Monitoring and early warning method for intelligent substation and control system thereof

Abstract

The invention relates to the technical field of transformer substation monitoring and early warning, in particular to a monitoring and early warning method for an intelligent transformer substation and a control system of the monitoring and early warning method. And comprehensive coverage of the equipment state and the environment data is realized. Time sequence modeling is carried out through the long short-term memory network (LSTM), the dynamic trend of operation parameters is effectively captured, potential anomalies are recognized, the accuracy of equipment anomaly judgment is remarkably improved by combining the semantic segmentation technology and equipment type feature extraction, and the false alarm rate and the missing report rate are reduced. Through combination of a rotatable camera and a convolutional neural network (CNN) and multi-modal data fusion, accurate personnel identity recognition and behavior path tracking are realized, non-staff and abnormal behaviors are intelligently detected, and potential security threats are early warned in time.

Description

Technical Field

[0001] The present invention relates to the technical field of substation monitoring and early warning, and in particular to a monitoring and early warning method and a control system for an intelligent substation. Background Art

[0002] With the rapid development of power systems, substations are key nodes for power transmission and distribution, and their operational reliability and safety are crucial to the stability of power systems. The monitoring and management of traditional substations mainly rely on manual inspections and single-point equipment monitoring systems. However, these methods have certain limitations. For example, manual inspections are difficult to achieve real-time monitoring of the operating status of equipment, and human factors may lead to monitoring omissions; while single-point equipment monitoring systems lack overall correlation analysis capabilities and are difficult to identify potential equipment failures or safety hazards in a timely manner. The data between the various monitoring modules in the existing system are independent of each other and lack comprehensive analysis capabilities. At the same time, abnormal judgments are made based on fixed thresholds, and the timing characteristics of equipment operation are not considered, which is prone to false alarms or missed alarms. The existing substation monitoring system mainly focuses on the operating status of the equipment, lacks intelligent monitoring of abnormal behaviors of on-site personnel or non-staff, and poses safety risks. Summary of the invention

[0003] In order to solve the above problems, the present invention provides a monitoring and early warning method and a control system for a smart substation, which realizes intelligent monitoring and early warning by combining the deployment of rotatable cameras, collecting the operating parameters of substation equipment and combining deep learning.

[0004] To achieve the above object, the technical solution adopted by the present invention is: A monitoring and early warning control system for a smart substation, comprising: a real-time monitoring module, an operation abnormality determination module, a behavior determination module and an early warning control module, wherein the operation abnormality determination module and the behavior determination module are respectively connected to the real-time monitoring module, and the early warning control module is respectively connected to the operation abnormality determination module and the behavior determination module; The real-time monitoring module is used to obtain real-time image data of different targets in the substation in a distributed manner, and is also used to obtain operating parameters of substation equipment; The operation abnormality determination module is used to determine the operation abnormality of the equipment through time series analysis according to the operation parameters of the substation equipment, perform semantic segmentation through real-time image data, extract features characterized by equipment type, and determine the abnormality of the equipment in the substation; The behavior determination module is used to determine the identity of a person based on real-time image data, match the person's authority based on the identity, and track the person's behavior path based on the real-time image data based on a convolutional neural network to determine whether the person's behavior is abnormal; The early warning control module is used to classify alarms and control the execution of alarm actions according to abnormal equipment operation, abnormal equipment representation or abnormal personnel behavior.

[0005] Furthermore, the distributed acquisition of real-time image data of different targets of the substation includes the following steps: Several rotatable cameras are set up at different locations within the substation area; Preset a plane rectangular coordinate system for the substation area in the cloud platform; Matching corresponding coordinates in a rectangular plane of a rectangular coordinate system according to the position of the rotatable camera; Based on the orientation and coordinates of the rotatable camera, the target position is determined and real-time images of different targets are obtained.

[0006] Furthermore, the operating parameters of the substation equipment include: electrical parameters, environmental parameters, vibration parameters, insulation state parameters and switch action parameters.

[0007] Further, determining equipment operation abnormality through timing analysis according to operating parameters of substation equipment includes: Perform segmented normalization processing on the real-time collected operating parameters to extract the changing trends of the operating parameters; The normalized operating parameters are modeled in time series based on the long short-term memory network model, and the parameter changes in the future are predicted by learning the characteristics of historical operating data; Dynamically compare the predicted parameters with the actual parameters collected in real time, and use the set abnormal threshold to determine the degree of deviation; Perform cluster analysis on operating parameters that exceed the threshold range, match the corresponding abnormal patterns in combination with the abnormal type library, generate operation abnormality judgment results, and associate the equipment components that cause the abnormality.

