layer and a wavelet CNN layer, and performs feature fusion using the graph convolutional network and the wavelet convolutional neural network; S02: Chebyshev polynomial approximation; S03: Generate a comprehensive feature representation.

[0019] The Chebyshev polynomial further includes that the calculation process of the constructed graph convolutional layer is completed by edge computing nodes. The computing nodes receive the data from each sensor in real time, and after graph convolution and Chebyshev polynomial operations, output the feature vectors of each sensor node; the polynomial operations include high-order graph convolution calculations of the graph convolutional layer constructed by the Chebyshev polynomial, and the formula is: T k =2⋅A⋅T k−1 −T k−2 , where: A is the adjacency matrix of the sensors in the reactor; T k is the graph convolution result of the k-th order Chebyshev polynomial approximation; T k−1 and T k−2 are the convolution results of the first two orders respectively.

[0020] Furthermore, by introducing new environmental parameters such as gas concentration and water immersion, the system further enhances the monitoring ability of the intelligent inspection system.

[0021] Furthermore, the intelligent inspection system integrated in the present invention, through the collaborative scheduling of multi-sensing terminal components, adopts a fusion model of reinforcement learning and game reasoning to automatically generate a scheduling strategy according to real-time task requirements, process section priorities, equipment loads, and path reachability during the production process. The system can support different types of sensing terminal components, including rail-mounted inspection sensing terminals, autonomous mobile vehicles, and robotic arms. Through the cooperation and local synchronization mechanism of these components, the task execution status is exchanged in real time, enabling the inspection task to be completed intelligently and efficiently, and adapting to the flexible and dynamic industrial field tasks.

[0022] Furthermore, the present invention is suitable for long-term stable operation in high-temperature, high-pressure, corrosive media, and flammable and explosive environments. For this purpose, the system adopts a high-strength encapsulation shell that complies with the protection standards of the chemical industry. The shell material is selected as corrosion-resistant alloy and high-temperature-resistant composite coating, and combined with thermal isolation design and multi-layer electromagnetic interference shielding technology to ensure that the core modules inside the electrical cabinet can operate stably in harsh environments and guarantee the long-term reliability of the system.

[0023] Furthermore, for the acquisition requirements of various heterogeneous process parameters existing in chemical production, including flow rate, pressure, temperature, component concentration, etc., the present invention optimizes the acquisition accuracy and stability of the sensor components, supports the unified parsing of analog signals and digital signals, and combines the graph convolutional network algorithm to realize the spatial correlation analysis of information between multiple nodes. In the sensor layout strategy, by combining mobile sensing terminals and fixed-node sensors, a dynamically adjustable data acquisition network is constructed to enhance the coverage ability of the system in multi-region and distributed reaction sections.

[0024] Preferably, the present invention further integrates a multi-modal sensor collaborative perception group, including a high-frequency pressure sampling module, a low-noise temperature and humidity monitor, an IMU inertial measurement unit, a volatile gas sensor, etc. The perception information is synchronized to the edge control core through the parallel communication method of the CAN bus and the I²C protocol. The system constructs a time series modeling layer based on the long short-term memory network (LSTM) and the gated recurrent unit (GRU), combines the CNN and Transformer fusion architecture to extract feature trends and abnormal points, and automatically adjusts sampling frequencies, analysis window sizes, dynamic path planning parameters of sensing terminal components, etc., to improve the adaptability and control ability of the system to complex dynamic reaction processes.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows: First, the present invention solves the problem of isolated sensor data. Through multimodal data fusion technology, data from different sensors can be effectively deeply fused. Through the graph convolutional network (GCN), information can be extracted from the spatial and temporal domains to build a global data model, thus realizing real-time global modeling of the operating status of the electrical cabinet. The system improves the monitoring accuracy of the equipment operating status and the comprehensive cognitive ability of the reaction process, especially in the scenarios of multi-parameter monitoring and high-dimensional data analysis.

[0026] Secondly, the present invention uses a long short-term memory network (LSTM) to perform time series analysis and prediction of temperature and humidity, and combines the closed-loop ventilation strategy of contact temperature and load current to achieve a dynamic control mechanism. The LSTM model learns the changing rules of temperature and humidity and predicts future trends to avoid equipment failures caused by excessive humidity or overheating. By dynamically adjusting the operating status of the dehumidifier and fan, the operating stability and fault prevention capabilities of the equipment are further improved.

[0027] Thirdly, the present invention solves the problems of inflexible task allocation, unavoidable path conflicts, and low execution efficiency in traditional inspection systems. By integrating reinforcement learning and game reasoning models, the present invention can optimize the scheduling and path planning of inspection tasks in real time according to task priority, process section importance, equipment load, and path accessibility. The system can not only automatically generate task scheduling strategies, but also adjust task execution plans according to inspection progress and real-time feedback, significantly improving the intelligence and automation level of the inspection system.

[0028] Finally, the present invention delegates data processing and reasoning tasks to the electrical cabinet site through edge computing nodes, performs local data processing and model reasoning, effectively reduces data transmission requirements, and avoids the bandwidth pressure of traditional centralized computing systems. Through edge computing technology, the system makes decisions within millisecond response time, improving the real-time performance and response speed of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or exemplary descriptions. Obviously, the drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the technical solutions shown in these drawings without creative work.

