User-side power safety hazard diagnosis system and method based on multi-channel communication

Abstract

This application discloses a user-side power safety diagnostic system and method based on multi-channel communication, belonging to the field of power equipment operation and maintenance and intelligent diagnostic technology. The system includes a multi-modal acquisition module, a time synchronization module, an edge processing module, a multi-channel communication module, a caching and retransmission module, and a cloud diagnostic module. These modules work together to form an architecture of real-time processing on the end side, in-depth analysis on the cloud, and multi-link collaborative transmission. The system achieves full-process diagnosis through multi-modal data acquisition and time synchronization, edge preprocessing and lightweight inference, multi-channel dynamic transmission, data caching and retransmission, and in-depth cloud diagnostics. This invention adopts a multi-channel redundant communication and dynamic switching strategy, combined with end-cloud model collaborative inference, to ensure stable and complete data transmission in complex scenarios, significantly improving the accuracy of hidden danger identification and real-time diagnostics. It can output fault type, causal chain, and risk level assessment, and is applicable to power equipment safety diagnosis in multiple scenarios such as substations and transmission lines.

Description

Technical Field

[0001] This application belongs to the field of power equipment inspection communication and intelligent diagnosis technology, and relates to a user-side power safety diagnosis system and method based on multi-channel communication. Background Technology

[0002] During power line inspections, power equipment testing, and energy system operation, it is necessary to collect multimodal data such as video streams, image frames, voice samples, electrical sensing data, and temperature information to comprehensively understand the operating status of power equipment. However, existing power line inspection systems generally suffer from the following problems: Single communication link: Most devices only use a single 4G or WiFi for data transmission. In complex scenarios such as underground shafts, shielded environments of electrical cabinets, and mountainous areas with weak coverage, communication interruptions or severe packet loss are likely to occur, leading to inspection failures.

[0003] Multimodal data is not processed in a hierarchical manner: Video data has high bandwidth requirements, while image and voice data are relatively lightweight. However, existing solutions often transmit all data to the cloud in a unified manner, which can easily cause network congestion and affect transmission efficiency.

[0004] Lack of intelligent processing capabilities on the edge: Traditional solutions rely too much on cloud models, requiring all raw data to be sent to the cloud for analysis, resulting in high network usage and high data transmission latency, which cannot meet the needs of real-time inspection.

[0005] The aforementioned problems make it difficult for existing technologies to guarantee the data continuity and real-time reliability of multimodal inspection systems in environments with weak communication or interference, which seriously affects the timely detection and handling of power safety hazards. Therefore, it is necessary to design a power safety hazard diagnosis system that can solve the above problems.

[0006] Invention / Invention Content To address the problems of existing technologies, this invention provides a user-side power safety diagnostic system based on multi-channel communication, including a multi-modal acquisition module, a time synchronization module, an edge processing module, a multi-channel communication module, a buffering and retransmission module, and a cloud diagnostic module. Each module works collaboratively through a data interaction link. The multimodal acquisition module is used to collect multimodal data of power equipment and its surrounding environment through various inspection devices. The multimodal data includes video data, image data, acoustic signal data, electrical parameter data, and environmental parameter data. The time synchronization module is used to uniformly synchronize the system clocks of each acquisition device in the multimodal acquisition module, and control the acquisition clock error of each device to not exceed a preset threshold, thereby ensuring the time synchronization of the acquired data. The edge processing module is deployed on the local data processing terminal and is used to perform preprocessing operations and lightweight model inference operations on the raw data collected by the multimodal acquisition module to generate structured abnormal events. The multi-channel communication module is configured with multiple types of communication links and is equipped with a link status detection unit, which is used to execute link selection strategy or link degradation strategy according to the link status detection result to achieve stable data transmission. The caching and retransmission module is used to store the original data and inference results when the communication link is abnormal, and to perform delayed retransmission in the order of timestamps after the link is restored. The cloud-based diagnostic module is deployed on the power operation and maintenance cloud server and is used to perform in-depth diagnostic operations on the uploaded data, outputting fault type, risk level and maintenance strategy.

[0007] Furthermore, the multimodal acquisition module includes any one or more of the following: a drone, a clamp power meter, an acoustic imager, an infrared imager, a thermometer and hygrometer, a multimeter, a clamp meter, a power meter, a transformer ratio tester, and VR glasses. The drone is used to collect image data, infrared thermal image data, and spatial positioning data of high-altitude or long-distance power equipment. The clamp-on power meter is used for non-contact acquisition of line current data, power factor data, and harmonic characteristic data; The acoustic imager is used to generate sound source imaging data through an array of microphones to achieve partial discharge localization. The infrared imager is used to collect surface temperature distribution data of power equipment and identify hot spots and overload conditions of the equipment. The thermometer and hygrometer are used to collect ambient temperature and humidity data. The multimeter is used to collect basic electrical quantity data and verify the continuity of the equipment and the correctness of the wiring; the clamp meter is used to perform high-frequency inspection current monitoring and collect high-frequency current data. The power meter is used to perform power quality analysis and output power quality related parameters. The transformer ratio tester is used to detect the winding ratio data and magnetic flux characteristic data of transformer equipment. The VR glasses are used to collect first-person perspective image data, eye movement information data, or voice data during manual inspections, and interact with the edge processing module through a local network.

[0008] Furthermore, the time synchronization module achieves time alignment through the NTP protocol or a hardware synchronization mechanism, and the preset threshold is set according to the requirements of multimodal data fusion analysis.

[0009] Furthermore, the edge processing module includes a multimodal preprocessing unit, a lightweight inference unit, a cache and timestamp synchronization unit, and an abnormal event triggering unit; The multimodal preprocessing unit is used to perform image decoding, temperature matrix filtering, acoustic point cloud aggregation, data denoising and normalization on the raw data; The lightweight inference unit deploys a lightweight intelligent model to perform real-time inference on the preprocessed data and outputs structured inference results containing category labels, confidence levels, and anomaly levels. The cache and timestamp synchronization unit is used to cache the original data and inference results locally and align them with a unified timestamp. The abnormal event triggering unit is used to continuously monitor the confidence level output by the lightweight inference unit. When the confidence level reaches a preset threshold or the abnormal duration exceeds a preset duration, it sends an emergency upload command to the multi-channel communication module.

[0010] Furthermore, the lightweight intelligent model includes any one of MobileNet, YOLO-nano, and TinyTransformer, and the real-time inference includes image target detection, acoustic anomaly recognition, temperature and humidity threshold exceeding judgment, and current waveform sudden change judgment.

[0011] Furthermore, the multi-channel communication module includes multiple types of communication links such as 4G link, WiFi module, Bluetooth module, Ethernet wired interface and Type-C wired interface; The link status detection unit is used to periodically monitor the signal strength, network latency, packet loss rate, and available bandwidth of each communication link; The link selection strategy is executed based on a preset link priority, which is Ethernet > Type-C > WiFi > Dual 4G > BT; The link degradation strategy is triggered when the main link meets the preset threshold conditions. The preset threshold conditions include at least one of the following: RSSI≤-90dBm, packet loss rate≥3%, upload latency≥300ms, TCP handshake failures≥2 consecutive times, and uplink bandwidth<minimum video encoding requirements. The 4G link is used for video streaming or high-priority data transmission during normal inspection processes; The WiFi module is used to enable wireless data interaction between the VR glasses and the local data processing terminal. The local data processing terminal acts as a WiFi hotspot, and the VR glasses act as a client connecting to the hotspot. The Bluetooth module is used for low-power data uploading from various acquisition terminals. The local data processing terminal acts as the master device, and multiple sensors or detection instruments act as slave devices, uploading the acquisition results in broadcast or short connection mode. The Type-C wired interface is used for high-speed wired download of video and image data after the drone inspection is completed.

[0012] Furthermore, the cloud-based diagnostic module includes a high-precision defect identification unit, a multimodal power feature fusion unit, a fault knowledge graph unit, a power fault mode matching unit, an operational risk quantification unit, and a maintenance strategy feedback unit. The high-precision defect identification unit is equipped with a Transformer or visual attention network to accurately locate and identify equipment defects in image data, infrared thermal imaging data and acoustic data. The multimodal power feature fusion unit uses a device-level feature fusion method to integrate different types of collected data at the device level and mine the correlation between the data. The fault knowledge graph unit stores typical fault types of power equipment, industry standard safety thresholds, the relationship between operating load and fault evolution, and environmental factor influence models, which are used for tracing and explaining the causes of faults. The power fault mode matching unit is used to compare the fused feature data with the preset fault mode and output the specific fault type and cause chain. The operational risk quantification unit is used to comprehensively assess the failure risk in four levels, from L1 to L4, based on the frequency of failure occurrence, equipment type, load level, environmental factors, and historical degradation data. The maintenance strategy feedback unit is used to output maintenance suggestions based on the deep diagnostic results and to send control commands to the multimodal acquisition module to adjust the data acquisition parameters or acquisition method.

[0013] Furthermore, the industry standards mentioned include GB / T, DL / T, and IEC standards. Level L1 indicates continued operation, with periodic reviews recommended; Level L2 indicates potential accident risks, with maintenance recommended; Level L3 indicates a serious fault, requiring immediate shutdown or replacement; and Level L4 indicates a threat to safe power supply, requiring emergency handling and evidence preservation.

[0014] Furthermore, the system is applicable to inspection and safety hazard diagnosis scenarios of substations, transmission lines, power cabinets, distribution ring network cabinets, main transformers and switchgear, and can be implemented in the form of single-machine integration or split deployment.