[0008] Furthermore, the semantic segmentation is performed through the real-time image data, and the feature extraction is performed through the device type to determine the abnormality of the device characterization in the substation, including: Based on the DeepLab semantic segmentation model, the real-time image data is segmented and processed, the image is divided into the background area and the device target area, and the contour and position features of the device target are extracted; Extract the shape, color and texture features of the equipment; The graph neural network is used to extract features from the segmented device target area, and the device’s pixel points are represented as graph nodes. The geometric feature vector of the device is generated by aggregating node features and adjacency relationships. The generated geometric feature vector is matched with the preset feature library of the device type, and based on the similarity measurement of the feature embedding space, it is determined whether there is any abnormality in the device representation.

[0009] Furthermore, in the graph neural network, the formula for feature updating is as follows: ; in, For each node Representation at the kth layer; Each node Representation at the k+1th layer; For Node The set of neighbor nodes; For Node Neighbor nodes of For Node and neighbor nodes The normalized weight between ; represents the k-th layer weight matrix; represents a non-linear activation function.

[0010] Furthermore, determining the identity of a person based on the real-time image data and matching the person's authority based on the identity includes: The feature points of the person's face and body contour are extracted by combining real-time image data with key point detection algorithms, and the feature vector of the person is generated based on a preset posture model; Use multimodal data fusion to correlate and match the extracted personnel feature vector with the identity feature information obtained based on device interaction data to generate a comprehensive identity judgment; Through the matched identity features, the behavior rules and equipment operation permissions corresponding to the identity are retrieved in the permission management module, including the accessible area, the type of operation equipment and the operation time limit.

[0011] Furthermore, the tracking of personnel behavior paths based on a convolutional neural network according to real-time image data and determining whether personnel behavior is abnormal include: S1. Based on real-time image data, a convolutional neural network is used to detect human targets and generate unique identification and motion trajectory points of the target personnel; S2, based on the orientation and coordinates of the rotatable camera, determine the target position and adjust the rotation angle, focal length and field of view of the camera in real time, continuously obtain the real-time image of the target person and update the motion trajectory points; S3, based on the updated trajectory points, calculate the movement direction and speed of the target person, and when the person is about to leave the field of view of the current camera, predict his / her expected position in the field of view of the next rotatable camera based on the coordinates of the rotatable camera, and pass the prediction parameters and target identity to the next camera; S4, after the next camera receives the parameters transmitted in step S3, it continues to track the target person in combination with the real-time image, generates a connected motion trajectory segment, and performs a fusion check of the timestamp and spatial position of the segment with the trajectory segment recorded by the previous camera to generate a complete behavior path; S5. Match and analyze the complete behavior path with the preset activity rules, and identify abnormal personnel behavior based on the path morphological characteristics, area coverage relationship and stop point distribution. The abnormal personnel behavior includes unauthorized area entry, abnormal stop and abnormal wandering.

[0012] A monitoring and early warning method for a smart substation, applied to any of the above-mentioned monitoring and early warning control systems for a smart substation, comprising: Distributed acquisition of real-time image data of different targets in the substation and the operating parameters of substation equipment; According to the operating parameters of the substation equipment, the equipment operation abnormality is determined through time series analysis, semantic segmentation is performed through real-time image data, and feature extraction is performed based on the equipment type to determine the abnormality of the equipment in the substation; Determine the identity of a person based on real-time image data, match the person's authority based on the identity, and track the person's behavior path based on real-time image data based on a convolutional neural network to determine if the person's behavior is abnormal; Classify alarms and control the execution of alarm actions based on abnormal equipment operation, abnormal equipment performance, or abnormal personnel behavior.

[0013] The beneficial effect of the present invention is that the scheme dynamically acquires real-time image data of different targets of the substation by deploying distributed rotatable cameras in multiple areas of the substation, and ensures comprehensive coverage of equipment status and environmental data by collecting equipment operating parameters. The acquired data is fused through a centralized processing module, breaking the limitation of independent monitoring module data in the traditional system and realizing comprehensive analysis within the entire station. By adopting a time series analysis algorithm and modeling the equipment operating parameters through a long short-term memory network (LSTM) model, the dynamic change trend of equipment operation can be effectively captured and potential anomalies can be identified. In addition, feature extraction is performed by combining semantic segmentation technology with equipment type, which further improves the accuracy of equipment anomaly determination and significantly reduces the incidence of false alarms and missed alarms. In view of the problem that the traditional system has weak ability to monitor personnel behavior, the rotatable camera combined with convolutional neural network (CNN) and multimodal data fusion technology can not only accurately identify the identity of personnel, but also dynamically monitor the activity trajectory of personnel through behavioral path tracking. In particular, the intelligent recognition function of non-staff or abnormal behavior can timely discover and warn potential security threats. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1It is a structural schematic diagram of a monitoring and early warning control system for a smart substation in the present invention.