[0030] Figure 1 Flow chart for environmental parameter monitoring; Figure 2 This is the edge computing and data fusion flow chart; Figure 3 It is the flow chart of LSTM temperature and humidity time series analysis and dehumidifier linkage control; Figure 4 It is a flow chart of temperature adaptive management; Figure 5 This is a hierarchical inspection architecture diagram;Figure 6 This is a schematic diagram of the Chebyshev polynomial state recognition model structure; Figure 7 This is a fault diagnosis and early warning flow chart; Figure 8 Intelligently distribute flowcharts for work orders; Figure 9 Iterate the flow chart for the knowledge base; Figure 10 This is the overall flow chart of the electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion; Figure 11 Flowchart for PPO path planning and multi-robot collaboration; Figure 12 It is the peak signal-to-noise comparison diagram; Figure 13 This is the prediction effect diagram. DETAILED DESCRIPTION

[0031] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0032] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0033] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0034] See also Figure 1 , is a flowchart of environmental parameter monitoring. The monitoring process starts with data collection. Temperature and humidity, condensation, SF6 / O2 concentration, smoke and water immersion sensors synchronously obtain environmental data. The collected data is then transmitted to the edge gateway for preprocessing, including feature extraction and data alignment, to improve the efficiency and accuracy of subsequent analysis. The preprocessed data enters the anomaly detection module to determine whether there is an anomaly based on the preset model. If an anomaly is detected, the system triggers a risk warning and links the control equipment (such as a dehumidifier or ventilation system) to intervene; if the data is normal, it is stored in the database for subsequent analysis and model optimization. This process realizes real-time monitoring and dynamic control of environmental parameters through a closed-loop management mechanism.

[0035] See also Figure 2, is a flowchart for edge computing and data fusion. Sensor data is processed in real time through an edge gateway. After extracting key features, high-value information is transmitted to the cloud for in-depth analysis. At the technical implementation level, the combination of edge computing and data fusion relies on efficient algorithms and architecture designs. For example, the multi-source data fusion algorithm based on spatio-temporal alignment solves the problems of differences in sampling frequencies and data formats among different sensors through timestamp alignment and spatial interpolation techniques. Its mathematical expression can be represented as: , where is the fused data, are the weight coefficients, is the -th raw data of the sensor. By dynamically adjusting the weight coefficients, the algorithm can adapt to data characteristics in different scenarios and ensure the accuracy of the fusion results.

[0036] Refer to Figure 3 , which is a flowchart for the time series analysis of temperature and humidity by LSTM and the linkage control of a dehumidifier.

[0037] When using a long short-term memory network (LSTM) for temperature prediction, a series of preset values need to be set for the model. The parameter settings of the LSTM network have an important impact on the prediction accuracy. The following is a table of the preset values of the LSTM model parameters for the temperature prediction of an electrical cabinet, as shown in Table 1.

[0038] Table 1

[0039] In this context, the forget gate, input gate, and output gate included in its structure enable the model to retain important information in multi-step predictions and avoid the problem of gradient disappearance faced by traditional recurrent neural networks.

[0040] In the prediction of condensation risk, the main task of the LSTM model is to infer the future environmental state and its impact on the condensation risk through historical temperature and humidity data. Given the time series data (where is the temperature and humidity data at a certain moment), the training process of the model minimizes the prediction error by continuously updating the network weights to optimize the prediction results. The structure of the LSTM determines that it can extract long-term dependencies from the sequence, which is particularly important for data types such as temperature and humidity that highly depend on historical states. Specifically, the LSTM performs forward propagation through the following process, and the steps are: is one of the core formulas of the LSTM, where: x t is the input data at the current moment; h t-1 is the state at the previous moment; W h is the weight matrix; b hThe bias is σ, and the activation function usually uses the sigmoid function to control the passing degree of information.

[0041] Specifically: S01 forgetting gate calculation: determines what past memories the network forgets ; S02 input gate calculation: controls the update of the current input information ; S03 candidate cell state calculation: generates the candidate memory at the current moment ; S04 cell state update: combines the results of the forgetting gate and the input gate to update the cell state ; S05 output gate calculation: determines the content to be output at the next moment ; S06 hidden state update: updates the current hidden state under the control of the output gate ; When the model training is completed, it can output the risk value at future moments, and this value can be compared with the safety threshold of the device. If the predicted condensation risk exceeds the safety threshold, the dehumidifier can be triggered to start, adjust the humidity of the environment, and reduce the occurrence probability of condensation. This linkage control mechanism realizes the dynamic adjustment of the device operating environment by combining temperature and humidity prediction with automatic control, which not only improves the safety of the device but also enhances the operating efficiency and response ability of the system.

[0042] Refer to Figure 4 for the temperature adaptive management flow chart. The core of this strategy lies in real-time monitoring of the load current and the contact temperature. By constructing a temperature-current relationship model, the adaptive adjustment of the temperature control system is realized. There is usually a certain functional relationship between the contact temperature and the load current. Especially under high load conditions, the increase in the load current will cause the device to generate more heat, thus affecting the increase in the contact temperature. Assume that the contact temperature of the device under the stable operating state and the load current I(t) can be approximated by the following linear regression model: , where is the gain coefficient between the load current and the contact temperature, is the initial temperature deviation of the system. Through the regression analysis of historical data, the coefficients in this relationship can be accurately estimated, providing a basis for temperature prediction. Based on this model, if the contact temperature exceeds the set safety threshold , the system can initiate corresponding ventilation or cooling measures.