[0015] A method for diagnosing user-side power safety hazards based on multi-channel communication, characterized by comprising: S1: Multimodal data acquisition and time synchronization: Through various inspection devices of the multimodal acquisition module, video data, image data, acoustic signal data, electrical parameter data and environmental parameter data of power equipment and its surrounding environment are collected; the time synchronization module performs a unified timestamp alignment operation on the data collected by each acquisition device to control the timestamp error of the output data of each acquisition device to not exceed the preset threshold. S2: Edge Preprocessing and Lightweight Inference: The edge processing module performs preprocessing operations and lightweight model inference operations on the collected raw data to generate structured anomaly events containing category labels, confidence levels, and anomaly levels. S3: Multi-channel data transmission: The multi-channel communication module monitors the operating status of various types of communication links through the link status detection unit, and executes link selection strategy or link degradation strategy according to the detection results, so as to stably transmit structured abnormal events and related raw data to the cloud diagnostic module. S4: Data caching and delayed retransmission: When the communication link is abnormal, the original data and inference results are stored through the caching and retransmission module; after the communication link is restored to normal, the delayed retransmission operation is automatically executed in the order of timestamps, and the cached data is uploaded to the cloud diagnostic module. S5: Cloud-based in-depth diagnostics and results output: The cloud-based diagnostic module performs in-depth diagnostic operations on the uploaded data and outputs the fault type, risk level, and maintenance strategy of the power equipment.

[0016] The beneficial effects that this application can produce include: (1) This application adopts a multi-channel communication link design, combined with link status monitoring and dynamic degradation strategy, which can maintain reliable communication in complex scenarios such as strong interference and weak coverage, and effectively avoid inspection failure due to communication interruption; (2) This application implements hierarchical transmission based on data bandwidth characteristics, reduces redundant data uploads by lightweight inference on the end side, reduces network congestion, and prioritizes the transmission of abnormal events to ensure timely reporting of emergency risks. (3) This application uses a caching and retransmission module to achieve local data caching in the case of weak network or disconnection, and automatically retransmits the data after the link is restored, thus completely solving the problem of data loss. (4) This application adopts a collaborative mode of real-time processing on the end side and in-depth analysis on the cloud. The end side responds quickly to simple anomalies, and the cloud accurately diagnoses complex faults. The interpretation of the diagnostic results is improved by combining fault knowledge graphs, and the quantitative classification of operational risks provides a scientific basis for operation and maintenance decisions. Attached Figure Description

[0017] Figure 1 This is a diagram of the overall system architecture of the present invention; Figure 2 This is a flowchart of the multimodal data acquisition and hierarchical transmission process of the present invention; Figure 3 This is a flowchart of the user-side power safety hazard diagnosis method based on multi-channel communication according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1-3 This invention provides a user-side power safety diagnostic system based on multi-channel communication, including a multi-modal acquisition module, a time synchronization module, an edge processing module, a multi-channel communication module, a buffer and retransmission module, and a cloud diagnostic module. Each module works collaboratively through a data interaction link. The multimodal acquisition module is used to collect multimodal data of power equipment and its surrounding environment through various inspection devices. The multimodal data includes video data, image data, acoustic signal data, electrical parameter data, and environmental parameter data. The time synchronization module is used to uniformly synchronize the system clocks of each acquisition device in the multimodal acquisition module, and control the acquisition clock error of each device to not exceed a preset threshold, thereby ensuring the time synchronization of the acquired data. The edge processing module is deployed on the local data processing terminal and is used to perform preprocessing operations and lightweight model inference operations on the raw data collected by the multimodal acquisition module to generate structured abnormal events. The multi-channel communication module is configured with multiple types of communication links and is equipped with a link status detection unit, which is used to execute link selection strategy or link degradation strategy according to the link status detection result to achieve stable data transmission. It should be noted that the multi-channel communication module is the core data transmission module of the power safety diagnostic system. It integrates multiple types of communication link hardware and has a built-in link status detection unit. Through the entire process logic of periodic quantitative monitoring of link status, intelligent policy determination, and adaptive link switching / parameter adjustment, it executes link selection or link degradation strategies to achieve stable, continuous, and real-time transmission of multimodal data in complex power inspection scenarios. At the same time, it works with a buffering and retransmission module to ensure the integrity of data transmission. Its specific implementation is as follows: Implementation of Multi-type Communication Link Configuration and Link Status Detection Unit The multi-channel communication module is configured with various communication links, including an Ethernet wired interface, a Type-C wired interface, a WiFi wireless communication module, dual 4G wireless communication links, and a BLE Bluetooth module. Each communication link is deployed independently and works in parallel to adapt to the differentiated transmission requirements of power inspection scenarios: the Ethernet wired interface is used for high-speed and stable data transmission in fixed scenarios such as substations; the Type-C wired interface is used for high-speed wired download of video / image data after drone inspection; the WiFi module is used to realize local wireless data interaction between VR glasses and the local data processing terminal, with the local data processing terminal acting as a WiFi hotspot and the VR glasses acting as a client; the dual 4G links are used for wireless transmission of video streams and high-priority structured abnormal event data during normal inspections; and the BLE Bluetooth module is used for low-power, small-volume data uploading from various acquisition terminals, with the local data processing terminal acting as the master device and each sensor / detection instrument acting as a slave device, completing data uploading in broadcast or short-connection mode.

[0020] The link status detection unit is a built-in functional unit of the multi-channel communication module, used to periodically and comprehensively monitor the real-time operating status of each communication link. The monitoring period can be configured as needed according to the inspection scenario. The monitoring parameters include the signal strength (RSSI), network latency (RTT), data packet loss rate, available uplink / downlink bandwidth, and TCP handshake success rate of each link. In addition, for wireless links such as WiFi, 4G, and BLE, the signal interference strength and base station / hotspot connection status are monitored. For wired links such as Ethernet and Type-C, the hardware connectivity status is monitored.

[0021] The specific monitoring process of the link status detection unit is as follows: Full data collection of the above parameters for each communication link is performed according to a preset cycle. Normalization and standardization processing is applied to the collected multi-dimensional parameters with different dimensions to eliminate dimensional interference. Failed data collection values ​​due to communication interruptions or hardware failures are marked as anomalies. Moving average filtering is applied to parameters with large fluctuations, such as signal strength and packet loss rate, to ensure the validity and accuracy of the monitoring data. The standardized monitoring parameters are compared with the preset normal operation threshold of the link. A binary judgment result of normal / abnormal is output for each link. Abnormal links are labeled with specific anomaly types. The judgment result is encapsulated in the format of a unique device identifier + unified timestamp and transmitted in real time to the strategy execution core of the multi-channel communication module, providing a quantitative basis for the execution of link selection and degradation strategies.

[0022] Implementation of link selection strategy The link selection strategy is the basic transmission strategy of the multi-channel communication module. It is automatically executed based on a combination of preset link priority and real-time judgment results from the link status detection unit. The core principle is to prioritize links with high priority and normal operating status as the primary transmission link, ensuring efficient and stable data transmission. The specific execution steps are as follows: The multi-channel communication module has a pre-stored fixed link priority rule, with the priority from high to low as follows: Ethernet wired interface > Type-C wired interface > WiFi module > dual 4G links > BLE Bluetooth module. This priority is set based on the transmission bandwidth, stability, and transmission delay characteristics of each link, adapting to the transmission needs of various power inspection scenarios. The strategy execution core receives the real-time judgment results from the link status detection unit, performs dual screening and sorting of all communication links based on status and priority, removes links with abnormal status, and sorts the remaining links with normal status from high to low according to the above-mentioned preset priority to form an available link priority queue. This queue is refreshed in real time as the link status detection results are updated. The highest priority link is selected from the available link priority queue as the current main transmission link. The multi-channel communication module sends a transmission enable command to the main transmission link and distributes the structured inference results and raw data output by the edge processing module to the main transmission link for transmission according to the data transmission priority. Among them, the emergency upload data sent by the abnormal event triggering unit has the highest transmission priority and is scheduled for transmission first. Ordinary inspection data is transmitted in the order of the queue. If the link status detection unit detects the addition of a normal link with a higher priority (such as the inspection equipment connecting to the Ethernet wired interface after entering the substation), the policy execution core immediately refreshes the available link priority queue and automatically switches to the higher priority link as the new main transmission link, realizing the link's seamless dynamic upgrade. At the same time, it restores the link's full-configuration transmission parameters to ensure high-speed transmission of all data.