[0015] Figure 2 It is a step flow chart of a monitoring and early warning method for a smart substation in the present invention. DETAILED DESCRIPTION

[0016] See also Figure 1-Figure 2 As shown, the present invention relates to a monitoring and early warning control system for a smart substation, comprising: a real-time monitoring module, an operation abnormality determination module, a behavior determination module and an early warning control module, wherein the operation abnormality determination module and the behavior determination module are respectively connected to the real-time monitoring module, and the early warning control module is respectively connected to the operation abnormality determination module and the behavior determination module; The real-time monitoring module is used to obtain real-time image data of different targets in the substation in a distributed manner, and is also used to obtain operating parameters of substation equipment; The operation abnormality determination module is used to determine the operation abnormality of the equipment through time series analysis according to the operation parameters of the substation equipment, perform semantic segmentation through real-time image data, extract features characterized by equipment type, and determine the abnormality of the equipment in the substation; The behavior determination module is used to determine the identity of a person based on real-time image data, match the person's authority based on the identity, and track the person's behavior path based on the real-time image data based on a convolutional neural network to determine whether the person's behavior is abnormal; The early warning control module is used to classify alarms and control the execution of alarm actions according to abnormal equipment operation, abnormal equipment representation or abnormal personnel behavior.

[0017] In some embodiments, the system realizes real-time monitoring of different targets of the substation through the hardware composition and integration of the real-time monitoring module. The real-time monitoring module is composed of multiple rotatable cameras deployed in a distributed manner, which have high-definition resolution and multi-dimensional adjustable mechanical pan-tilt. Through the built-in motion control unit, the camera can realize rotation, pitch and focal adjustment according to the instructions of the central control system to accurately cover the key equipment area and pedestrian flow channel of the substation. The camera and multiple environmental sensors (including temperature sensors, humidity sensors, vibration sensors and current and voltage detection modules) form an integrated acquisition network to collect equipment operating parameters and environmental status in real time. The operation abnormality judgment module mainly relies on a high-performance edge computing unit for data processing, and its hardware structure includes a multi-core CPU and an efficient GPU to operate in coordination. After the operating parameters are collected by the sensor, the edge computing unit completes data cleaning, format conversion and compression, and then inputs them into the timing analysis model. The model runs based on an embedded FPGA acceleration unit and uses a long short-term memory network (LSTM) to model the equipment operating parameters in real time. Through this module, the dynamic operation trend of the equipment can be captured, and the feature vector of the equipment representation can be automatically extracted by combining the historical abnormal feature library and the semantic segmentation algorithm, and the equipment representation can be analyzed and determined whether it is abnormal. The hardware configuration of the behavior determination module includes a dedicated AI processing chip (such as the NVIDIA Jetson platform), which is equipped with a deep learning acceleration unit for running a convolutional neural network (CNN) model. The image captured by the camera in real time is parsed by the neural network in the module to extract the key point features of the personnel. Combined with multimodal sensor data, the system tracks the movement trajectory of personnel in real time through the behavior path generation algorithm. In particular, the behavior determination module has a built-in permission rule memory and identity comparison unit, which can determine the identity and authority of personnel through rapid matching with the database. For unauthorized personnel or detected abnormal behaviors, the module generates an early warning signal and sends it to the early warning control module. The hardware of the early warning control module consists of a central control unit (MCU) based on the ARM architecture and a programmable logic controller (PLC). After receiving the abnormal signal, the early warning control module determines the alarm priority according to the classification algorithm, and completes the alarm action through a variety of execution units (such as sound and light alarms, remote control switches or communication interfaces). At the same time, the warning information will be uploaded to the cloud platform storage module synchronously to provide real-time feedback and post-analysis basis for operation and maintenance personnel.

[0018] Furthermore, the distributed acquisition of real-time image data of different targets of the substation includes the following steps: Several rotatable cameras are set up at different locations within the substation area; Preset a plane rectangular coordinate system for the substation area in the cloud platform; Matching corresponding coordinates in a rectangular plane of a rectangular coordinate system according to the position of the rotatable camera; Based on the orientation and coordinates of the rotatable camera, the target position is determined and real-time images of different targets are obtained.

