Video monitoring equipment anomaly detection method based on artificial intelligence transformer substation

Abstract

The invention discloses a video monitoring equipment anomaly detection method based on an artificial intelligence transformer substation. The method comprises the following steps: 1, determining power equipment needing to be monitored and a specific anomaly type thereof, and designing a cloud-side-end collaborative system architecture; 2, synchronously acquiring visible light and infrared video data of the equipment, and constructing a high-quality data set subjected to accurate labeling and enhancement processing; 3, for different anomalies, developing or training a special deep learning model by modules, so that the AI learns to identify various specific fault modes; 4, integrating the optimized lightweight model into local edge computing equipment of the transformer substation, constructing an automatic video analysis assembly line, and realizing real-time detection and field alarm; 5, the system continuously collects feedback in actual operation, continuous learning and iterative upgrading are performed on the model through manual case rechecking, so that the detection capability is continuously evolved in practice, and the detection capability is more and more accurate; the method has the advantages of automation, precision, real-time performance and evolutionary performance.

Description

An abnormal detection method for video surveillance equipment in substations based on artificial intelligence Technical Field

[0001] This invention belongs to the field of substation inspection and maintenance technology, specifically relating to an abnormal detection method for video surveillance equipment in substations based on artificial intelligence. Background Technology

[0002] Artificial intelligence substations refer to modern substations that utilize artificial intelligence technologies (such as machine learning, computer vision, natural language processing, knowledge graphs, etc.) to deeply process and analyze massive amounts of substation operation data (from online monitoring, video, infrared, inspection robots, protection devices, etc.), thereby achieving advanced applications such as intelligent equipment operation and maintenance, intelligent operation control, intelligent safety management and control, and autonomous inspection operations.

[0003] In the process of generating massive operational data in AI-powered substations, video surveillance is one of the most intuitive inspection and maintenance technologies. However, current video surveillance systems still have some problems, as follows:

[0004] 1. Manual inspection and processing of video images is inefficient and inconsistent;

[0005] 2. Anomaly detection relies on expert experience, resulting in a high rate of missed detections and false positives;

[0006] 3. Fault warnings are delayed, and most issues are addressed after the fact;

[0007] To address the aforementioned issues, it is essential to develop an anomaly detection method for video surveillance equipment in substations based on artificial intelligence. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide an automated, precise, real-time, and evolvable method for detecting anomalies in video surveillance equipment in substations based on artificial intelligence, which can fundamentally improve the reliability, economy, and security of substation equipment status monitoring.

[0009] The objective of this invention is achieved as follows: a method for anomaly detection in video surveillance equipment of substations based on artificial intelligence, comprising the following steps:

[0010] S1, System Design and Goal Definition: Clearly define the power equipment that needs to be monitored and its specific anomaly types, and design a cloud-edge-device collaborative system architecture to create a blueprint for the entire project;

[0011] S2, Data Acquisition and Preprocessing: Simultaneously acquire visible light and infrared video data from the device, and construct a high-quality dataset that has been accurately labeled and enhanced, laying a solid foundation for training AI models;

[0012] S3, Core Algorithm Model Development: For different anomalies, develop or train dedicated deep learning models in modules to enable AI to learn to identify various specific fault modes.

[0013] S4, System Integration and Edge Deployment: Integrate the optimized lightweight model into the local edge computing device of the substation to build an automated video analysis pipeline and realize real-time detection and on-site alarm;

[0014] S5, Closed-loop verification and iterative optimization: The system continuously collects feedback during actual operation, and continuously learns and iteratively upgrades the model through manual review of cases, so that the detection capability continues to evolve in practice and becomes more and more accurate with use.

[0015] Furthermore, step S1 includes the following steps:

[0016] S11, clearly define the detection target and anomaly type:

[0017] S111, Testing equipment: transformers, circuit breakers, disconnect switches, voltage / current transformers, surge arresters, insulators, control cabinets;

[0018] S112, Exception type:

[0019] Appearance defects: breakage, rust, oil leaks, detached parts, hanging foreign objects;

[0020] Abnormal status: Incorrect switch position, indicator light malfunction;

[0021] Instrument malfunction: Pointer / digital meter readings exceed limits, are unclear, or are obstructed;

[0022] Thermal defects: Abnormally high localized temperatures at electrical connection points or within the equipment body;

[0023] S12, Design System Architecture:

[0024] End-side: Deploy high-definition visible light cameras, infrared thermal imagers, and dual-spectrum PTZ cameras on the equipment side to be responsible for high-quality video and image data acquisition;

[0025] Edge computing devices are deployed locally in the substation to carry core AI algorithm models and realize real-time video stream analysis, anomaly detection and local alarms;

[0026] Cloud-based: Deployed in the monitoring center, responsible for aggregating data from multiple stations, centralized model training and optimization, global alarm management, report generation, and remote model distribution.