[0023] Implementation of link degradation strategy The link degradation strategy is a fault tolerance guarantee strategy for the multi-channel communication module. When the current primary transmission link is determined to be abnormal and there is no higher-priority normal link in the available link priority queue, it is automatically triggered by the strategy execution core. Through suboptimal link switching and adaptive adjustment of transmission parameters, it ensures uninterrupted data transmission while taking into account the transmission priority of core data. The specific execution steps are as follows: Degradation strategy trigger determination: The strategy execution core monitors the status determination result of the current main transmission link in real time. If the monitoring parameters of the main transmission link meet any of the preset threshold conditions for link anomalies, and there are no normal links with higher priority in the available link priority queue, the link degradation strategy is immediately triggered, and the fault-tolerant transmission mode is entered. The preset threshold conditions for link anomalies include: RSSI≤-90dBm, data packet loss rate≥3%, upload latency≥300ms, TCP handshake failures≥2 consecutive times, and uplink bandwidth<minimum video encoding requirements. Second-best primary transmission link selection and switching: The policy execution core selects the highest priority normal link from the available link priority queue as the degraded primary transmission link. The multi-channel communication module sends a transmission enable command to the degraded primary transmission link and at the same time disables the transmission function of the original abnormal primary transmission link to achieve seamless automatic switching of links. If all wireless links are determined to be abnormal, only the BLE Bluetooth module is kept in working state to ensure low-power transmission of core structured data. Transmission parameters adaptive adjustment: The multi-channel communication module adaptively adjusts the data transmission parameters according to the characteristics of the bandwidth, transmission delay, signal stability and other characteristics of the main transmission link after the downgrade, and matches the transmission capacity of the main transmission link after the downgrade, so as to avoid transmission congestion and data loss due to insufficient link capacity. The specific adjustment rules are as follows: (1) If downgraded to WiFi module or dual 4G link, the encoding bit rate and resolution of video / infrared image data are automatically reduced, only key frame images are transmitted, large volume original video data is temporarily suspended, and structured abnormal event data containing defect category, confidence level, device ID and abnormal level are transmitted first; (2) If downgraded to BLE Bluetooth module, large volume original data such as images and videos are stopped, and only the core structured reasoning results of power equipment fault diagnosis are transmitted to ensure that the key information of power equipment safety hazards is not lost; (3) In all downgrade scenarios, the data to be transmitted is scheduled and transmitted in a unified timestamp order to ensure the consistency of the time sequence of uploaded data and provide a basis for the multimodal data fusion analysis of the cloud diagnosis module. Link recovery and automatic upgrade: The link status detection unit continuously monitors the operating status of the original abnormal main transmission link. If all its monitored parameters recover to the preset normal operating threshold range, and this normal state remains stable for a preset duration, it is determined that the original abnormal main transmission link has recovered to normal. The policy execution core immediately adds the link back to the available link priority queue. If its priority is higher than the current degraded main transmission link, it automatically switches back to the link as the new main transmission link, realizing a seamless automatic upgrade of the link, while restoring the link's full-configuration transmission parameters. The multi-channel communication module synchronously sends a retransmission command to the buffer and retransmission module, which retransmits the large volume of raw data temporarily stored during the link abnormality to the cloud diagnostic module in timestamp order, ensuring the integrity of data transmission.

[0024] Link strategy and its coordination with other modules in the system The link selection and link degradation strategies of the multi-channel communication module are not executed independently, but are deeply integrated with the system's edge processing module, buffering and retransmission module to form a closed-loop collaboration of data acquisition, edge processing, link transmission, and buffering and retransmission, further ensuring the stability, real-time performance, and integrity of data transmission. Linked with the abnormal event triggering unit of the edge processing module: When the abnormal event triggering unit detects that the output confidence of the lightweight inference unit reaches a preset threshold (0.85) or the abnormal duration exceeds a preset duration, and sends an emergency upload command, the multi-channel communication module marks the emergency data as the highest transmission priority. Regardless of whether the current link selection strategy or link degradation strategy is executed, the emergency data transmission is prioritized in the main transmission link to ensure that emergency information on power equipment safety hazards is reported to the cloud diagnostic module as soon as possible. In conjunction with the caching and retransmission module: When the link degrades to a low-bandwidth link and cannot transmit large volumes of raw data, the multi-channel communication module forwards this part of the data to the caching and retransmission module for local encryption and caching. After the link recovers and completes automatic upgrade, a retransmission command is triggered, and the caching and retransmission module automatically retransmits the temporarily stored data to the cloud diagnostic module in the order of timestamps, completely solving the problem of data loss caused by link anomalies in complex scenarios.

[0025] The caching and retransmission module is used to store the original data and inference results when the communication link is abnormal, and to perform delayed retransmission in the order of timestamps after the link is restored. The cloud-based diagnostic module is deployed on the power operation and maintenance cloud server and is used to perform in-depth diagnostic operations on the uploaded data, outputting fault type, risk level and maintenance strategy.

[0026] Furthermore, the multimodal acquisition module includes any one or more of the following: a drone, a clamp power meter, an acoustic imager, an infrared imager, a thermometer and hygrometer, a multimeter, a clamp meter, a power meter, a transformer ratio tester, and VR glasses. The drone is used to collect image data, infrared thermal image data, and spatial positioning data of high-altitude or long-distance power equipment. The clamp-on power meter is used for non-contact acquisition of line current data, power factor data, and harmonic characteristic data; The acoustic imager is used to generate sound source imaging data through an array of microphones to achieve partial discharge localization. The infrared imager is used to collect surface temperature distribution data of power equipment and identify hot spots and overload conditions of the equipment. The thermometer and hygrometer are used to collect ambient temperature and humidity data. The multimeter is used to collect basic electrical quantity data and verify the continuity of the equipment and the correctness of the wiring; the clamp meter is used to perform high-frequency inspection current monitoring and collect high-frequency current data. The power meter is used to perform power quality analysis and output power quality related parameters. The transformer ratio tester is used to detect the winding ratio data and magnetic flux characteristic data of transformer equipment. The VR glasses are used to collect first-person perspective image data, eye movement information data, or voice data during manual inspections, and interact with the edge processing module through a local network.

[0027] Furthermore, the time synchronization module achieves time alignment through the NTP protocol or a hardware synchronization mechanism, and the preset threshold is set according to the requirements of multimodal data fusion analysis.

[0028] It should be noted that the time synchronization module uses the NTP protocol or hardware synchronization mechanism to perform unified timestamp alignment on the data collected by each acquisition device in the multimodal acquisition module, and controls the timestamp error of the output data of each acquisition device to not exceed 10ms, so as to ensure the temporal consistency of multimodal data of different acquisition channels and different types, and provide a foundation for the fusion analysis and correlation reasoning of multimodal data.

[0029] Furthermore, the edge processing module includes a multimodal preprocessing unit, a lightweight inference unit, a cache and timestamp synchronization unit, and an abnormal event triggering unit; The multimodal preprocessing unit is used to perform image decoding, temperature matrix filtering, acoustic point cloud aggregation, data denoising and normalization on the raw data; The lightweight inference unit deploys a lightweight intelligent model to perform real-time inference on the preprocessed data and outputs structured inference results containing category labels, confidence levels, and anomaly levels. The inference covers image target detection (such as defect recognition), acoustic anomaly recognition (such as partial discharge feature extraction), temperature and humidity threshold exceeding judgment, and current waveform sudden change detection. The cache and timestamp synchronization unit is used to cache the original data and inference results locally and align them with a unified timestamp. The abnormal event triggering unit is used to continuously monitor the confidence level output by the lightweight inference unit. When the confidence level reaches a preset threshold or the abnormal duration exceeds a preset duration, it sends an emergency upload command to the multi-channel communication module.

[0030] Furthermore, the lightweight intelligent model includes any one of the following: a lightweight convolutional neural network (MobileNet), an object detection algorithm (YOLO-nano), and a deep learning model based on a self-attention mechanism (TinyTransformer). The real-time inference includes image object detection, acoustic anomaly recognition, temperature and humidity threshold exceeding judgment, and current waveform change judgment, wherein the preset threshold is 0.85. It should be noted that the process of identifying abnormal transformer joint temperatures based on the YOLO-nano model in this lightweight inference unit is completed using the edge processing module of the power safety diagnostic system. This edge processing module is deployed on the local data processing terminal and is signal-connected to the multi-modal acquisition module, time synchronization module, and multi-channel communication module. The lightweight inference unit identifies abnormal transformer joint temperatures using the YOLO-nano model, specifically including: The data acquisition and time synchronization multimodal acquisition module acquires infrared thermal image data of the transformer equipment through an infrared imager, and simultaneously acquires the ambient temperature and humidity parameters around the transformer equipment through a thermometer and hygrometer. The infrared thermal image data includes the surface temperature distribution characteristics of the transformer joints. The time synchronization module performs a unified timestamp alignment operation on the infrared thermal image data and ambient temperature and humidity parameters through the NTP protocol or hardware synchronization mechanism, controls the timestamp error to not exceed a preset threshold, forms the original acquisition data with consistent time sequence, and transmits it to the edge processing module.

[0031] The multimodal data preprocessing edge processing module performs preprocessing operations on the raw acquired data, including image decoding and temperature matrix filtering of the infrared thermal image data to eliminate temperature noise caused by environmental interference and equipment vibration, and restore the true temperature distribution characteristics of the transformer joint; at the same time, the infrared thermal image data is normalized to adjust to the input resolution adapted to the YOLO-nano model, and the temperature value of the infrared thermal image and the ambient temperature and humidity parameters are mapped to the 0-1 range to complete the normalization, and the standardized processed data is output.

[0032] The lightweight inference unit of the YOLO-nano model's lightweight inference and target localization edge processing module calls the pre-trained YOLO-nano lightweight target detection model to perform real-time inference on the standardized data: 3.1 The YOLO-nano model extracts spatial features and temperature anomaly visual features of transformer joints through its lightweight backbone network. The spatial features include the location, shape, and contour features of the transformer joints, and the temperature anomaly visual features include high-brightness pixel clusters in the transformer joint area of ​​the infrared thermal image and temperature color difference features with the normal area; 3.2 Based on the preset anchor frame size of the transformer joint in the power equipment inspection scenario, anchor frame matching and target prediction are performed on the extracted features, and the bounding box coordinates and category confidence of the transformer joints are output to achieve accurate localization of all transformer joint targets in the image and distinguish between normal temperature joints and temperature anomaly joints; 3.3 Non-maximum suppression is performed to remove duplicate bounding box prediction results and retain the optimal localization box of each transformer joint target to ensure the uniqueness and accuracy of transformer joint localization.

[0033] The lightweight inference unit for temperature anomaly quantification and confidence output combines the temperature rise safety thresholds for transformer joints in the GB / T and DL / T standards of the power industry to perform temperature anomaly quantification on the located transformer joints: it extracts the average / maximum temperature of the transformer joint area within the bounding box and compares it with the ambient temperature. If the transformer joint temperature is greater than the ambient temperature +25~35℃, it is determined to be a temperature anomaly joint. At the same time, the YOLO-nano model outputs the identification confidence of the "temperature anomaly" of the transformer joint. The confidence is determined by combining the accuracy of the transformer joint location and the significance of the temperature anomaly.