[0019] In some embodiments, several rotatable cameras are deployed in the substation area, and these cameras have precise positioning capabilities and programmable rotation control functions. The position of each camera is stored in the plane rectangular coordinate system preset by the cloud platform in the form of three-dimensional coordinates. To ensure the accuracy of the coordinate system, the system uses a camera self-correction algorithm to automatically adjust the installation angle and coordinate record of the camera by comparing the shooting results of the known reference points with the preset reference coordinates to eliminate the offset caused by the installation error. In the process of real-time image acquisition of the target, the system first calculates the theoretical position of the target object in the camera field of view based on the position coordinates and orientation information of the rotatable camera using a coordinate mapping algorithm based on reverse projection. The algorithm converts the global coordinates of the target into the field of view coordinates of the camera, and calculates the relative position of the target in the imaging plane by combining the pitch angle and focal length information of the camera. In this process, in order to reduce the camera rotation control error, the system adjusts the rotation speed and angle of the camera in real time through the proportional-integral-differential (PID) control algorithm to ensure that the target is always in the center of the camera's optimal field of view. When the field of view of multiple cameras covers the same area, in order to avoid data redundancy, the system introduces an image feature fusion algorithm. Based on the weighted feature matching method, the algorithm combines images from different cameras through geometric correction of feature points and weight fusion of feature vectors to generate a panoramic view. This method not only solves the problem of single camera viewing angle limitation, but also enhances the integrity and consistency of image data. In addition, to improve the accuracy of target detection, the system combines deep learning models for target detection and region segmentation during the real-time image acquisition stage. The YOLOv5 (You Only Look Once, Version 5) model is used to classify and locate targets in real-time images. The model is optimized using a large amount of actual substation scene data during the training phase to improve the recognition of common substation equipment (such as transformers, switchgear) and dynamic targets (such as personnel and vehicles). In real-time operation, the model processes the data streams of multiple cameras through parallel acceleration algorithms (such as TensorRT optimization) to ensure real-time responsiveness.

[0020] Furthermore, the operating parameters of the substation equipment include: electrical parameters, environmental parameters, vibration parameters, insulation state parameters and switch action parameters.

[0021] Specifically, electrical parameters include core data such as current, voltage, power factor, frequency and total harmonic distortion. The system uses high-precision current transformers and voltage transformers to collect these signals in real time and digitize them through a multi-channel analog-to-digital conversion module (ADC). To ensure the stability and anti-interference ability of the data, the sensor is equipped with a low-noise operational amplifier and combined with differential signal processing technology to eliminate common-mode interference. In addition, the electrical parameters extract harmonic characteristics through an algorithm based on Fourier transform for analyzing power quality and abnormal load conditions. Environmental parameters include temperature, humidity and air pressure in the area where the equipment is located. The system obtains these data through a distributed environmental monitoring sensor network. These sensors have wide range and high sensitivity and can operate stably under extreme conditions such as high temperature and high humidity. Environmental parameter data is uploaded to the edge computing node through the LoRa wireless communication module to analyze the impact of environmental changes on equipment operation in real time and generate trend forecasts. Vibration parameters are obtained through high-frequency acceleration sensors, which are arranged on the mechanical parts of key equipment (such as transformers and switch cabinets). The vibration signal collected by the sensor is processed by a bandpass filter to remove low-frequency noise, and key vibration features (such as amplitude and frequency) are extracted through time domain and frequency domain analysis algorithms. These features are combined with fault diagnosis models to determine whether there are potential abnormalities in the mechanical parts of the equipment, such as looseness or fatigue. The insulation state parameters involve the insulation resistance value and dielectric loss factor of the equipment. The system uses online insulation monitoring equipment to regularly collect insulation characteristic data of the equipment based on high-voltage testing and dielectric response analysis methods. By modeling the time series of insulation parameter change trends, the system can predict the insulation aging process and potential breakdown risks. The switch action parameters include the opening and closing time of the circuit breaker, the action current and the current waveform characteristics. The system uses a high-frequency sampling unit to record the dynamic current data during the switch operation, and compares it with the normal switch action reference waveform through a waveform matching algorithm to quickly identify the abnormal operation of the switch mechanical parts or electromagnetic mechanisms.

[0022] Further, determining equipment operation abnormality through timing analysis according to operating parameters of substation equipment includes: Perform segmented normalization processing on the real-time collected operating parameters to extract the changing trends of the operating parameters; The normalized operating parameters are modeled in time series based on the long short-term memory network model, and the parameter changes in the future are predicted by learning the characteristics of historical operating data; Dynamically compare the predicted parameters with the actual parameters collected in real time, and use the set abnormal threshold to determine the degree of deviation; Perform cluster analysis on operating parameters that exceed the threshold range, match the corresponding abnormal patterns in combination with the abnormal type library, generate operation abnormality judgment results, and associate the equipment components that cause the abnormality.