[0027] Furthermore, step S2 includes the following steps:

[0028] S21, Multispectral Data Acquisition:

[0029] S211, synchronously acquire visible light images and infrared thermal images from the same device, and perform time and space registration to ensure that the same physical point corresponds in the two images;

[0030] S212, different acquisition frequencies are set for different detection targets;

[0031] S22, Building a high-quality dataset:

[0032] S221, For visible light images, use annotation tools to mark the positions of equipment components, instrument panels, and indicator lights, and indicate their normal / abnormal status;

[0033] S222, mark abnormal heating areas in infrared images;

[0034] S223, Data Augmentation: Expand the dataset by rotating, scaling, adding noise, and simulating different lighting and weather conditions to improve model robustness.

[0035] Furthermore, step S3 is performed according to the following table:

[0036]

[0037] Furthermore, the target detection model employs YOLOv8 and Faster R-CNN.

[0038] Furthermore, the image segmentation employs U-Net.

[0039] Furthermore, the image classification employs ResNet.

[0040] Furthermore, the anomaly detection model employs an autoencoder.

[0041] Furthermore, step S4 includes the following steps:

[0042] S41, Model Lightweighting and Optimization: The trained model is compressed using pruning, quantization, and knowledge distillation techniques to adapt to the limited computing power of edge computing devices;

[0043] S42, Building a video analytics pipeline:

[0044] S421, deploying an algorithm pipeline on an edge computing device: video stream access → key frame extraction → multi-module parallel analysis → result fusion and decision-making → generating alarms / reports;

[0045] S422, alarm rule engine: Set up multi-level alarms, including early warning, alarm, and emergency, and support multiple alarm push methods, including sound and light, SMS, and platform pop-up.

[0046] Furthermore, step S5 includes the following steps:

[0047] S51, Online Testing and Feedback: During system trial operation, false alarms and missed alarms are collected to form a feedback loop;

[0048] S52, Continuous Model Learning: New abnormal samples verified by humans are sent back to the cloud training platform for incremental training or version iteration of the model, continuously improving the accuracy of the model in real complex environments.

[0049] S53, Performance Evaluation: Continuously monitor the core indicators of the system, including accuracy, recall, false alarm rate, and average processing time, to ensure that it meets the practical requirements of substation operation and maintenance.

[0050] Due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0051] (1) This invention deploys visible light and infrared video monitoring equipment and combines edge computing to achieve “7×24-hour automatic inspection”, using machines to replace manual labor to complete repetitive and high-risk on-site inspections, realizing unmanned and efficient inspection, freeing personnel from repetitive labor, greatly improving inspection efficiency, and is not affected by time or weather.

[0052] (2) By applying customized deep learning models such as target detection and image segmentation and multispectral fusion analysis, this invention enables AI to learn to accurately identify instrument readings, equipment status and infrared thermal defects, thereby achieving precision and objectivity in detection, greatly improving the accuracy of key equipment status identification, and significantly reducing false alarms and missed alarms.

[0053] (3) By establishing a real-time analysis closed loop of cloud-edge collaboration, this invention can detect anomalies and issue alarms within the station in milliseconds, and at the same time feed the data back to the cloud platform for trend analysis and model iteration, thereby realizing early warning and foresight of operation and maintenance, changing passive emergency repair to proactive early warning, and can detect latent defects such as overheating of contacts in advance to avoid the expansion of faults.

[0054] In summary, this invention has the advantages of automation, precision, real-time performance, and evolvability, fundamentally improving the reliability, economy, and security of substation equipment status monitoring. Attached Figure Description

[0055] Figure 1 is a flowchart of the method steps of the present invention.

[0056] Figure 2 is a deployment topology diagram of a specific embodiment of the present invention. Detailed Implementation

[0057] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0058] As shown in Figure 1, this invention provides an anomaly detection method for video surveillance equipment in substations based on artificial intelligence. This method follows the principles of "data-driven, intelligent analysis, and closed-loop management," and includes the following steps:

[0059] S1, System Design and Goal Definition: Clearly define the power equipment that needs to be monitored and its specific anomaly types, and design a cloud-edge-device collaborative system architecture to create a blueprint for the entire project.