[0043] In the closed-loop regulation process of the temperature control system, the response time of the change in the load current to the change in temperature is crucial. By introducing the PID control algorithm, the precise regulation of the ventilation system can be realized. When the contact temperature When it exceeds the set value, the control system calculates the adjustment range of the ventilation system according to the real-time current I(t) and the temperature change. The basic form of the PID controller is: ; where is the temperature deviation, , , are the proportional, integral, and differential gains respectively. This formula dynamically adjusts the operating state of the ventilation system to ensure that the equipment remains within the safe operating range under load fluctuations and temperature changes.

[0044] The advantage of the closed-loop control strategy lies in its efficient real-time response ability, which can adjust instantaneously when the load current changes and the ambient temperature fluctuates.

[0045] Refer to Figure 5 , which is the hierarchical inspection architecture diagram. It includes: the perception layer and the decision-making layer. The perception layer is the basis of the intelligent inspection system, mainly relying on the collaborative work mode of fixed sensors and mobile robots to achieve comprehensive and accurate monitoring of equipment status. Fixed sensors are deployed around electrical equipment to continuously monitor environmental parameters and equipment operating status, mainly involving two key technologies: infrared thermal imaging and ultrasonic partial discharge detection. Infrared thermal imaging can capture the temperature changes on the surface of the equipment in real time, helping to identify thermal abnormal areas and thus discover potential overload or short-circuit problems. Ultrasonic partial discharge detection can effectively identify potential partial discharge phenomena inside the equipment. Especially in high-voltage equipment, partial discharge is an important precursor to equipment aging and failure.

[0046] At the same time, as a flexible supplement to inspection, mobile robots have the ability of autonomous movement and intelligent perception, and can penetrate into areas that cannot be covered by traditional fixed sensors. By carrying infrared and ultrasonic sensors, mobile robots can conduct more detailed inspections on equipment, especially in complex and inaccessible environments, showing advantages that are difficult to replace by traditional inspection methods. This collaborative work mode not only improves the breadth and depth of equipment monitoring, but also effectively makes up for the limitations of manual inspection.

[0047] In the decision-making layer of the intelligent inspection system, using AI algorithms for defect diagnosis is the core means to achieve automatic fault warning. By constructing a combination of a convolutional neural network (CNN) model and a support vector machine (SVM) model, it is possible to effectively identify the aging of switch contacts and diagnose the trend of insulation deterioration.

[0048] Specifically include: CNN image recognition of switch contact aging: The convolutional neural network (CNN) is widely used in the field of image processing and automatically extracts the spatial features of images through its hierarchical structure. Let the input image be , where H and W are the height and width of the image respectively, and C is the number of channels of the image (such as the three RGB channels). Through the convolution operation, local features of the image are extracted, and the convolution kernel performs convolution on the input image I to obtain a feature map:

[0049] After the generation of the feature map, an activation function (such as ReLU) is used to increase the non-linear expression ability. After multiple calculations of convolution, pooling, and fully connected layers, the final predicted output y is obtained, which is the classification result of whether there is aging of the switch contact. If we set as the binary classification result, where 0 represents normal and 1 represents aging; the loss function usually selects the cross-entropy loss: ; where, is the probability value predicted by the CNN model.

[0050] SVM diagnoses the insulation degradation trend: The support vector machine (SVM) is a binary classification model that constructs a hyperplane to separate data of different classes. In the insulation degradation diagnosis, it is assumed that the input data is representing the monitoring features of the device, and the output represents the insulation state (-1 represents normal and 1 represents degradation). The goal of SVM is to find an optimal hyperplane to maximize the margin: Among them, is the normal vector and b is the bias. Through the constraint condition: ; the optimization goal is to maximize the margin , that is, to minimize the objective function: Through the Lagrange multiplier method, the Lagrange multiplier is introduced to solve the constrained optimization problem. Finally, the obtained support vector solution can be used for classification prediction: where are the weight coefficients of the support vectors, and the prediction result is used to judge whether the device has insulation degradation.

[0051] The diagnostic system combining CNN and SVM can perform multi-dimensional analysis on the health status of electrical equipment. CNN identifies the aging of switch contacts through image data, while SVM identifies the degradation trend of insulation performance based on sensor data. The combination of the two enhances the fault diagnosis ability of the intelligent inspection system, improves the accuracy of equipment status prediction, and reduces the deficiencies and delays of traditional manual inspections.

[0052] Refer to Figure 6, combined with Chebyshev polynomials, approximate high-order convolutions, optimize the feature fusion of each node of the sensors in the reactor through local information transfer, and obtain the key indicators of the reaction process in real time. The data provided by the sensor network is complex and diverse. To accurately predict potential faults or optimize production during the reaction process, it is necessary to process and fuse data from different types of sensors. Graph Convolutional Networks (GCNs) are ideal tools for handling such tasks, but in traditional GCNs, as the depth of the graph structure increases, the computational complexity increases sharply, and at the same time, the calculation of high-order convolutions may also lead to information loss. Therefore, we introduce Chebyshev polynomials to approximate high-order convolutions, thereby reducing the computational overhead and improving the accuracy of the model. The specific steps are as follows: Initialize and import the required libraries, including calculating the output of the graph convolutional layer: the adjacency matrix of the graph (N * N), the node feature matrix (N * F), and the new node feature matrix (N * O); graph convolutional operation: multiplying the node features by the adjacency matrix; the implementation of approximating convolutions with Chebyshev polynomials includes, in sequence: calculating the Laplacian matrix of the graph, initializing Chebyshev polynomials, calculating the high-order terms of Chebyshev polynomials, the Laplacian matrix of the graph, constructed through the adjacency matrix A. Recursively calculate the high-order results of Chebyshev polynomials. Call the sensor function to standardize the data from different sensors and fuse them into a feature matrix, standardize the sensor data to the range of [0, 1]. The order of Chebyshev polynomials for two-layer graph convolution to process graph-structured data. The first layer uses the adjacency matrix approximated by Chebyshev polynomials to perform convolution operations, and the second layer further processes and generates the final output.