[0023] In some embodiments, first, for the real-time collected operating parameter data (including electrical parameters, environmental parameters, vibration parameters, etc.), the system uses a segmented normalization processing algorithm to standardize the original data. The normalization process scales the data to a uniform numerical range (such as [0,1]) through a linear transformation formula to eliminate the model processing deviation caused by different device parameter dimensions or numerical range differences. At the same time, the system performs segmented analysis on the operating parameters according to the time period characteristics, extracts the change trend of each segment of data, and uses a sliding window mechanism to generate a smoothed trend sequence for subsequent modeling. In the time series modeling stage, the system uses a long short-term memory network (LSTM) model to construct a time series prediction framework. The LSTM model can effectively handle the long-term dependency and short-term change characteristics of the equipment operating parameters through its gated unit structure. The input of the model is the normalized parameter sequence, and the output is the parameter prediction value for a specified time period in the future. During the training process, the system uses large-scale historical operating data to supervise the LSTM model. The objective function is to minimize the mean square error (MSE) between the predicted value and the actual value, and the prediction accuracy of the model is improved by iteratively optimizing the weights. Dynamically comparing the predicted parameters with the actual parameters collected in real time is a key step in anomaly detection. The system comparison process adopts a dynamic threshold setting method. The threshold setting is based on the historical distribution characteristics of the operating parameters and the current operating conditions. It is dynamically adjusted through statistical methods (such as standard deviation calculation) to adapt to the diversity of equipment operation. When the comparison results show that the actual parameters deviate from the predicted parameters and exceed the threshold range, the system marks the parameter as an abnormal candidate value. For the operating parameters marked as abnormal, the system further identifies the abnormal pattern through a clustering analysis algorithm. Specifically, the system uses a density-based clustering method (such as DBSCAN) to cluster the abnormal parameter data to separate isolated abnormal points and dense abnormal clusters. The clustering results are matched with the pattern features in the abnormal type library to determine the abnormal type. Combined with the analysis of the abnormal pattern, the system associates the abnormal data with the equipment components that may cause the abnormality through a decision tree model to generate the final operation abnormality determination result.

[0024] Furthermore, the semantic segmentation is performed through the real-time image data, and the feature extraction is performed through the device type to determine the abnormality of the device characterization in the substation, including: Based on the DeepLab semantic segmentation model, the real-time image data is segmented and processed, the image is divided into the background area and the device target area, and the contour and position features of the device target are extracted; Extract the shape, color and texture features of the equipment; The graph neural network is used to extract features from the segmented device target area, and the device’s pixel points are represented as graph nodes. The geometric feature vector of the device is generated by aggregating node features and adjacency relationships. The generated geometric feature vector is matched with the preset feature library of the device type, and based on the similarity measurement of the feature embedding space, it is determined whether there is any abnormality in the device representation.

[0025] In some embodiments, first, the system uses the DeepLab semantic segmentation model to process the real-time image data and divide the image into a background area and a device target area. The DeepLab model uses the Atrous Convolution technology to enhance the receptive field of feature extraction while ensuring high resolution. Through deep learning model pre-training, the system can identify common equipment (such as transformers, switch cabinets) and background elements (such as the ground and walls) in the substation scene. When the model is running, it receives the image captured by the real-time camera, extracts the contour and position features of the device target area, and generates a mask for subsequent feature processing. After the segmentation is completed, the system further extracts the shape features, color features and texture features of the device target. The shape features are extracted by an algorithm based on shape descriptors, such as calculating parameters such as the area, perimeter and aspect ratio of the device outline. Color feature extraction is based on the statistical method of color histogram to describe the color distribution on the device surface; texture features are modeled by algorithms such as local binary patterns (LBP) and directional gradient histograms (HOG) to obtain texture information on the device surface. These features together constitute a multidimensional description of device representation. Subsequently, the system uses a graph neural network (GNN) to extract high-level features from the segmented device target area. The pixels in the device area are represented as graph nodes, and the connection relationship between nodes is established based on the proximity of pixel positions. GNN aggregates and updates node features through a message passing mechanism to generate a geometric feature vector of the device. Node features include pixel color values, texture feature vectors, and local geometric information, and the aggregation process uses a weighted graph convolution operation. Through a multi-layer graph neural network, the system can capture the global structural information of the device area. The generated geometric feature vector is used to match with a preset feature library of the device type. The feature library is built based on standard images of specific devices and contains geometric descriptions of different device types under ideal conditions. During the matching process, the system uses an embedding space measurement method based on cosine similarity to calculate the similarity between real-time features and preset features. If the similarity is lower than the set threshold, the system determines that the device has a representation anomaly.

[0026] Furthermore, in the graph neural network, the formula for feature updating is as follows: ; in, For each node Representation at the kth layer; Each node Representation at the k+1th layer; For Node The set of neighbor nodes; For Node Neighbor nodes of For Node and neighbor nodes The normalized weight between ; represents the k-th layer weight matrix; represents a non-linear activation function.

[0027] Specifically, the core process of feature update is completed by linear transformation of each node's own features and weighted aggregation of its neighbor node features. In each update, the feature representation of the node consists of two parts: one is the information contribution of the node itself, which reflects the role of the node's original features in the overall structure of the graph; the other is the information contribution of the neighbor nodes, which reflects the strength of the relationship between the node and the surrounding nodes and the structural influence. This process is repeated in each layer of the graph neural network to gradually expand the receptive field of each node, from the description of local relationships to the representation of global structures. To ensure the effectiveness of feature aggregation and the stability of graph structure information, the connection relationship between each node and its neighbor nodes is assigned a normalized weight to adjust the intensity of information propagation between nodes. The normalized weight is calculated by the number of connections of the node (i.e., the degree of the node) to balance the deviation of the relationship between nodes and prevent the excessive influence of highly connected nodes on feature updates. In the initial stage of the graph neural network, the node features are composed of the color value, texture feature and geometric position information of the pixel points in the device area. These features are updated layer by layer, continuously integrating local and global topological structure information, and finally generating a high-level feature representation of the node. In the last layer of the network, the feature vectors of all nodes in the graph are combined into a geometric feature representation of the entire device through global feature aggregation operations (such as averaging or maximum pooling). The resulting geometric feature vector is used for similarity matching with the preset device feature library. By measuring the similarity between the real-time extracted features and the standard features, the system can accurately determine whether there are any abnormalities in the representation of the device. This feature extraction method based on graph neural networks greatly enhances the ability to describe the morphology and structural features of the device, providing a reliable basic support for device anomaly detection and fault warning.