[0034] 5.1 The lightweight inference unit organizes the inference results into structured inference data. The structured inference data includes at least category labels (transformer joint temperature anomaly), transformer joint bounding box pixel coordinates, temperature anomaly identification confidence level, actual temperature value of the abnormal joint, and environmental temperature and humidity parameters, and is associated with a unified timestamp. 5.2 The edge processing module's caching and timestamp synchronization unit caches the structured inference data and the corresponding raw infrared thermal image data locally. 5.3 The edge processing module's abnormal event triggering unit continuously monitors the temperature anomaly identification confidence level. When the confidence level is ≥ the system preset threshold (0.85) or the duration of the temperature anomaly exceeds the preset duration, it immediately sends an emergency upload command to the multi-channel communication module. The multi-channel communication module selects the optimal communication link based on the link status detection results to prioritize uploading the structured inference data and the corresponding infrared thermal image data to the cloud diagnostic module.

[0035] The YOLO-nano model described in the model scene adaptation optimization is a dedicated model pre-trained and fine-tuned for power equipment inspection scenarios. Its training dataset contains infrared thermal image data of transformer joint temperature anomalies for different types of transformers, different operating conditions, and different degrees of anomalies. At the same time, the model is optimized for small target detection layer, network pruning and channel compression are performed, retaining only the core feature extraction layer for transformer joint detection, eliminating redundant network structures, adapting to the edge computing power requirements of local data processing terminals, and ensuring real-time inference.

[0036] Furthermore, the multi-channel communication module includes multiple types of communication links such as 4G link, WiFi module, Bluetooth module, Ethernet wired interface and Type-C wired interface; The link status detection unit is used to periodically monitor the signal strength (RSSI), network latency (RTT), packet loss rate, and available bandwidth of each communication link; and to provide real-time feedback on the link health status. The link selection strategy is executed based on a preset link priority, which is Ethernet > Type-C > WiFi > Dual 4G > BT; The link degradation strategy is triggered when the main link meets preset threshold conditions. The preset threshold conditions include at least one of the following: RSSI≤-90dBm, packet loss rate≥3%, upload latency≥300ms, TCP handshake failures≥2 consecutive times, and uplink bandwidth<minimum video encoding requirements.

[0037] It should be noted that the link selection and degradation strategy is as follows: The link selection strategy is executed based on the preset link priority (Ethernet > Type-C > WiFi > Dual 4G > BT), prioritizing links with stable transmission and sufficient bandwidth; when the primary link meets the preset threshold conditions (RSSI≤-90dBm, packet loss rate≥3%, upload latency≥300ms, TCP handshake failure ≥2 consecutive times or uplink bandwidth < minimum video encoding requirements), the link degradation strategy is triggered, automatically switching to the backup link and adjusting the data transmission parameters (such as reducing the video bitrate and only transmitting structured inference results) to ensure the continuity and stability of data transmission; Furthermore, the 4G link is used for video streaming or high-priority data transmission during normal inspection processes; The WiFi module is used to enable wireless data interaction between the VR glasses and the local data processing terminal. The local data processing terminal acts as a WiFi hotspot, and the VR glasses act as a client connecting to the hotspot. The Bluetooth module is used for low-power data uploading from various acquisition terminals. The local data processing terminal acts as the master device, and multiple sensors or detection instruments act as slave devices, uploading the acquisition results in broadcast or short connection mode. The Type-C wired interface is used for high-speed wired download of video and image data after the drone inspection is completed.

[0038] Furthermore, the cloud-based diagnostic module includes a high-precision defect identification unit, a multimodal power feature fusion unit, a fault knowledge graph unit, a power fault mode matching unit, an operational risk quantification unit, and a maintenance strategy feedback unit. The high-precision defect identification unit is equipped with a Transformer or visual attention network to accurately locate and identify equipment defects in image data, infrared thermal imaging data, and acoustic data; it can identify small defects of 2-5mm (such as joint erosion and slight contamination of insulators). The multimodal power feature fusion unit uses a device-level feature fusion method to integrate different types of collected data at the device level, and mines the correlation between data (such as the correlation between temperature rise and partial discharge signal enhancement), avoiding information loss caused by simple data splicing. It should be noted that the multimodal power feature fusion unit is the core data processing unit of the cloud diagnostic module. In response to the multi-source data correlation analysis needs of power equipment operation and maintenance scenarios, it adopts the device-level feature fusion method to integrate and mine the features of different types of collected data, break the information barriers of single-dimensional data, mine the causal correlation and coupling relationship between cross-type data features, and output a structured and integrated device-level fusion feature set. This provides comprehensive and accurate feature support for subsequent fault mode matching and cause tracing, and avoids the problems of information loss and feature misalignment caused by simple data splicing.

[0039] The device-level feature fusion process of this unit takes a single power device as the sole core associated entity and fusion benchmark, and relies on the unified timestamp alignment results of the time synchronization module to complete the cross-dimensional feature integration of multiple types of collected data. Specifically, it includes the following steps: Single device feature extraction and classification Using the unique identifier of power equipment as the filtering condition, all types of feature data under the same equipment and the same timestamp are extracted. The feature data is synchronously output by the high-precision defect identification unit and the multi-modal acquisition module, and is classified by data type into image defect features, infrared temperature rise features, acoustic partial discharge features, electrical parameter features, and environmental parameter features. Among them, image defect features include defect semantic labels, location coordinate boxes, and identification confidence; infrared temperature rise features include the highest / average temperature of the equipment area, temperature rise trend, and three-phase temperature difference; acoustic partial discharge features include partial discharge type, signal amplitude / frequency, and vibration spectrum features; electrical parameter features include core operating parameters such as three-phase current / voltage, power factor, harmonic content, and leakage current, as well as threshold judgment results; and environmental parameter features include environmental impact parameters such as temperature, humidity, salt spray concentration, and atmospheric pressure around the equipment.

[0040] Feature standardization The multi-type feature data after the above classification are subjected to dimensional standardization and dimensional normalization: all feature values ​​with dimensions are mapped to the 0-1 interval to eliminate the interference of different dimensions such as temperature, current, and amplitude; all types of feature data are uniformly mapped to the preset feature dimension space, missing feature dimensions are padded with 0, and redundant feature dimensions are reduced, finally obtaining a single device multi-type standardized feature vector with the same dimension and no dimensions, ensuring that cross-type features can be correlated and fused for analysis.

[0041] Device-level feature association fusion Based on the operating patterns and fault diagnosis logic of power equipment, a device-specific feature weight model and a cross-type feature correlation matrix are constructed to achieve deep fusion of standardized feature vectors of multiple types for a single device. (1) Assign exclusive weights to features based on the type of power equipment (transformer, insulator, switchgear, busbar, etc.), and give higher weights to features that play a core role in equipment fault diagnosis (such as temperature rise features and electrical parameter features of transformers, partial discharge features and environmental humidity features of insulators). The weight values ​​are dynamically generated by the "equipment type-fault type" association model of the fault knowledge graph unit. (2) Calculate the correlation coefficient between any two types of features, construct a cross-type feature correlation matrix, and explore the causal relationship and coupling relationship between features, such as the correlation between increased ambient humidity and enhanced partial discharge signal of insulators, the correlation between three-phase current imbalance and abnormal temperature rise of transformer windings, and the correlation between increased contact resistance and excessive temperature rise of joints. (3) Based on the feature-specific weights and correlation matrix, weighted summation and correlation feature aggregation are performed on all standardized feature vectors to generate a single device fusion feature vector. This vector not only retains the core state information of each type of data, but also reflects the inherent correlation between cross-type features. It is a comprehensive digital representation of the operating status of a single power device.

[0042] Structured device-level fusion feature set generation The fused single-device feature vector is integrated with the unique identifier of the power equipment, the unified timestamp, the original core feature information of each type, and the feature association conclusion to generate a structured device-level fused feature set. This feature set is a standardized data format that can be directly called by the power fault mode matching unit and the operation risk quantification unit. Its core content includes at least the unique identifier of the equipment, the unified timestamp, the fused feature vector, the description of each type of core feature, and the feature association conclusion.

[0043] The device-level feature fusion method in this unit differs from traditional indiscriminate data splicing and feature fusion. It features the technical characteristics of using devices as anchors, time series as benchmarks, and correlation as the core, ensuring that all fused features point to the same device and the same runtime, effectively avoiding feature confusion across devices and time periods. At the same time, by mining the causal relationships between cross-type features, it provides interpretable correlation evidence for fault cause tracing, breaking the "black box" limitation of pure algorithm diagnosis, and significantly improving the accuracy of subsequent fault mode matching and the scientific nature of fault diagnosis.

[0044] The device-level fusion feature set output by this unit can be adapted to the diagnostic needs of various power equipment such as substations, transmission lines, power cabinets, distribution ring network cabinets, main transformers, and switchgear. It dynamically adjusts feature weights and association rules according to the operating characteristics and fault patterns of different equipment, realizing intelligent and scenario-based fusion analysis of multi-type collected data.

[0045] The fault knowledge graph unit stores typical fault types of power equipment, industry standard safety thresholds, the relationship between operating load and fault evolution, and environmental factor influence models, which are used for tracing and explaining the causes of faults. The power fault mode matching unit is used to compare the fused feature data with the preset fault mode, output the specific fault type and cause chain (such as "temperature rise of disconnector joint and local corrosion leading to poor contact"), and directly generate executable maintenance suggestions. The operational risk quantification unit is used to comprehensively assess the failure risk in four levels, from L1 to L4, based on the frequency of failure occurrence, equipment type, load level, environmental factors, and historical degradation data. The maintenance strategy feedback unit is used to output maintenance suggestions based on the deep diagnostic results and to send control commands to the multimodal acquisition module to adjust data acquisition parameters (such as sampling frequency) or acquisition methods (such as close-range supplementary shooting by drone).