[0053] Refer to Figure 7 , which is the flowchart for fault diagnosis and early warning. The fault diagnosis and early warning model plays a crucial role in equipment health management, especially in complex systems. By efficient probabilistic reasoning, potential fault risks can be identified in advance to avoid the spread of faults and catastrophic consequences. Under the framework of Bayesian networks, the fault states of equipment are described by a set of random variables to describe, and each represents the fault state of a device component, and , where 0 indicates that the device is operating normally, and 1 indicates that the component has failed. Through the structure of the Bayesian network, the occurrence probability of equipment faults is affected by the conditional dependence relationships between various components. The conditional probability of each node depends on the state of its parent nodes, indicating the probability that the current node fails given the conditions of its parent nodes.

[0054] The joint probability density function of the Bayesian network is expanded through the chain rule ; where, represents node The set of parent nodes, represents a node The conditional probability distribution of the node under the given conditions of its parent nodes. The structure of the Bayesian network makes it concise and efficient to calculate the probability of the overall state of the computing device.

[0055] The core of the fault diagnosis process is the inference problem, that is, how to infer the fault probability of the device based on real-time monitoring data and historical data. Let the observed data be , and the posterior probability can be obtained through Bayes' theorem, that is, the probability that a certain component of the device fails under the condition of known observed data: ; where is the likelihood function, which represents the probability of observing the data under the given fault state ; is the prior probability, which represents the prior estimate of the occurrence of the fault; and is the marginal likelihood of the observed data, which serves as a normalization constant.

[0056] Through real-time data update, the Bayesian network can perform adaptive inference at each monitoring. As new data is added, the posterior probability will be continuously adjusted to reflect the current health state of the device. On this basis, if the posterior probability of a certain component exceeds the set threshold, the fault warning mechanism will be triggered to prompt the possible occurrence of a fault. This method has high robustness, can effectively cope with the uncertainties in device management, and provides a scientific basis for decision-making.

[0057] By combining historical data, real-time monitoring and device ledger information, the fault diagnosis and warning model based on the Bayesian network can make real-time responses in a dynamic environment, ensure more intelligent device health management, prevent the occurrence of fault risks, and optimize maintenance strategies.

[0058] Based on the fault probability inference of the Bayesian network, it can dynamically evaluate the probability of device faults on the basis of integrating historical data, real-time monitoring data and device ledgers, and provide a scientific basis for decision-making.

[0059] Refer to Figure 8 , Figure 9 , Figure 8 is the intelligent work order dispatching flow chart, Figure 9This is the flow chart for knowledge base iteration. When a device failure is detected, the system first captures the failure information in real time through integrated sensors and monitoring platforms. This failure data includes information such as the device location and failure type. Subsequently, the GIS positioning system is triggered to determine the specific location where the failure occurred. This positioning data is the core basis for subsequent dispatching, ensuring that maintenance personnel can reach the failure site accurately and quickly. At the same time, the system will analyze and match the skill tags of all maintenance personnel in real time, and select the most suitable staff to dispatch tasks in combination with the specific type of failure and required skills.

[0060] The continuous iteration and update of the knowledge base are the core components of an efficient fault diagnosis system. Especially when facing a complex and ever-changing device environment, it can improve the diagnosis accuracy and processing efficiency of the system in a timely manner.

[0061] The automatic archiving mechanism for defect handling cases, through the real-time collection and analysis of historical failure data, systematically stores information such as processing results, failure causes, and solutions in the database. This process is not limited to simple data recording. More importantly, through the archiving, classification, and summarization of each type of failure, it provides valuable input data for the training of subsequent diagnostic models. Through the precise tagging of the archived data, the information in the knowledge base is continuously enriched and improved, thus providing experience accumulation and technical support for the rapid diagnosis and repair of future similar failures.

[0062] At the same time, the online learning mechanism of the diagnostic model enables the system to self-adjust and optimize when facing newly emerging failures. Whenever a new failure case is processed, the system not only archives it in the knowledge base but also incorporates it into the existing diagnostic model through incremental learning algorithms. This online learning method can continuously adjust the parameters and weights of the model, making the diagnostic model more suitable for the current device operating state and failure patterns. For example, a deep learning-based fault diagnosis model adjusts the weights in the convolutional neural network (CNN) through the feedback of each round of processing results to improve the diagnostic accuracy and processing efficiency for future similar failures.

[0063] Through the knowledge base iteration method that combines automatic archiving and online learning, the system can achieve knowledge update from local to global. It can not only improve the efficiency of single-failure handling but also form a highly intelligent fault warning and diagnosis ability in large-scale device operations, promoting the intelligent operation and maintenance system to develop towards a more adaptive and efficient direction.

[0064] See Figure 10, is the overall flowchart of the electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion. The workflow of this system starts with the system startup. When the device starts, the intelligent control system is activated and begins to receive real-time data from various sensors. The sensors transmit the collected physical quantities to the data acquisition module through different protocols (such as I2C, UART, Modbus, etc.), and the system starts to process these raw data. In the data processing stage, the system preprocesses the collected data, including data cleaning, denoising, time alignment, etc., to ensure the consistency of data in space and time and the accuracy of subsequent data analysis.