[0028] Furthermore, determining the identity of a person based on the real-time image data and matching the person's authority based on the identity includes: The feature points of the person's face and body contour are extracted by combining real-time image data with key point detection algorithms, and the feature vector of the person is generated based on a preset posture model; Use multimodal data fusion to correlate and match the extracted personnel feature vector with the identity feature information obtained based on device interaction data to generate a comprehensive identity judgment; Through the matched identity features, the behavior rules and equipment operation permissions corresponding to the identity are retrieved in the permission management module, including the accessible area, the type of operation equipment and the operation time limit.

[0029] In some embodiments, the facial and body contour feature points of the person are first extracted using real-time image data combined with a key point detection algorithm. Specifically, the system deploys a high-performance target detection model (such as OpenPose or Mediapipe) to separate human targets from the image and identify key points, including facial feature points (such as eyes, nose, corners of the mouth) and body feature points (such as shoulders, elbows, knees). These key points are geometrically normalized to remove the scale and position differences caused by different camera angles or distances, and generate a standardized feature point set. Based on these key points, the system models the posture information of the person in combination with a preset posture model to generate a set of high-dimensional feature vectors to characterize the facial and body characteristics of the individual. On this basis, the system further associates and matches the person feature vector extracted from the image with the identity feature information generated based on the device interaction data through multimodal data fusion technology. The device interaction data includes the identity data generated when a person passes through the access control, device control terminal or other identity authentication device in the substation. These data are mapped and compared with the image feature vector through a unified embedding space. Specifically, the system uses deep neural networks to build a feature embedding model, maps multimodal data to the same high-dimensional space, and performs feature matching through cosine similarity or Euclidean distance to achieve comprehensive identity determination. During the matching process, the system will assign a confidence score to the matching result. If the confidence is lower than the set threshold, the system will trigger a secondary verification process to ensure the accuracy of the identity determination. When the system successfully matches the identity, it enters the permission verification stage. Permission verification relies on the permission management module, which dynamically loads the behavior rules and operation permissions of the person by retrieving the permission rules associated with the identity in the database. The permission rules include the area range that the person is allowed to enter, the category of the operable equipment, and the allowed operation time period. For example, for an operation and maintenance engineer, the permission rules may include maintenance permissions and limited time for specific equipment; while for a visitor, their permissions may be limited to a short stay in a public area. The system determines whether the current person's activities comply with their permission rules through logical matching. If a violation is found, such as entering an unauthorized area or operating a restricted device, the system generates an early warning signal and notifies the relevant management personnel.

[0030] Furthermore, the tracking of personnel behavior paths based on a convolutional neural network according to real-time image data and determining whether personnel behavior is abnormal include: S1. Based on real-time image data, a convolutional neural network is used to detect human targets and generate unique identification and motion trajectory points of the target personnel; S2, based on the orientation and coordinates of the rotatable camera, determine the target position and adjust the rotation angle, focal length and field of view of the camera in real time, continuously obtain the real-time image of the target person and update the motion trajectory points; S3, based on the updated trajectory points, calculate the movement direction and speed of the target person, and when the person is about to leave the field of view of the current camera, predict his / her expected position in the field of view of the next rotatable camera based on the coordinates of the rotatable camera, and pass the prediction parameters and target identity to the next camera; S4, after the next camera receives the parameters transmitted in step S3, it continues to track the target person in combination with the real-time image, generates a connected motion trajectory segment, and performs a fusion check of the timestamp and spatial position of the segment with the trajectory segment recorded by the previous camera to generate a complete behavior path; S5. Match and analyze the complete behavior path with the preset activity rules, and identify abnormal personnel behavior based on the path morphological characteristics, area coverage relationship and stop point distribution. The abnormal personnel behavior includes unauthorized area entry, abnormal stop and abnormal wandering.