[0046] It should be noted that the cloud-based diagnostic module includes a high-precision defect identification unit, a multimodal power feature fusion unit, a fault knowledge graph unit, a power fault mode matching unit, an operational risk quantification unit, and a maintenance strategy feedback unit. Each unit achieves linkage and collaboration through data interaction, and sequentially completes accurate defect identification, deep feature fusion, fault cause tracing, risk level quantification, and operation and maintenance. The high-precision defect identification unit is configured with a deep learning model cluster to target typical defects in power equipment. It performs specialized identification processing on the uploaded image data, infrared thermal imaging data, and acoustic signal data, specifically including: Image Defect Recognition Subunit: Employing a Transformer network or a deep learning network based on visual attention mechanisms, this subunit accurately locates and semantically labels typical defects in power equipment images. These typical defects include broken conductor strands, corrosion of metal components, oxidation of contact points, loose busbar connections, bushing damage, surge arrester leakage, porcelain insulator breakage, insulator contamination, circuit breaker contact carbonization, and arc-extinguishing chamber ablation. The network model supports pixel-level localization of small defect areas with a minimum size of 2-5 mm (such as joint ablation marks), and outputs image recognition results including defect semantic labels, location coordinate boxes, and recognition confidence scores. Infrared thermal imaging analysis subunit: For different types of power equipment such as transformers, switchgear, cable joints, and busbars, specific temperature rise threshold models and time-series temperature rise evolution models are established. The judgment rules of the temperature rise threshold model include: When the temperature at the equipment connection point is greater than the ambient temperature + ΔTcrit (ΔTcrit ranges from 25 to 35°C), it is judged as a poor contact defect. When the temperature difference between phases of a three-phase device is asymmetrically distributed, it is determined to be an imbalance of phase current or an abnormal load of opposite phases. When the temperature rise of a cable joint deviates from the preset standard curve as the operating load increases, it is determined to be a defect of increased internal impedance. The infrared thermal imaging analysis subunit enables early identification and trend prediction of slowly deteriorating thermal defects by performing thermal migration trajectory analysis and multi-time period temperature rise data modeling. It should be noted that the implementation method of thermal migration trajectory analysis and multi-time period temperature rise data modeling of the infrared thermal imaging analysis subunit is as follows: Based on the equipment-specific temperature rise threshold model and the time-series temperature rise evolution model, the infrared thermal imaging analysis subunit achieves spatial evolution tracking and temporal trend prediction of thermal defects in power equipment through the coordinated execution of thermal migration trajectory analysis and multi-time period temperature rise data modeling. It accurately captures the complete evolution process of thermal defects from their inception to their manifestation, providing spatial and temporal feature support for early warning of slowly deteriorating thermal defects. The specific implementation method is as follows: Implementation process of thermal migration trajectory analysis Thermal migration trajectory analysis focuses on the spatial evolution characteristics of surface temperature distribution in power equipment. By tracking the dynamic changes in the spatial location, area, and temperature gradient of hot spots, it identifies the diffusion and migration patterns of thermal defects, determines the type and deterioration trend of thermal defects, and the specific execution steps are as follows: Pixel-level hotspot localization and feature extraction: After the infrared thermal images of power equipment acquired by the infrared imager are processed by the multimodal preprocessing unit to perform temperature matrix filtering, noise reduction, and normalization, the pixel-level temperature threshold segmentation algorithm is used to extract hotspot areas in the thermal image whose temperature exceeds the environmental reference value. The core spatial features of each hotspot area are output, including: hotspot pixel coordinate set, hotspot geometric center, effective area of ​​hotspot (number of pixels), highest / average temperature in the hotspot area, temperature gradient between the hotspot and the normal area of ​​the equipment, and temperature attenuation coefficient around the hotspot.

[0047] The registration of hot zones in multiple frames of thermal images of the same equipment is based on the unique identifier of the power equipment. For infrared thermal images of the same equipment at different acquisition time periods, the equipment contour feature registration is performed. Based on the fixed geometric features of the equipment body (such as the position of transformer terminals, busbar contour, and switchgear cabinet structure), a spatial registration reference coordinate system is established to eliminate the spatial position deviation of hot zones caused by acquisition angle, slight vibration of equipment, and displacement of inspection equipment, so as to ensure that the spatial position of the same hot spot area in different frames of thermal images can be compared and tracked.

[0048] The thermal migration trajectory quantization construction uses the registered spatial reference coordinate system as the carrier, and takes the geometric center of the hot spot area in each acquisition period as the trajectory node. The nodes are connected in the order of acquisition timestamp to form the thermal defect migration trajectory curve. At the same time, the characteristic change values ​​of the hot area in each period are quantified and calculated, including: geometric center offset distance (pixels / physical size), hot spot area expansion / contraction rate, temperature gradient change rate, and temperature decay coefficient change value, forming a quantitative feature set of thermal migration trajectory.

[0049] The matching of thermal migration features and defect types is based on the association model of thermal defects and thermal migration features of power equipment stored in the fault knowledge graph unit. The quantified thermal migration trajectory features are matched with a preset feature library to determine the defect type and the degree of deterioration. If the geometric center of the hot spot does not shift significantly, the hot spot area continues to expand, and the temperature gradient gradually increases, it is determined to be a local contact defect (such as increased contact resistance due to joint oxidation, and the hot area gradually spreading). If the geometric center of the hot spot extends along the insulating component of the equipment and the temperature gradient decreases linearly, it is determined to be an insulation deterioration defect (such as surface creepage caused by insulator contamination, with the hot zone migrating along the insulation surface). If hot spots show multi-center diffusion and a sudden drop in temperature decay coefficient, it is judged as an internal fault spillover defect (such as local overheating of transformer windings, with the hot area spreading from the inside of the equipment to multiple locations on the surface).

[0050] The thermal migration risk level is determined based on the quantitative characteristics of the thermal migration trajectory. Thresholds for determining thermal migration risk are set, including: geometric center offset distance ≥ 5mm in a single time period, hot spot area expansion rate ≥ 20% / acquisition cycle, and temperature gradient growth rate ≥ 10% / acquisition cycle. If any threshold is met, it is determined that the spatial evolution of thermal defects is accelerated, and a thermal migration risk warning signal is output.

[0051] Implementation process of multi-time period temperature rise data modeling Multi-time period temperature rise data modeling focuses on the time-dimensional temperature rise characteristics of hot spots in power equipment. It integrates factors such as equipment operating load and ambient temperature and humidity to construct a quantitative correlation model between temperature rise data and influencing factors. This enables the prediction of the deterioration trend of thermal defects over time. The specific execution steps are as follows: The construction and annotation of the multi-time-period temperature rise dataset is based on key monitoring points of power equipment (such as joints, windings, and insulation components). Temperature rise data from multiple consecutive collection cycles are collected to form a basic time series dataset. The dataset is sorted by timestamp, and each data point is associated with multi-dimensional annotation information, including: temperature rise value at the monitoring point (monitoring point temperature - ambient temperature), real-time load rate of the equipment at the time of collection, ambient temperature / humidity, atmospheric pressure, and equipment operating time. At the same time, outliers in the dataset (such as temperature jumps caused by collection interference) are removed by the 3σ criterion and completed by moving average to ensure the validity and continuity of the dataset.

[0052] The feature quantification and correlation analysis of the factors affecting temperature rise perform feature quantification and normalization on the non-temperature rise features (load rate, ambient temperature, etc.) in the dataset, mapping all influencing factors to the 0-1 interval; through Pearson correlation coefficient and partial correlation analysis, the correlation between each influencing factor and the temperature rise value is calculated, and significant influencing factors are screened out (such as the significant influencing factors of transformer joint temperature rise are load rate and ambient temperature), and interference factors without significant correlation are eliminated to reduce the modeling complexity.

[0053] The construction and training of the multi-dimensional temperature rise prediction model uses selected significant influencing factors as input features and temperature rise values ​​as output labels to construct a time-series-based multi-dimensional temperature rise prediction model. The model is based on a Long Short-Term Memory (LSTM) network framework and is customized and optimized in combination with the thermal characteristics of power equipment. The input layer is a fusion feature of significant influencing factors over multiple time periods and historical temperature rise values; A thermal inertia constraint layer for power equipment is added to the hidden layer to adapt to the slow change characteristics of the temperature rise of power equipment and avoid the model from overfitting short-term temperature fluctuations. The output layer outputs the predicted temperature rise value and prediction error range for a preset future time period (e.g., 1 hour, 6 hours, 24 hours). The model is trained using historical temperature rise monitoring data of the equipment, and the model parameters are optimized iteratively through the loss function to ensure the accuracy of the model's prediction of temperature rise trends under different equipment and operating conditions, with a prediction error ≤ ±2℃.

[0054] The quantitative determination and prediction of temperature rise evolution trends involves inputting real-time collected temperature rise data and influencing factors into a multi-dimensional temperature rise prediction model that has been trained. This model outputs short-term / medium-to-long-term temperature rise evolution trend curves and simultaneously quantifies the deterioration state of thermal defects from a time perspective. If the deviation between the actual temperature rise and the model prediction continues to increase, and the rate of temperature rise growth is greater than the preset threshold (e.g., 2℃ / h), it is determined to be an abnormal acceleration of temperature rise, which is a mid-term warning of a slowly deteriorating type of thermal defect. If the model predicts that the temperature rise will reach the critical value of the equipment's dedicated temperature rise threshold model within a preset time period, it will output the predicted time window of the over-threshold, clarifying the expected time of temperature rise exceeding the threshold and the corresponding load / environmental conditions, providing a time reference for operation and maintenance. If the temperature rise trend curve shows the characteristics of "fluctuating rise - stable - sudden rise", it is determined that the thermal defect has entered the explicit stage, and a high-priority warning is immediately triggered.