[0060] S11, clearly define the detection target and anomaly type:

[0061] S111, Testing equipment: Transformers, circuit breakers, disconnect switches, voltage / current transformers, surge arresters, insulators, control cabinets and other major equipment.

[0062] S112, Exception type:

[0063] Appearance defects: breakage, rust, oil leaks, detached parts, hanging foreign objects;

[0064] Abnormal status: Incorrect switch position, indicator light malfunction;

[0065] Instrument malfunction: Pointer / digital meter readings exceed limits, are unclear, or are obstructed;

[0066] Thermal defects: Abnormal temperature rise (overheating) at electrical connection points or in localized areas of the equipment body.

[0067] S12, Design System Architecture:

[0068] End-side (acquisition layer): Deploy high-definition visible light cameras, infrared thermal imagers, and dual-spectrum PTZ cameras on the equipment side to be responsible for high-quality video and image data acquisition, thereby achieving 24 / 7 uninterrupted automatic inspection and freeing up manpower;

[0069] Edge computing (analysis layer): Deploy edge computing devices (AI hosts) locally in the substation to carry core AI algorithm models and realize real-time video stream analysis, anomaly detection and local alarms;

[0070] Cloud-based (management level): Deployed in the monitoring center, responsible for aggregating data from multiple stations, centralized model training and optimization, global alarm management, report generation, and remote model distribution.

[0071] S2, Data Acquisition and Preprocessing: Simultaneously acquire visible light and infrared video data from the device and construct a high-quality dataset that has been precisely labeled and enhanced, laying a solid foundation for training AI models.

[0072] S21, Multispectral Data Acquisition:

[0073] S211 simultaneously acquires visible light images and infrared thermal images from the same device and performs temporal and spatial registration to ensure that the same physical point corresponds in the two images. By fusing visible light and infrared information, it achieves more comprehensive detection dimensions and more accurate positioning.

[0074] S212 allows for setting different acquisition frequencies for different detection targets (e.g., instruments can be measured in minutes, while thermal imaging can be measured in hours).

[0075] S22, Building a high-quality dataset:

[0076] S221. For visible light images, use annotation tools to mark the positions (boundaries) of equipment components, instrument panels, indicator lights, etc., and mark their normal / abnormal status (such as open, closed, normal reading, over-limit).

[0077] S222, mark abnormal heating areas in infrared images;

[0078] S223, Data Augmentation: Expand the dataset by rotating, scaling, adding noise, and simulating different lighting and weather conditions (fog, rain, snow) to improve model robustness.

[0079] S3, Core Algorithm Model Development: Develop or train dedicated deep learning models for different anomalies in modules, enabling AI to learn to identify various specific fault modes, as shown in the table below:

[0080]

[0081] For object detection, YOLOv8 or Faster R-CNN can be used; for image segmentation, U-Net can be used; for image classification, ResNet can be used; and for anomaly detection, an autoencoder can be used.

[0082] S4, System Integration and Edge Deployment: Integrate the optimized lightweight model into the local edge computing equipment of the substation to build an automated video analysis pipeline and realize real-time detection and on-site alarms.

[0083] S41, Model Lightweighting and Optimization: The trained model is compressed using techniques such as pruning, quantization, and knowledge distillation to adapt to the limited computing power of edge computing devices.

[0084] S42, Building a video analytics pipeline:

[0085] S421 deploys an algorithm pipeline on the edge AI host: video stream access → key frame extraction → multi-module parallel analysis → result fusion and decision-making → alarm / report generation, thereby reducing alarm latency to the second level through edge analysis and achieving rapid response;

[0086] S422, alarm rule engine: set up multi-level alarms (such as early warning, alarm, emergency) and support multiple alarm push methods (audio and visual, SMS, platform pop-up).

[0087] S5, Closed-loop verification and iterative optimization: The system continuously collects feedback during actual operation, and continuously learns and iteratively upgrades the model through manual review of cases, so that the detection capability continues to evolve in practice and becomes more and more accurate with use.

[0088] S51, Online Testing and Feedback: During system trial operation, false alarms and missed alarms are collected to form a feedback loop. For example, the system alarms are compared and recorded with the results of manual confirmation. The closed-loop feedback mechanism enables the system to have continuous learning capabilities and become more intelligent with use.