[0065] After completing the data preprocessing, the system enters the data fusion stage. Data from multiple sources are merged through the data fusion module. Different data sources of each sensor are weighted to generate a unified feature vector, reducing noise interference and improving the data alignment accuracy. These fused data features are passed to the subsequent modules as key inputs.

[0066] Next, based on the fused feature vector, the system uses the dynamic path planning algorithm to determine the optimal path for the inspection task. The path planning algorithm uses real-time sensor data and environmental information, and calculates the most suitable inspection path in combination with the requirements of the inspection task. The path planning module predicts the best path through real-time analysis of the environment and equipment, avoids path conflicts, and optimizes the use of task resources to ensure the efficient completion of the inspection task.

[0067] Once the path planning is completed, the system enters the abnormal diagnosis stage. Through machine learning and abnormal diagnosis models, the system compares and analyzes real-time data and historical data to detect potential equipment failures or abnormal operations in a timely manner. Through the abnormal diagnosis module, the system can not only identify whether there are faults in the equipment, but also give fault warnings according to the results output by the model to ensure preventive measures are taken before the faults occur.

[0068] In the task execution stage, the system starts the inspection task according to the path planning and abnormal diagnosis results. The mobile sensor component executes the inspection task according to the system instructions along the calculated optimal path. During the task execution, the system will update the task status in real time according to the sensor data and make feedback adjustments to ensure that the task is executed according to the optimal plan.

[0069] At the same time, the system will store all the collected and processed data in the local database and synchronize it to the cloud system for remote management. Through this data storage and synchronization mechanism, the system ensures the security and traceability of the data and provides support for remote monitoring and operation. The cloud system can also remotely control and schedule the local system as needed to ensure the efficient operation of the equipment at a remote location.

[0070] In the edge computing module, sensor data is first initially inferred and analyzed locally, reducing data transmission latency and ensuring that the system can respond and react in a timely manner. The local calculation results of the data are sent to the cloud for in-depth analysis to make final decisions and policy adjustments.

[0071] Finally, the system optimizes the operation results through a feedback mechanism, adjusting the working state of the device in real time. The intelligent decision-making module adjusts the device state based on the feedback data to ensure that the device is always in the best operating state. This closed-loop feedback mechanism guarantees the flexibility and efficiency of the system, enabling it to effectively cope with complex industrial environments and continuously optimize work efficiency.

[0072] See Figure 11 , in dynamic path planning, reinforcement learning is often used to guide an agent, such as a robot, to choose the optimal path through a reward mechanism. In a multi-robot system, a combined method of reinforcement learning (RL) and game reasoning is adopted to achieve dynamic path planning through the interaction between agents (robots). Specifically, it includes: Value function (V(s)): , where V(s) is the value function in state s; Rt is the reward at time t; γ is the discount factor. Policy gradient: The action policy is optimized through the policy gradient, and the formula is as follows: ; where, is the probability of taking action a in state s; is the action-state value function. The action-state value function is: , and the obtained value represents the expected reward that can be obtained by taking action a in state s. The robot selects the optimal action according to the Q value. For example, during the inspection process of the robot, if a certain action, such as changing the inspection path, has a relatively high Q value, then the robot will choose to execute that action. By continuously updating the Q value, the robot can gradually improve the effectiveness of its actions and thus optimize the overall inspection path.

[0073] In reinforcement learning, the calculated and the reward are usually stored in the memory bank. The memory bank stores the state, actions taken, rewards obtained, and state transition information of the robot at each moment for future training and policy updates.

[0074] This information constitutes a time step sequence. By analyzing these sequences, the robot can learn how to choose the optimal action according to the current state.

[0075] The update of the Q value is usually based on the Bellman equation. The Bellman equation expresses the relationship between the long-term return of a state-action pair and the immediate reward and the value of the next state. The update rule of the Bellman equation is as follows: ; where: immediate reward : the immediate return obtained by the robot after taking action a t For in a multi-robot system, the failure probability value directly affects the priority and urgency of the task. The higher the failure probability, the higher the urgency of the task, and a greater reward should be given to encourage the robot to process these tasks first. In the Q-value update, the immediate reward Rt is dynamically adjusted according to the failure probability . Therefore, the immediate reward can be expressed as ; where is the basic reward for task completion; α is a weight factor that controls the impact of the failure probability on the reward; is the priority factor of the task; Discount factor γ: The discount factor represents the weight of future rewards. Value of state V(st+1): The value of the new state reached by the robot after taking an action, usually the estimated long-term return of the next state. The formula updates the Q-value at each time step, reflecting the robot's evaluation of the current action. When the robot obtains a new reward or updates the path, the Q-value is updated accordingly, thereby adjusting the strategy for the robot to select actions. Considering the failure probability in the Q-value update rule The function formula is: ; where, including the immediate reward part: The immediate reward here, in addition to including the basic reward of the task, also dynamically adjusts the reward according to the failure probability and the priority of the task. The higher the failure probability , the greater the reward value of the task; Future rewards still affect through the discount factor γ, but the failure probability value will affect the current decision through the immediate reward.

[0076] The Q-value helps the robot select the optimal task by evaluating the long-term return of the task. For tasks with a higher failure probability, the system will increase the priority of their selection through the update of the Q-value. The PPO algorithm optimizes the task selection strategy through policy gradients to ensure that the robot selects tasks with high Q-values and improves the task execution efficiency. The policy update of PPO is carried out through gradient ascent, and the optimization objective is: According to the Q-value and the policy probability, the PPO algorithm will adjust the task selection strategy so that the robot is more inclined to select tasks with high Q-values in similar states.