[0031] Specifically, the convolutional neural network (CNN) is first used to detect targets in real-time images. The model adopts the YOLOv5 (You Only Look Once Version 5) architecture. Through its single-step detection capability, the target person is quickly located in each frame and a bounding box and category label are generated. The input of the model is a normalized and resized camera image, and the output is the spatial coordinates and confidence of the target person. For each detected target person, the system assigns a unique identity identifier (ID) and generates the corresponding motion trajectory points of the frame, including the coordinate position and timestamp information of the target. Then, the system dynamically adjusts its orientation and focal length through the built-in rotation control module of the camera to ensure that the target person is always in the center of the field of view. This process is implemented by the proportional-integral-differential (PID) control algorithm. The PID control algorithm receives the offset between the real-time position of the target person and the center of the field of view, and dynamically adjusts the rotation angle and pitch angle of the camera based on error feedback. At the same time, in order to improve the accuracy of the target position calculation, the system combines the built-in geometric correction model of the camera to convert the bounding box position into a position in the global coordinate system to ensure the consistency of the trajectory points between different cameras. When the target person is about to leave the field of view of the current camera, the system uses the motion prediction model to calculate his next position. Motion prediction is based on the Kalman filter, which predicts the target person's motion trajectory by updating the position and velocity estimates in real time. The system combines the fixed layout of the cameras in the substation scene with the field of view coverage to calculate the expected position of the target in the field of view of the next camera. The prediction parameters (including the target identifier, the expected position coordinates and the velocity vector) are transmitted to the next camera through the local area network (LAN) to start the real-time tracking of the next camera. After the next camera receives the prediction parameters, it continues to track the target person in combination with the real-time image data in its field of view to generate a new motion trajectory segment. In order to ensure the continuity of the trajectory of multiple cameras, the system uses timestamp and spatial verification algorithms to fuse the trajectory segments generated by different cameras. Specifically, the timestamp verification ensures the temporal continuity of the trajectory segment by comparing the time intervals of the data transmission of adjacent cameras; the spatial verification ensures the accuracy of the position matching by comparing the positions of the start and end points of the trajectory segment in the global coordinate system. During the fusion process, the system interpolates and smoothes the trajectory points to eliminate the jump errors that may be introduced by camera switching and generate a complete behavior path. After the complete behavior path is generated, the system matches and analyzes the path with the preset activity rules. The activity rules are defined by the substation's safety management policy, including the boundaries of the authorized area, the allowed dwell time, and the reasonable movement mode. The path analysis is carried out in the following three dimensions: Path morphological characteristics: By analyzing the trajectory curvature, angle changes, and consistency of movement direction, it is possible to identify whether the target person has abnormal behavior (such as frequent return or non-straight path).Area coverage relationship: Determine whether a person has entered an unauthorized area by the intersection of the path and the boundary of the authorized area. If the path covers the area around sensitive equipment and does not comply with the permission rules, the system marks it as abnormal. Stay point distribution: Detect whether a person has stayed in a specific area for a long time by clustering the stay points. Trigger an alarm for unauthorized stays or behaviors that exceed the set time.

[0032] The present invention also includes a monitoring and early warning method for a smart substation, which is applied to any of the above-mentioned monitoring and early warning control systems for a smart substation, comprising: Distributed acquisition of real-time image data of different targets in the substation and the operating parameters of substation equipment; According to the operating parameters of the substation equipment, the equipment operation abnormality is determined through time series analysis, semantic segmentation is performed through real-time image data, and feature extraction is performed based on the equipment type to determine the abnormality of the equipment in the substation; Determine the identity of a person based on real-time image data, match the person's authority based on the identity, and track the person's behavior path based on real-time image data based on a convolutional neural network to determine if the person's behavior is abnormal; Classify alarms and control the execution of alarm actions based on abnormal equipment operation, abnormal equipment performance, or abnormal personnel behavior.

[0033] The above implementation modes are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering and technical personnel in the field shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A monitoring and early warning control system for a smart substation, characterized in that: include: Real-time monitoring module, used for distributed acquisition of real-time image data of different targets in the substation, and also for obtaining the operating parameters of substation equipment; The operation abnormality determination module is used to determine the operation abnormality of the equipment through time series analysis according to the operation parameters of the substation equipment, perform semantic segmentation through real-time image data, extract features represented by equipment types, generate geometric morphological feature vectors of the equipment, match the generated geometric morphological feature vectors with the preset feature library of the equipment type, and determine whether there is an abnormality in the equipment representation based on the similarity measurement of the feature embedding space; The behavior determination module is used to determine the identity of a person based on real-time image data, match the person's authority based on the identity, track the person's behavior path based on the real-time image data based on the convolutional neural network, and determine whether the person's behavior is abnormal; The early warning control module is used to classify alarms and control the execution of alarm actions according to abnormal equipment operation, abnormal equipment performance or abnormal personnel behavior.

2. A monitoring and early warning control system for a smart substation according to claim 1, characterized in that: The distributed acquisition of real-time image data of different targets in the substation includes the following steps: Several rotatable cameras are set up at different locations within the substation area; Preset a plane rectangular coordinate system for the substation area in the cloud platform; Matching corresponding coordinates in a rectangular plane of a rectangular coordinate system according to the position of the rotatable camera; Based on the orientation and coordinates of the rotatable camera, the target position is determined and real-time images of different targets are obtained.