[0055] Collaborative execution logic of thermal migration trajectory analysis and multi-time period temperature rise data modeling Thermal migration trajectory analysis and multi-time-period temperature rise data modeling are spatial-temporal dual-dimensional analysis methods for the infrared thermal imaging analysis subunit. Both share infrared thermal image preprocessing results and equipment monitoring benchmarks, and are executed collaboratively according to the logic of "spatial features supporting temporal modeling, and temporal modeling verifying spatial evolution." Specifically: The infrared thermal imaging analysis subunit simultaneously extracts the spatial features of the thermal zone and collects the temperature rise values ​​for each frame of the infrared thermal image it acquires. The spatial features are stored in the thermal migration trajectory dataset, and the temperature rise values ​​and related factors are stored in the multi-time period temperature rise dataset. After the data accumulation of the preset collection cycle (such as 3 or more) is completed, the thermal migration trajectory analysis and multi-time period temperature rise data modeling are started simultaneously. The thermal defect type judgment result output by the thermal migration trajectory analysis provides defect type constraints for multi-time period temperature rise data modeling, so that the model adopts differentiated modeling parameters for different types of thermal defects (poor contact, insulation degradation). The temperature rise evolution trend prediction results output by multi-time period temperature rise data modeling provide time dimension verification for heat migration trajectory analysis. If the temperature rise is predicted to be accelerated deterioration, the risk judgment threshold of heat migration trajectory is adaptively tightened to improve the sensitivity of spatial evolution analysis. The analysis results of the two are integrated to form a thermal defect spatial-temporal dual-dimensional evolution report. The report includes: thermal defect spatial migration trajectory, quantitative value of thermal zone characteristic changes, temperature rise time evolution curve, temperature rise exceeding threshold prediction time, thermal defect type and deterioration level. This report, together with the judgment results of the equipment-specific temperature rise threshold model and the time series temperature rise evolution model, is transmitted to the multimodal power feature fusion unit to provide complete thermal defect evolution characteristics for equipment-level feature fusion.

[0056] Model adaptation and results application The registration benchmark and trajectory judgment threshold for thermal migration trajectory analysis, as well as the selection of influencing factors and model parameters for multi-time period temperature rise data modeling, can all be adapted to different equipment types through the fault knowledge graph unit. For the thermal characteristics of different power equipment such as transformers, switching equipment, cable joints, and busbars, the corresponding feature association models and modeling parameters can be retrieved to ensure the relevance and accuracy of the analysis results. The analysis results provide core basis for early warning of the infrared thermal imaging analysis subunit on the one hand, and on the other hand, they are fused with the electrical parameters and acoustic partial discharge characteristics of the equipment at the equipment level to explore the correlation between "thermal defect spatial-temporal evolution" and "electrical parameter fluctuation" and "partial discharge signal enhancement". This provides key thermal evolution feature basis for the power fault mode matching unit to determine the cause of the fault and sort out the cause chain. For example, the hot zone of the cable joint migrates along the cable axis and the temperature rise continues to accelerate. Combined with the three-phase current imbalance characteristics, it can be accurately determined that the cause of the fault is "poor cable joint crimping leads to increased contact resistance, causing thermal defects to spread along the cable insulation layer".

[0057] Acoustic and Partial Discharge Diagnostic Subunit: Based on acoustic spectrum feature extraction and pattern matching algorithms, it identifies the type of partial discharge through the spectral characteristics of acoustic signals. Specific matching rules include: When high-frequency discrete spectral peak characteristics are detected, it is determined to be air corona type partial discharge; When broadband noise superimposed with high-frequency spikes is detected, it is determined to be surface creepage type partial discharge; When periodic vibration spectrum characteristics are detected, it is determined to be abnormal vibration caused by loose flexible connection or mechanical defect of busbar; Based on the IEC60270 industry standard, the subunit combines the amplitude, frequency and energy attenuation characteristics of the acoustic signal to quantitatively infer the actual amplitude and released energy of the partial discharge signal and output a partial discharge diagnostic report.

[0058] It should be noted that the infrared thermal imaging analysis subunit constructs equipment-specific temperature rise threshold models and time-series temperature rise evolution models for different types of power equipment, such as transformers, switchgear, cable joints, and busbars, based on their thermal characteristics and operating patterns. Through a synergistic approach of static threshold determination and dynamic trend analysis, it achieves accurate identification of thermal defects in power equipment, early warning of slowly deteriorating thermal defects, and prediction of thermal defect development trends. This overcomes the technical deficiency of single static threshold determination in missing potential thermal defects. Specific implementation details are as follows: Construction and Implementation of Equipment-Specific Temperature Rise Threshold Model The equipment-specific temperature rise threshold model is a customized static temperature rise anomaly judgment model for different types of power equipment / key components. Based on GB / T, DL / T, and IEC power industry standards, and combined with equipment structural characteristics, rated operating parameters, load capacity, and installation environment, the model sets differentiated and refined temperature rise judgment thresholds and multi-dimensional judgment rules for each type of power equipment / its key connection components. The specific implementation steps are as follows: Model layered construction: Based on the type of power equipment and the function of its components, the model is divided into the equipment body temperature rise threshold layer and the equipment connection point temperature rise threshold layer. The equipment body temperature rise threshold layer is adapted to equipment body parts such as transformer tanks and switch cabinets, while the equipment connection point temperature rise threshold layer is adapted to easily heated connection parts such as transformer joints, disconnector contacts, cable joints, and busbar connection points. Each layer independently sets the temperature rise judgment benchmark. Customized threshold settings: For different types of equipment, a basic temperature rise threshold value ΔTcrit is set for the connection point temperature rise threshold layer, with a range of 25~35℃, and differentiated values ​​are assigned according to the equipment type. An absolute temperature rise threshold under rated operating conditions is set for the equipment body temperature rise threshold layer; for example, the transformer body temperature rise threshold is set to ≤65K according to GB / T1094.2 standard. Simultaneously, a phase-to-phase temperature difference balance threshold is added for three-phase power equipment, with a preset phase-to-phase temperature difference exceeding 8℃ as the abnormal judgment benchmark. Multi-rule anomaly judgment: The model has three core judgment rules built in, which make static and real-time judgments on the thermal state of power equipment. Specifically: (1) Connection point temperature rise judgment: If the temperature of the connection point of the equipment is greater than the ambient temperature + the corresponding ΔTcrit, it is directly judged as a visible thermal defect of poor contact; (2) Phase-to-phase temperature difference judgment: If the phase-to-phase temperature of the three-phase power equipment is asymmetrically distributed, and the temperature difference between any two phases exceeds the preset phase-to-phase temperature difference balance threshold, it is judged as a thermal defect of unbalanced phase current or abnormal load of different phases; (3) Body temperature rise judgment: If the body temperature rise value of the equipment exceeds the absolute temperature rise threshold under rated working conditions, or the temperature of the local area of ​​the equipment is far greater than the average temperature of the body, it is judged as a thermal defect of insulation deterioration and internal impedance rise of the equipment body. Dynamic load adaptation: The model is configured with load linkage adjustment logic, which can receive real-time load data of the device uploaded by the multimodal acquisition module. When the device operating load exceeds 80% of the rated load, the temperature rise judgment threshold is automatically relaxed adaptively to avoid misjudgment caused by normal load fluctuations and improve the accuracy of the model in judgment under different load conditions.

[0059] Construction and Implementation of Time Series Temperature Rise Evolution Model The time-series temperature rise evolution model is a dynamic trend analysis model based on temperature rise monitoring data throughout the entire life cycle of power equipment. The model uses continuous temperature rise data from multiple time periods at key monitoring points of the equipment as its foundation, and combines this with related influencing factors such as ambient temperature and real-time equipment load to uncover the changing patterns of temperature rise data over time. This enables early identification and prediction of the development trend of slowly deteriorating thermal defects. The specific implementation steps are as follows: Construction of standardized time series dataset: Based on the unique identifier of power equipment, periodic continuous temperature rise data of the same key monitoring point of the equipment are collected, and the ambient temperature, real-time load of the equipment, ambient humidity and other data of the corresponding time period are collected synchronously. The temperature rise data is subjected to outlier removal and smoothing and noise reduction to eliminate the temperature rise data deviation caused by environmental interference and equipment vibration, and form a standardized temperature rise time series dataset bound to the equipment ID and unified timestamp. Temperature rise baseline curve fitting: A time series fitting algorithm is used to fit the standardized temperature rise time series dataset to generate a temperature rise change baseline curve for the equipment monitoring points. This curve accurately represents the normal change law of temperature rise value with time, load and ambient temperature under normal operating conditions, and serves as the benchmark for subsequent abnormal trend judgment. Multi-dimensional abnormal trend judgment: Substitute the real-time temperature rise data of the equipment and the related influencing factors into the temperature rise change benchmark curve, and realize the abnormal trend judgment through triple comparison. Even if the real-time temperature rise value does not exceed the critical value of the equipment's exclusive temperature rise threshold model, potential thermal defects can still be identified. Specifically: (1) Curve deviation judgment: If the curve of the real-time temperature rise value changes significantly with the increase of the operating load, it is judged that there are potential thermal defects such as the increase of internal impedance; (2) Rate growth judgment: If the temperature rise growth rate of the equipment monitoring point shows a continuous upward trend, and the temperature rise growth rate is >2℃ / h for three or more consecutive collection cycles, it is judged that there are slow deterioration type thermal defects; (3) Related factor abnormal judgment: If the temperature rise value fluctuates irregularly and the fluctuation amplitude exceeds the preset range when there is no significant change in equipment load and ambient temperature, it is judged that there are signs of early evolution of thermal defects. Thermal defect trend prediction: Based on the established temperature rise evolution model, combined with the future operating load planning data of power equipment and environmental prediction data, the model extrapolates and calculates to predict the deterioration rate of thermal defects and outputs the expected time window when the temperature rise value reaches the critical value of the equipment's specific temperature rise threshold model, providing a time reference for operation and maintenance planning.