[0077] At this time, PPO sets the maximum allowable amplitude of the policy update to avoid over-updating the policy, that is, "clipping". Specifically: ; : represents the product of the ratio between the current policy and the old policy, measuring the advantage of an action. If the ratio is close to 1, it means that the policy change is small, and the advantage function Contributes significantly to the objective function and the policy update is normal.

[0078] The cropped part If it exceeds the range , the clipping operation will limit the ratio within this range. This prevents the impact of overly large updates on the policy and ensures that the magnitude of each update is limited.

[0079] The MQTT (Message Queuing Telemetry Transport) protocol uses the publish / subscribe model. In this model, the robot publishes its data to a topic through an MQTT client, and other robots or control systems receive this data as subscribers. In this way, all devices in the system can share data in real time, greatly improving the collaborative efficiency of the multi-robot system. Specifically: Real-time data transmission: Each robot regularly uploads the sensor data and task status collected in real time to the central control system through MQTT. The control system can then receive this data in real time and understand the current status of each robot.

[0080] Task and path update: The robot subscribes to messages from the control system or other robots to receive the latest task instructions, path planning, or adjustment instructions. During execution, the robot continues to perform path planning and task adjustment based on the received feedback. Real-time communication ensures the collaboration and synchronization of the robots, avoiding path conflicts and task duplication.

[0081] See Figure 12 , which is a peak signal-to-noise ratio comparison graph. The performance of the convolutional neural network (CNN) in the image denoising task is verified through the quantitative analysis of the peak signal-to-noise ratio (PSNR). It can be seen from the line graph that as the number of sampling points increases, the PSNR values under different learning rates show significant differences. The PSNR of the original signal-to-noise value fluctuates between 32 dB and 34 dB, indicating obvious noise interference in the unprocessed image. In contrast, the CNN model with a learning rate of 0.10 performs well in the denoising effect. Its PSNR value gradually increases from the initial 32 dB to 36 dB and stabilizes after the sampling point reaches 30. This result shows that a higher learning rate can accelerate the convergence of the model, thereby achieving significant noise suppression in a shorter time.

[0082] Although the initial PSNR value of the CNN model with a learning rate of 0.01 is slightly lower than that of the model with a learning rate of 0.10, its denoising effect gradually approaches and even surpasses the latter after the sampling points exceed 20, and finally stabilizes at around 36.5 dB. This phenomenon indicates that although a lower learning rate has a slower convergence speed, it can optimize the model parameters more finely in the later stage of training, thus obtaining higher denoising accuracy. From the overall trend, the denoising effect of the CNN model gradually improves with the increase of sampling points and stabilizes after the sampling points reach 30. This result is closely related to the deep feature extraction ability of CNN. Through the combination of multiple layers of convolution and non-linear activation functions, it can effectively capture and suppress the noise patterns in the image. In addition, the PSNR change trends at different learning rates also reflect the importance of learning rate selection in the model training process. A higher learning rate can accelerate convergence, while a lower learning rate helps to improve the final performance of the model.

[0083] See Figure 13 , which is the prediction effect diagram. It can be seen from the line chart of the temperature prediction results that the CNN-LSTM model shows advantages in prediction accuracy and trend fitting. Compared with the prediction values of CNN-XGBoost and temperature weighted fusion, the prediction curve of CNN-LSTM fits better with the true temperature value (purple solid line), especially in the middle and late stages of the time series (40 to 80 minutes), and its prediction error is significantly lower than other methods. This phenomenon indicates that by combining the spatial feature extraction ability of the convolutional neural network (CNN) and the time series modeling ability of the long short-term memory network (LSTM), CNN-LSTM can more accurately capture the dynamic law of temperature change.

[0084] During the prediction process, CNN-LSTM not only considers the temperature features at the current moment but also retains the historical temperature information through the memory unit of LSTM. In the time period of 60 to 80 minutes, the true temperature value shows small fluctuations, and the prediction value of CNN-LSTM (purple dashed line) can better reflect this change trend, while the prediction values of CNN-XGBoost (yellow dashed line) and temperature weighted fusion (purple star line) show certain lag and deviation. In addition, the advantage of CNN-LSTM in prediction accuracy is also reflected in its response to the warning value of the measured point temperature (red dashed line), and its prediction value always remains below the warning value, avoiding false alarms and missed alarms.

[0085] From the overall performance, the prediction error of CNN-LSTM is significantly lower than other methods, especially at the inflection points and fluctuation regions of temperature change, where the deviation between its prediction curve and the true value is the smallest. This result verifies the effectiveness of CNN-LSTM in complex time series data modeling. Through its multi-level network structure and joint optimization strategy, it realizes high-precision prediction of temperature change.

[0086] Instructions for Use: The hardware platform has an IP65 protection rating, ensuring that the device can operate stably for a long time in harsh environments, especially suitable for the monitoring requirements of equipment in complex industrial environments. This design effectively improves the reliability of the system and ensures that the device can work properly under high temperature, humidity or other harsh environments, as shown in Table 2.

[0087] Table 2

[0088] The software system architecture adopts a microservices architecture and 3D visualization technology.

[0089] This system consists of multiple hardware and software modules. In the hardware part, temperature and humidity sensors, partial discharge sensors, smoke and water immersion sensors, etc. are installed inside and outside the electrical cabinet. These sensors are connected to the sensor data acquisition module through communication interfaces (such as I2C, UART, and Modbus protocols). The transmission of sensor data uses wired or wireless networks, and the appropriate connection method is selected according to the location of the sensors and the network environment. The status monitoring of the electrical cabinet is through real-time data feedback, enabling the system to continuously monitor the working status inside and outside the electrical cabinet.