3. A monitoring and early warning control system for a smart substation according to claim 1, characterized in that: The operating parameters of the substation equipment include: electrical parameters, environmental parameters, vibration parameters, insulation state parameters and switch action parameters.

4. A monitoring and early warning control system for a smart substation according to claim 1, characterized in that: Determining equipment operation abnormality through timing analysis based on operating parameters of substation equipment includes: Perform segmented normalization processing on the real-time collected operating parameters to extract the changing trends of the operating parameters; The normalized operating parameters are modeled in time series based on the long short-term memory network model, and the parameter changes in the future are predicted by learning the characteristics of historical operating data; Dynamically compare the predicted parameters with the actual parameters collected in real time, and use the set abnormal threshold to determine the degree of deviation; Perform cluster analysis on operating parameters that exceed the threshold range, match the corresponding abnormal patterns in combination with the abnormal type library, generate operation abnormality judgment results, and associate the equipment components that cause the abnormality.

5. A monitoring and early warning control system for a smart substation according to claim 1, characterized in that: The method performs semantic segmentation through real-time image data, extracts features represented by device types, and generates geometric morphological feature vectors of the devices; specifically includes: Based on the DeepLab semantic segmentation model, the real-time image data is segmented and processed, the image is divided into the background area and the device target area, and the contour and position features of the device target are extracted; Extract the shape, color and texture features of the equipment; The graph neural network is used to extract features from the segmented device target area, and the pixel points of the device are represented as graph nodes. The geometric feature vector of the device is generated by aggregating the node features and adjacency relationships.

6. A monitoring and early warning control system for a smart substation according to claim 5, characterized in that: In the graph neural network, the formula for feature update is as follows: ； in, For each node Representation at the kth layer; Each node Representation at the k+1th layer; For Node The set of neighbor nodes of For Node Neighbor nodes of For Node and neighbor nodes The normalized weight between ; represents the k-th layer weight matrix; represents a non-linear activation function.

7. A monitoring and early warning control system for a smart substation according to claim 1, characterized in that: Determining the identity of a person based on real-time image data and matching the person's authority based on the identity includes: The feature points of the person's face and body contour are extracted by combining real-time image data with key point detection algorithms, and the feature vector of the person is generated based on a preset posture model; Use multimodal data fusion to correlate and match the extracted personnel feature vector with the identity feature information obtained based on device interaction data to generate a comprehensive identity judgment; Through the matched identity features, the behavior rules and equipment operation permissions corresponding to the identity are retrieved in the permission management module, including the accessible area, the type of operation equipment and the operation time limit.

8. A monitoring and early warning control system for a smart substation according to claim 7, characterized in that: The tracking of personnel behavior paths based on the convolutional neural network according to the real-time image data and determining that the personnel behavior is abnormal include: S1. Based on real-time image data, a convolutional neural network is used to detect human targets and generate unique identification and motion trajectory points of the target personnel; S2, based on the orientation and coordinates of the rotatable camera, determine the target position and adjust the rotation angle, focal length and field of view of the camera in real time, continuously obtain the real-time image of the target person and update the motion trajectory points; S3, based on the updated trajectory points, calculate the movement direction and speed of the target person, and when the person is about to leave the field of view of the current camera, predict his / her expected position in the field of view of the next rotatable camera based on the coordinates of the rotatable camera, and pass the prediction parameters and target identity to the next camera; S4, after the next camera receives the parameters transmitted in step S3, it continues to track the target person in combination with the real-time image, generates a connected motion trajectory segment, and performs a fusion check of the timestamp and spatial position of the segment with the trajectory segment recorded by the previous camera to generate a complete behavior path; S5. Match and analyze the complete behavior path with the preset activity rules, and identify abnormal personnel behavior based on the path morphological characteristics, area coverage relationship and stop point distribution. The abnormal personnel behavior includes unauthorized area entry, abnormal stop and abnormal wandering.

9. A monitoring and early warning method for a smart substation, applied to a monitoring and early warning control system for a smart substation as claimed in any one of claims 1 to 8, characterized in that: include: Distributed acquisition of real-time image data of different targets in the substation and the operating parameters of substation equipment; According to the operating parameters of substation equipment, equipment operation abnormalities are determined through time series analysis, semantic segmentation is performed through real-time image data, and feature extraction is performed based on the equipment type to determine the abnormalities of equipment characterization in the substation; Determine the identity of a person based on real-time image data, match the person's authority based on the identity, and track the person's behavior path based on real-time image data based on a convolutional neural network to determine if the person's behavior is abnormal; Classify alarms and control the execution of alarm actions based on abnormal equipment operation, abnormal equipment performance, or abnormal personnel behavior.

Images