[0060] Cooperative operation logic of the two models The equipment-specific temperature rise threshold model and the time-series temperature rise evolution model are complementary algorithm models of the infrared thermal imaging analysis subunit. They share the temperature rise data acquisition and preprocessing results from the infrared thermal imaging analysis subunit and operate collaboratively according to the logic of first static judgment and then dynamic analysis. Specifically: The surface temperature distribution data of power equipment collected by the infrared imager is processed by the multimodal preprocessing unit to perform temperature matrix filtering and normalization, and then simultaneously input into the equipment-specific temperature rise threshold model and the time series temperature rise evolution model. The equipment-specific temperature rise threshold model prioritizes static real-time judgment on the preprocessed temperature rise data. If the judgment result is that the temperature rise value exceeds the threshold, the explicit thermal defect judgment result is immediately output, including information such as defect type, location, and temperature rise exceeding the threshold. The result is then transmitted to the multi-modal power feature fusion unit. If the equipment's dedicated temperature rise threshold model determines that there are no obvious thermal defects, the infrared thermal imaging analysis subunit automatically inputs the temperature rise data into the time series temperature rise evolution model to perform dynamic trend analysis and uncover potential slowly deteriorating thermal defects; if it is determined that there is an abnormal trend, it outputs an early warning signal, including information such as the type of potential defect, the rate of trend deterioration, and the expected time to exceed the threshold, and transmits the results to the multimodal power feature fusion unit. Both models' judgment / early warning results carry a unique equipment identifier, a unified timestamp, and original temperature rise data. These are then fused with equipment-level features such as image defect features, acoustic partial discharge features, and electrical parameter features to provide a complete thermal defect feature basis for subsequent fault mode matching and operational risk quantification.

[0061] Model scene adaptation and updates Both models in the infrared thermal imaging analysis subunit are equipped with scene-based adaptation and self-updating mechanisms. They can make personalized parameter adjustments based on the model, operating years, and installation environment of the power equipment. At the same time, they can receive historical fault diagnosis data from the cloud-based diagnostic module to iteratively optimize the model's judgment threshold and fitting algorithm. For different application scenarios such as substations, transmission lines, and power cabinets, the fault knowledge graph unit can retrieve the evolution law of equipment thermal defects in the corresponding scenario and make targeted adjustments to the model's trend prediction logic to ensure that the model has high judgment accuracy and high early warning sensitivity under different scenarios and equipment operating conditions.

[0062] The multimodal power feature fusion unit adapts to the actual needs of power equipment operation and maintenance scenarios, adopting a device-level feature fusion architecture rather than simple data splicing processing. The fusion architecture takes a single power device as the core associated entity, and integrates the image defect features, infrared temperature rise features, and acoustic partial discharge features output by the high-precision defect identification unit with the electrical parameter features and environmental parameter features uploaded by the multimodal acquisition module across dimensions. It explores the causal relationship between different types of features (such as the correlation between humidity parameters and insulator contamination defects, and the coupling relationship between load parameters and abnormal temperature rise), and outputs a structured device-level fusion feature set. This avoids information loss caused by single-dimensional data analysis and provides comprehensive feature support for subsequent fault mode matching.

[0063] The fault knowledge graph unit constructs and maintains a four-layer association knowledge graph of "equipment type - fault type - operational behavior - evolution result". The core data stored in the knowledge graph includes: A database of typical power equipment fault types, covering known fault modes and characteristics of various types of equipment such as cables, transformers, switchgear, distribution ring main units, and surge arresters; The industry standard safety threshold dataset integrates the safety thresholds, defect judgment standards and maintenance requirements for power equipment operating parameters as specified in GB / T series, DL / T series and IEC series standards; A model of the relationship between load and fault evolution was developed to quantify the impact of different load intensities and durations on the probability of occurrence and the rate of deterioration of various faults. An environmental factor impact model that covers the correlation mechanism between environmental factors such as salt spray, air humidity, diurnal temperature range, and extreme weather and equipment failure; During the fault diagnosis process, the fault knowledge graph unit, based on high-precision defect identification results and multimodal fusion features, enables fault cause tracing (explaining "why this fault was identified"), historical case matching (querying "whether this fault has occurred and how it was handled"), and deterioration trend prediction (analyzing "whether the fault will continue to worsen and its potential consequences"). This breaks the "black box" limitation of pure algorithm diagnosis and improves the interpretability and credibility of the diagnostic results.

[0064] The power fault mode matching unit compares the device-level fused feature set output by the multimodal power feature fusion unit with the preset fault modes in the fault knowledge graph unit for similarity comparison and feature matching, and outputs a fault diagnosis result containing the specific fault category, matching confidence level, and complete causal chain; the causal chain is presented in a logical structure of "feature manifestation - direct cause - indirect impact", and typical examples include: Abnormal temperature rise and localized corrosion characteristics of the disconnector connector indicate a fault category of poor contact. High transformer oil temperature + abnormal vibration signal, corresponding fault category: loose winding; The cable terminal hotspot phenomenon combined with high ambient humidity data indicates the following fault categories: insulation degradation / water treeing aging. The surge arrester leakage current fluctuation and slight abnormal infrared temperature rise correspond to the fault category: potential internal breakdown. Unlike conventional AI diagnostic modules, this unit generates actionable preliminary maintenance suggestions while outputting fault diagnosis results, clearly defining the core objectives and basic procedures of maintenance operations, and providing direct guidance to maintenance personnel.

[0065] The operational risk quantification unit, based on the diagnostic results of the power fault mode matching unit and combined with multi-dimensional influencing factors, performs a graded and quantitative assessment of fault risk, outputting four risk levels from L1 to L4. The definitions of each level are as follows: Level L1: The impact of the fault is minimal, the equipment can continue to operate normally, and the risk assessment result is "recommend regular review", specifying the review cycle and monitoring focus; Level L2: There is a potential accident risk, and the fault may gradually worsen over time. The risk assessment result is "recommended planned maintenance", with clear maintenance windows and priorities. Level L3: This is a serious fault that has affected the normal operation of the equipment. Continued operation may lead to an escalation of the fault. The risk assessment result is "immediate shutdown or replacement of relevant components is required". Level L4: The fault directly threatens the stability of safe power supply and may cause serious consequences such as large-scale power outages and equipment burnout. The risk assessment result is "emergency handling and preservation of fault evidence is required". The risk quantification process comprehensively considers the frequency of fault occurrence, the importance of the equipment involved in the fault (equipment type weight), the current operating load level, the degree of influence of environmental factors, and the historical deterioration data of the equipment. It uses a weighted scoring method to calculate the risk value and map it to the corresponding risk level, providing a scientific quantitative basis for power maintenance resource allocation and operation and maintenance decision-making.

[0066] The maintenance strategy feedback unit, based on the classification results of the operational risk quantification unit and the maintenance suggestions of the power fault mode matching unit, generates targeted closed-loop operation and maintenance control commands, which are then sent back to the multimodal acquisition module and edge processing module. This dynamically adjusts data acquisition parameters and methods, forming a closed-loop operation and maintenance system of "data acquisition—edge analysis—cloud diagnostics—operation and maintenance control—secondary acquisition." Typical scenarios for these control commands include: If the diagnosis is "poor contact", issue a command for the UAV to acquire close-range high-definition images, focusing on the faulty area to obtain more detailed visual data; If the diagnosis result is "partial discharge", issue an acoustic sampling frequency increase command or a direction finding array switching command to enhance the acquisition accuracy and positioning accuracy of the partial discharge signal; If the diagnosis result is "turn ratio deviation", issue a secondary sampling command for multimeter / turn ratio tester to collect additional key electrical parameters to verify the diagnosis result; If the diagnosis result is "slowly deteriorating temperature rise fault", a long-term monitoring instruction is issued to the edge end, requiring the edge processing module to establish a special thermal time series database to continuously track the temperature rise trend; Through the aforementioned closed-loop control, the precise allocation of inspection resources and the dynamic optimization of the data collection process are achieved, thereby improving the operational adaptability and continuous optimization capabilities of the entire diagnostic system.

[0067] Furthermore, the industry standards mentioned include GB / T, DL / T, and IEC standards. Level L1 indicates continued operation, with periodic reviews recommended; Level L2 indicates potential accident risks, with maintenance recommended; Level L3 indicates a serious fault, requiring immediate shutdown or replacement; and Level L4 indicates a threat to safe power supply, requiring emergency handling and evidence preservation.

[0068] Furthermore, the system is applicable to inspection and safety hazard diagnosis scenarios of substations, transmission lines, power cabinets, distribution ring network cabinets, main transformers and switchgear, and can be implemented in the form of single-machine integration or split deployment; A user-side power safety hazard diagnosis method based on multi-channel communication includes: S1: Multimodal data acquisition and time synchronization: Through various inspection devices of the multimodal acquisition module, video data, image data, acoustic signal data, electrical parameter data and environmental parameter data of power equipment and its surrounding environment are collected; the time synchronization module performs a unified timestamp alignment operation on the data collected by each acquisition device to control the timestamp error of the output data of each acquisition device to not exceed the preset threshold. S2: Edge Preprocessing and Lightweight Inference: The edge processing module performs preprocessing operations and lightweight model inference operations on the collected raw data to generate structured anomaly events containing category labels, confidence levels, and anomaly levels. S3: Multi-channel data transmission: The multi-channel communication module monitors the operating status of various types of communication links through the link status detection unit, and executes link selection strategy or link degradation strategy according to the detection results, so as to stably transmit structured abnormal events and related raw data to the cloud diagnostic module. S4: Data caching and delayed retransmission: When the communication link is abnormal, the original data and inference results are stored through the caching and retransmission module; after the communication link is restored to normal, the delayed retransmission operation is automatically executed in the order of timestamps, and the cached data is uploaded to the cloud diagnostic module. S5: Cloud-based in-depth diagnostics and results output: The cloud-based diagnostic module performs in-depth diagnostic operations on the uploaded data and outputs the fault type, risk level, and maintenance strategy of the power equipment.