[0090] The mobile robot is one of the core execution units of this system and is equipped with sensors for inspection. The robot transmits data and receives task instructions through wireless communication with the computing node. The robot can move autonomously according to the instructions of the path planning system and real-time feedback its current position and task completion status. The robot itself is equipped with a wheeled drive system, and the wheeled drive is connected to the motor control system. After the motor receives the instructions generated by the path planning module, it starts the corresponding drive mode to complete the inspection task. Relying on the built-in edge computing ability, the robot can process part of the data locally, make a preliminary judgment, and then feedback to the central control system as needed.

[0091] The software part mainly includes data processing, analysis, decision-making, and execution systems. The data processing module receives the data uploaded by the sensors and performs necessary data cleaning, filtering, and standardization processing. Through data fusion algorithms, the system can integrate data from different sensors to generate a comprehensive operating status report of the electrical cabinet. After these data are processed by the graph convolutional network (GCN), features are further extracted for fault identification and prediction. The fault diagnosis and early warning model uses machine learning-based algorithms, combines historical data and real-time data, and dynamically evaluates the health status of the equipment. Through methods such as Bayesian inference, the system can early warn of potential faults and trigger maintenance notifications.

[0092] The path planning and task allocation module, based on a reinforcement learning model, intelligently schedules mobile robots to perform inspection tasks. By combining PPO path planning, the MQTT protocol, and digital signature technology, the system achieves dynamic path planning and task execution for multi-robot collaboration. When the system starts, robots obtain environmental data in real time through sensors, including temperature, humidity, location, and other information, which serves as the basis for subsequent path planning and task execution. The acquired sensor data is input into the PPO algorithm, which optimizes the robot's action strategy according to the principles of reinforcement learning and calculates the optimal path. The robot executes tasks according to this path to ensure the efficient completion of inspection and other tasks. During the execution of path planning, the robot publishes information such as its own status and task progress to the control system or other robots in real time through the MQTT protocol. The control system or other robots subscribe to this data to obtain the latest status of the robot and adjust tasks or path planning as needed. In this way, multiple robots can stay synchronized and collaborate to complete tasks, thus avoiding path conflicts and duplicate work.

[0093] To ensure the security of the system and the integrity of data, all messages sent by robots are verified through digital signatures. When each robot publishes a message, it signs the data using its private key, and the recipient verifies the signature using the public key to ensure that the message has not been tampered with and the source is reliable. This mechanism effectively prevents data tampering and malicious attacks and ensures the secure operation of the system.

[0094] The system also designs a real-time feedback mechanism to monitor the task execution of robots. If a robot detects a deviation or conflict in the path during execution, it sends feedback information to the control system to adjust the path planning in a timely manner. This closed-loop mechanism enables the robot to make flexible adjustments according to changes in the environment and ensures the smooth progress of tasks.

[0095] Data generated by robots during task execution, such as path planning and task status, is stored in a local database and synchronized to the cloud for backup and management. Cloud synchronization enables data to be stored securely and provides a basis for subsequent data analysis and decision-making.

[0096] After the mobile robot completes the inspection task, it feeds back its status to the central control system, and the system adjusts tasks or assigns new tasks according to the feedback. When a fault is detected, the system automatically generates a maintenance work order, locates the fault location through the GIS system, and intelligently assigns it to maintenance personnel with appropriate skills and permissions. After receiving the task, the maintenance personnel perform equipment maintenance according to the given instructions. After the maintenance is completed, the system records relevant data and conducts performance verification to ensure that the equipment resumes normal operation.

Claims

1. The electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion is characterized by: The system includes: Intelligent monitoring subsystem for the operating environment of electrical cabinets and intelligent inspection subsystem for terminal status; The electrical cabinet operating environment intelligent monitoring subsystem includes: Electrical cabinet body; A multi-parameter sensor network component is deployed inside and outside the electrical cabinet body, including: an environmental parameter monitoring submodule, which detects temperature, humidity, condensation, gas concentration, smoke, and water immersion parameters inside the electrical cabinet; an electrical parameter monitoring submodule, which monitors partial discharge, contact temperature, three-phase voltage, current, harmonics, and three-phase imbalance of the running equipment in the electrical cabinet; The edge computing node is deployed in the electrical cabinet body or a nearby location, receives the data collected by the environmental parameter monitoring module and the electrical parameter monitoring module, and performs data preprocessing, feature extraction, fusion analysis and state recognition tasks; A data fusion mechanism, integrated in the edge computing node, includes: a convolutional neural network (CNN), the convolutional neural network (CNN), a Transformer architecture for extracting local features of data, and the Transformer architecture for performing correlation modeling of multimodal data to achieve fusion of different sensor data in the time domain and the space domain; The dynamic control module includes: a temperature and humidity time series analysis submodule based on a long short-term memory network (LSTM), which learns the historical change pattern of temperature and humidity in the electrical cabinet and predicts future short-term trends; a dehumidifier linkage control unit, which automatically adjusts the start and stop state of the dehumidifier based on the prediction results of the long short-term memory network (LSTM) model to achieve dynamic control of the humidity inside the cabinet; a closed-loop ventilation control submodule, which dynamically adjusts the fan operation mode in combination with the real-time monitoring results of the load current and contact temperature to achieve refined cooling of the heat accumulation area, and the control strategy is constructed based on weighted logic rules and PID feedback adjustment algorithm; The terminal status intelligent inspection subsystem includes: a layered inspection architecture, including: The perception layer uses fixed sensors and mobile robot sensors to collect environmental parameters of electrical cabinets and operating status data of electrical equipment; the data transmission layer transmits the data collected by the perception layer to the upper-level decision analysis module. The transmission protocols include Modbus, I2C, UDP, and CAN, and use local area networks or 5G networks for data transmission; The decision layer uses edge computing nodes and cloud platforms to perform real-time analysis and fault diagnosis on sensor data, supports local and global data fusion, and performs status assessment and predictive analysis based on deep learning and expert knowledge base; the decision layer includes: fault diagnosis and early warning models, including: spatial feature recognition of switch contact aging based on Bayesian network path images, and diagnosis of insulation degradation trends based on SVM; Closed-loop operation and maintenance process, including: A work order generation and scheduling submodule, which automatically generates maintenance work orders based on real-time monitoring and fault diagnosis results, and dispatches tasks based on the severity of equipment failures and the skills of maintenance personnel; Maintenance task execution submodule, which guides maintenance personnel to the designated location for on-site maintenance based on GIS positioning and work order task scheduling information, and uses mobile terminals and the system for real-time data update and feedback; A feedback optimization submodule uses real-time operation and maintenance data to provide feedback and optimize the fault analysis model.