[0089] S52, Continuous Model Learning: New abnormal samples (especially difficult examples) verified by humans are sent back to the cloud training platform to perform incremental training or version iteration on the model, continuously improving the model's accuracy in real-world complex environments.

[0090] S53, Performance Evaluation: The core indicators of the continuous monitoring system include accuracy, recall, false alarm rate and average processing time, to ensure that it meets the practical requirements of substation operation and maintenance (such as the State Grid's relevant standards requiring a typical defect identification accuracy of >95%).

[0091] It is worth mentioning that the method of this invention is a systematic framework, and in actual projects, detailed adjustments and parameter optimizations need to be made according to the specific equipment type, site environment and business needs.

[0092] The following is a specific embodiment of the present invention for transformers. Transformers have complex structures and high value, and their monitoring requires comprehensive information from multiple dimensions such as appearance, status, and temperature. This embodiment applies the aforementioned method to the main transformer of a substation, demonstrating a complete closed-loop solution from equipment deployment to algorithm analysis.

[0093] I. Definition of Monitoring Targets and Equipment Deployment Plan

[0094] First, identify the key parts of the transformer that need to be monitored and equip them with appropriate "sensing organs," as shown in the table below:

[0095]

[0096] Then, as shown in Figure 2, the deployment of the data acquisition equipment is completed.

[0097] II. Core Algorithm Strategy and Anomaly Detection Process

[0098] The edge AI host performs parallel analysis on the incoming video streams using the following strategy:

[0099] Step 1: Scheduled task triggering and video stream acquisition

[0100] The system automatically controls the pan-tilt unit to rotate to a preset position according to a preset cycle (e.g., oil level / oil temperature every 2 hours, infrared temperature measurement every 4 hours, fan status every hour), and captures high-definition images or short video streams.

[0101] Key: Dual-spectrum equipment simultaneously acquires visible light images and infrared thermal images of the same location and performs pixel-level registration.

[0102] Step 2: Multi-module parallel AI analysis

[0103] For different parts, a pre-trained lightweight model is used for analysis, as shown in the table below:

[0104]

[0105] Step 3: Results Fusion and Alarm Decision Making

[0106] The AI ​​host summarizes and logically correlates the analysis results of each module, and generates tiered alarms: Level 1 alarm (urgent): oil temperature seriously exceeds the limit + overheating is found in the same area, the system immediately pushes SMS and audible and visual alarms; Level 2 alarm (important): oil level is abnormal or the temperature difference of a single hot spot exceeds the standard, pushes to the operation and maintenance platform, and requires planned inspection; Level 3 alarm (notification): fan stops or suspected oil stains are found, and an inspection work order is generated.

[0107] Step 4: Generate report and conduct manual review

[0108] The system automatically generates standardized reports containing abnormal images / thermal images, data, location, and time. These reports are pushed to the central control center's cloud platform for operation and maintenance personnel to review and confirm. The confirmation results (true / false alarm) are then fed back to the system as key data.

[0109] III. Continuously optimize the closed loop:

[0110] 1. Difficult Case Collection: Operation and maintenance personnel mark false alarms and missed alarms on the platform, and the system automatically stores the corresponding original video data and annotations into the "Difficult Case Library".

[0111] 2. Model Iteration: The model is incrementally trained regularly in the cloud using the "difficult example library" and new normal data to generate new versions of the model.

[0112] 3. OTA (Over-The-Air Technology) update: Securely download the optimized model to the edge AI host within the site to complete the algorithm upgrade, thereby achieving self-evolution that becomes more accurate with use.

[0113] In summary, this invention utilizes computer vision technology based on visible light and infrared fusion and deep learning technology based on machine learning to automatically analyze the visible light and infrared video streams of key equipment in substations. This enables real-time and accurate detection of abnormal conditions, and can detect early thermal defects that are difficult for the human eye to detect. It realizes the transformation from post-maintenance to pre-prediction, and provides a solid guarantee for the safe and stable operation of core substation assets.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for anomaly detection in video surveillance equipment of a substation based on artificial intelligence, characterized in that... The process includes the following steps: S1, System Design and Goal Definition: Clearly define the power equipment to be monitored and its specific anomaly types, and design a cloud-edge-device collaborative system architecture to create a blueprint for the entire project; S2, Data Acquisition and Preprocessing: Simultaneously acquire visible light and infrared video data from the equipment, and construct a high-quality dataset that has undergone precise annotation and enhancement processing to lay a solid foundation for training the AI ​​model; S3, Core Algorithm Model Development: Develop or train dedicated deep learning models for different anomalies in modules, enabling the AI ​​to learn to identify various specific fault modes; S4, System Integration and Edge Deployment: Integrate the optimized lightweight model into the local edge computing devices of the substation to build an automated video analysis pipeline, achieving real-time detection and on-site alarms; S5, Closed-Loop Verification and Iterative Optimization: The system continuously collects feedback during actual operation, and continuously learns and iteratively upgrades the model through manual review of cases, so that the detection capability continues to evolve in practice and becomes more accurate with use.