2. The electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion according to claim 1 is characterized in that: The environmental parameter monitoring submodule detects temperature and humidity using MEMS sensors, condensation monitoring uses distributed optical fiber temperature measurement components, gas concentration is detected by electrochemical gas sensors, and smoke and water immersion events are detected by optical sensors and conductive electrodes.

3. The electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion according to claim 1 is characterized in that: The electrical parameter monitoring submodule uses ultrasonic sensors and high-frequency current transformers to detect partial discharge, and uses non-invasive wireless temperature measurement technology to measure contact temperature. The sensor signal is transmitted based on the LoRa communication protocol. Harmonic and three-phase imbalance monitoring is achieved using high-frequency current transformers, combined with FFT algorithm for spectrum analysis.

4. The electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion according to claim 1 is characterized in that: The basic form of the PID feedback regulation algorithm is: ;in, is the temperature deviation, , , are the proportional, integral and derivative gains respectively.

5. The electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion according to claim 1 is characterized in that: The convolutional neural network (CNN) specifically includes the following operation steps: S01: using convolution operation to extract local features of the image; S02: convolution kernel Convolve the input image to get the feature map: Convolve the input image to get the feature map: ; Where I is the input image and F is the feature map; S03: Use activation function to increase nonlinear expression ability; S04: Predict output y.

6. The electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion according to claim 5 is characterized in that: The predicted output y uses the loss function , the loss function is cross entropy loss: ;in, is the probability value predicted by the CNN model.

7. The electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion according to claim 1 is characterized in that: The SVM diagnosis steps include: S01: finding an optimal hyperplane to maximize the interval; S02: through the constraint conditions: ; The optimization goal is to maximize the interval , that is, minimize the objective function: ; S03: Introducing Lagrange multipliers , the obtained support vector solution is used for classification prediction: in is the weight coefficient of the support vector, and the prediction result f(x) determines whether the equipment has insulation degradation.

8. The electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion according to claim 1 is characterized in that: The feedback optimization submodule includes knowledge base iteration, specifically including: The data acquisition interface receives status data and text description information from work order execution results and inspection feedback, and performs semantic parsing and structured processing; the label calibration submodule evaluates and updates the confidence of fault labels based on the deviation between manual feedback and automatic judgment results, and uses the time window backtracking mechanism to correct label drift; the knowledge graph update submodule uses the semantic relationship extraction method to extract new fault events, trigger conditions and disposal plans from the maintenance log, and integrates them into the knowledge graph nodes and edge weight structure based on the graph structure; the rule reconstruction submodule; the knowledge base iteratively outputs the updated fault causal relationship, response strategy and model parameters to the diagnostic model in the decision layer.

9. The electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion according to claim 5 is characterized in that: The convolutional neural network (CNN) includes Chebyshev polynomial approximation; a graph convolution layer constructed by the Chebyshev polynomials, which is based on the graph structure of sensor data in the production process and combines Chebyshev polynomials to perform approximate calculations on high-order convolutions, optimizes the feature fusion of each sensor node in the reactor through local information transmission, and obtains key indicators of the reaction process in real time.

10. The electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion according to claim 9 is characterized in that: The Chebyshev polynomial further includes that the calculation process of the constructed graph convolution layer is completed by the edge computing node, the computing node receives data from each sensor in real time, and outputs the feature vector of each sensor node after graph convolution and Chebyshev polynomial operation; the polynomial operation includes the high-order graph convolution calculation of the graph convolution layer constructed by the Chebyshev polynomial.

11. The electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion according to claim 1 is characterized in that: The electrical cabinet operating environment intelligent monitoring subsystem specifically includes: PPO algorithm, the algorithm includes: S01: initialization environment and task setting; S02: calculation of value function (V(s)); S03: strategy gradient Optimization; S04: Action-State Value Function Evaluation; S05: Training and optimization of PPO.

12. The electrical cabinet status monitoring and intelligent inspection system based on multi-source data fusion according to claim 11 is characterized in that: The PPO algorithm uses the MQTT protocol; specifically, the MQTT protocol is used as a message transmission protocol, and the data transmission of the MQTT protocol uses a digital signature.

Images