2. The anomaly detection method for video surveillance equipment in an AI-based substation according to claim 1, characterized in that, Step S1 includes the following steps: S11, Defining the object to be tested and the type of abnormality: S111, Testing equipment: transformers, circuit breakers, disconnect switches, voltage / current transformers, surge arresters, insulators, control cabinets; S112, Types of abnormalities: Appearance defects: damage, corrosion, oil leakage, component detachment, foreign objects hanging; Status abnormalities: incorrect opening / closing position of switch, abnormal indicator lights; Instrument abnormalities: pointer / digital meter readings exceeding limits, unclear, obstructed; Thermal defects: Abnormal temperature rise at electrical connection points and in localized areas of the equipment body; S12, System architecture design: End side: Deploy high-definition visible light cameras, infrared thermal imagers, and dual-spectrum PTZ cameras on the equipment side to be responsible for high-quality video and image data acquisition; Edge computing devices are deployed locally at substations to carry core AI algorithm models, enabling real-time video stream analysis, anomaly detection, and local alarms. Cloud computing devices are deployed at the monitoring center to aggregate data from multiple stations, centrally train and optimize models, manage alarms globally, generate reports, and remotely distribute models.

3. The anomaly detection method for video surveillance equipment in an AI-based substation according to claim 1, characterized in that, Step S2 includes the following steps: S21, Multispectral data acquisition: S211, Simultaneously acquire visible light images and infrared thermal images of the same device, and perform temporal and spatial registration to ensure that the same physical point corresponds in the two images; S212, Set different acquisition frequencies for different detection targets; S22, Construct a high-quality dataset: S221, For visible light images, use annotation tools to mark the positions of device components, dashboards, and indicator lights, and mark their normal / abnormal states; S222, For infrared images, mark abnormal heating areas; S223, Data augmentation: Expand the dataset by rotating, scaling, adding noise, and simulating different lighting and weather conditions to improve the robustness of the model.

4. The anomaly detection method for video surveillance equipment in an AI-based substation according to claim 1, characterized in that, Step S3 is performed according to the following table:

5. The anomaly detection method for video surveillance equipment in an AI-based substation according to claim 4, characterized in that: The target detection model uses YOLOv8 and Faster R-CNN.

6. The anomaly detection method for video surveillance equipment in an AI-based substation according to claim 4, characterized in that: The image segmentation uses U-Net.

7. The anomaly detection method for video surveillance equipment in an AI-based substation according to claim 4, characterized in that: The image classification uses ResNet.

8. The anomaly detection method for video surveillance equipment in an AI-based substation according to claim 1, characterized in that: The anomaly detection model employs an autoencoder.

9. The anomaly detection method for video surveillance equipment in an AI-based substation according to claim 1, characterized in that, Step S4 includes the following steps: S41, Model lightweighting and optimization: The trained model is compressed using pruning, quantization, and knowledge distillation techniques to adapt to the limited computing power of edge computing devices; S42, Building a video analysis pipeline: S421, Deploying the algorithm pipeline on the edge computing device: Video stream access → Keyframe extraction → Multi-module parallel analysis → Result fusion and decision-making → Generation of alarms / reports; S422, Switching the alarm rule engine: Setting up multi-level alarms, including early warning, alarm, and emergency, and supporting multiple alarm push methods, including sound and light, SMS, and platform pop-ups.

10. The anomaly detection method for video surveillance equipment in an AI-based substation according to claim 1, characterized in that, Step S5 includes the following steps: S51, Online testing and feedback: The system is tested and false alarms and missed alarms are collected to form a feedback loop; S52, Continuous model learning: New abnormal samples verified by manual verification are sent back to the cloud training platform to perform incremental training or version iteration on the model, continuously improving the accuracy of the model in real complex environments; S53, Performance evaluation: The core indicators of the system are continuously monitored, including accuracy, recall, false alarm rate and average processing time, to ensure that it meets the practical requirements of substation operation and maintenance.

Images