[0069] Example 1: Deployment of Substation Equipment Inspection Scenario 1. Deployment Environment: This system is deployed within a substation and includes a local data processing terminal (embedded industrial computer), various inspection devices (drone, clamp power meter, acoustic imager, infrared imager, thermometer / hygrometer, multimeter, clamp meter, power meter, transformer ratio tester, VR glasses); and a power maintenance cloud server running a cloud diagnostic module. The local data processing terminal has 4G / WiFi / BT / Ethernet / Type-C communication interfaces and runs an edge analysis module. A time synchronization module ensures timestamp alignment across all data acquisition devices.

[0070] 2. Diagnostic process Step 1: The multimodal acquisition module operates according to the preset inspection plan. The drone collects images and infrared thermal images of high-altitude equipment in the substation (such as busbars and surge arresters); the acoustic imager monitors the partial discharge signals of the equipment; the infrared imager collects the surface temperature of transformers and switchgear; VR glasses assist manual inspection and collect first-person view images of key areas; the data collected by each device are aligned with the timestamp by the time synchronization module.

[0071] Step 2: The edge processing module preprocesses the collected data, such as infrared image temperature matrix filtering and acoustic signal denoising; the lightweight inference unit identifies the transformer joint temperature abnormality through the YOLO-nano model with a confidence level of 0.91, and the abnormal event triggering unit sends an emergency upload command.

[0072] Step 3: The multi-channel communication module detects that the Ethernet link is in normal condition and prioritizes uploading the structured results of the abnormal event and the corresponding infrared image through the Ethernet link; at the same time, other non-urgent data is transmitted through the WiFi link according to normal priority.

[0073] Step 4: After receiving the data, the cloud-based diagnostic module uses a high-precision defect identification unit to accurately locate the abnormal joint position; the multimodal feature fusion unit combines temperature and humidity data with historical operating data to find a correlation between abnormal temperature and ambient humidity; the fault knowledge graph unit calls the joint temperature rise threshold in the DL / T standard to trace the cause to possible poor contact; the power fault mode matching unit outputs the fault type "poor joint contact"; and the operation risk quantification unit comprehensively considers the equipment load level and fault frequency to assess it as an L2 level risk.

[0074] Step 5: The maintenance strategy feedback unit outputs a maintenance suggestion of "arranging joint tightening and temperature measurement re-inspection within 72 hours" and simultaneously issues an instruction to the infrared imager to increase the sampling frequency of the joint (from 10 minutes / time to 5 minutes / time); maintenance personnel receive the information through the dispatch platform and carry out targeted maintenance.

[0075] Example 2: Transmission Line Safety Diagnosis (Mountainous Area Weak Network Scenario): System Deployment The system adopts a split deployment approach, with a local data processing terminal mounted on the inspection vehicle, along with two drones (supporting relay communication), an infrared imager, a clamp meter, and other data acquisition equipment. The multi-channel communication module enables 4G links, satellite backup links, and Bluetooth modules. The power maintenance cloud server runs a cloud-based diagnostic module.

[0076] 2. Diagnostic process Step 1: The drone inspects the power transmission line, collecting images and infrared thermal images of the conductors and insulators, and simultaneously collecting line current data; in some mountainous areas, the 4G signal is weak, so the collected data is synchronized by the time synchronization module and then transmitted to the local data processing terminal.

[0077] Step 2: After the edge processing module preprocesses the data, it identifies slight contamination (3mm) on the insulator using the TinyTransformer model with a confidence level of 0.87, triggering an emergency upload command.

[0078] Step 3: The multi-channel communication module detects that the 4G link RSSI is -95dBm and the packet loss rate is 5%, triggering the link degradation strategy and switching to the satellite backup link; at the same time, non-urgent raw data (such as complete video) is stored by the caching and retransmission module; abnormal event packets are uploaded to the cloud via the satellite link first.

[0079] Step 4: The cloud-based diagnostic module combines the fault knowledge graph with meteorological data (high humidity in mountainous areas) to determine that there is a risk of flashover due to insulator contamination. The power fault mode matching unit outputs the cause chain of "insulator contamination leads to a decrease in insulation performance", and the operation risk quantification unit assesses it as a level L3 risk.

[0080] Step 5: The maintenance strategy feedback unit outputs a maintenance suggestion of "immediately arrange insulator cleaning operation", issues instructions to the drone, adjusts the inspection route, and takes close-up photos of the insulators in the area to obtain more detailed data; when the inspection vehicle travels to an area with good 4G signal, the cache and retransmission module automatically retransmits the stored original data to ensure the comprehensiveness of cloud diagnosis.

[0081] The feasibility of this invention is verified by comparing its performance and advantages. Table 1 shows the statistical table of performance advantages of this solution compared with the prior art: Table 1

[0082] It is worth noting that (1) this invention uses multi-channel redundant communication and dynamic switching to solve the problem of unreliable communication in complex scenarios. It adopts multi-channel redundant communication combined with real-time monitoring of link status and dynamic degradation strategy to build a communication guarantee system without dead ends. The link status detection unit periodically monitors the signal strength (RSSI), network latency (RTT), packet loss rate and available bandwidth of each link. When the main link meets any abnormal threshold (RSSI≤-90dBm, packet loss rate≥3%, upload latency≥300ms, etc.), the system automatically triggers the degradation strategy, switches to the backup link and dynamically adjusts the transmission parameters, such as reducing the video bitrate in weak network environment and only transmitting structured inference results to ensure uninterrupted data transmission. In mountainous power transmission line inspection scenarios, when the 4G signal is weak, the system can switch to the satellite backup link to prioritize the transmission of abnormal event data, while non-urgent raw data is temporarily stored by the cache module and retransmitted according to the timestamp after the link is restored; in scenarios where wired networks are available, such as substations, high-speed transmission of high-definition video and critical data is achieved through Ethernet links; in shielded environments such as inside electrical cabinets, the Bluetooth module ensures stable uploading of sensor data in low-power mode.

[0083] (2) This application adopts hierarchical processing and edge-side intelligent inference to overcome the dilemma of network congestion and insufficient real-time performance. It constructs a hierarchical processing architecture of edge-side preprocessing, lightweight inference, and cloud-based deep analysis to optimize transmission efficiency from the data source while ensuring real-time diagnostics. The edge processing module is equipped with lightweight models such as MobileNet, YOLO-nano, and TinyTransformer. It first performs preprocessing such as image decoding, temperature matrix filtering, and signal denoising on the raw data, and then identifies abnormal events through real-time inference—including image target detection, acoustic anomaly recognition, and temperature and humidity exceeding limits. Only the structured results containing category labels, confidence levels, and anomaly levels are uploaded first, rather than the complete raw data. When the confidence level of the inference result reaches the preset threshold of 0.85 or the duration of the anomaly exceeds the limit, the system triggers an emergency upload command to ensure that critical hidden dangers are reported as soon as possible. Non-urgent data is transmitted with normal priority, and high-bandwidth data such as video is only uploaded in full when the link conditions are good. (3) This application adopts a cache and retransmission and time synchronization mechanism to ensure data integrity and the effectiveness of multimodal fusion. When the communication link is abnormal, the cache and retransmission module automatically stores the original data and inference results. After the link is restored, the data is retransmitted in the order of timestamps to ensure that no data is lost. The time synchronization module controls the timestamp error of each acquisition device within a preset threshold (usually ≤10ms) through the NTP protocol or hardware synchronization mechanism, laying the foundation for the association inference of multimodal data. In the inspection of power transmission lines in mountainous areas, the insulator pollution images, acoustic signals and environmental temperature and humidity data collected by UAVs are time-synchronized to form a time-consistent dataset. Even if the 4G link is interrupted in the middle, the cache module can completely save the data and automatically retransmit it when the vehicle reaches the area with good signal, ensuring the comprehensiveness of cloud diagnosis. This design not only completely solves the data loss problem, but also enables the multimodal feature fusion unit to accurately mine data association. (4) This application adopts cloud-based deep diagnosis and closed-loop operation and maintenance to improve the accuracy of diagnosis and the scientific nature of decision-making. The cloud diagnosis module integrates high-precision defect identification, multi-modal feature fusion, fault knowledge graph and other units to realize the whole chain diagnosis of identification, tracing, evaluation and decision-making. The high-precision defect identification unit is based on Transformer or visual attention network and can accurately locate small defects of 2-5mm (such as joint ablation, slight contamination of insulators). The accuracy rate of obvious hidden danger identification is ≥98.5%, and the missed detection rate of small defects is ≤0.8%. The fault knowledge graph unit stores industry standards such as GB / T, DL / T, IEC and fault evolution relationship, and can trace the fault cause chain (such as "temperature rise and local corrosion of switch joint → poor contact"). The operation risk quantification unit assesses the risk according to the four levels of L1-L4, providing a clear basis for operation and maintenance decision-making. (5) The present invention has strong inspection initiative: closed-loop control is achieved through the maintenance strategy feedback unit, and the data acquisition parameters and methods are dynamically adjusted according to the diagnostic results to achieve proactive optimization of the inspection process and improve inspection efficiency and intelligence level.

[0084] The above description is merely a few embodiments of this application and is not intended to limit this application in any way. Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any changes or modifications made by those skilled in the art without departing from the scope of the technical solution of this application using the disclosed technical content are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A user-side power safety diagnostic system based on multi-channel communication, characterized in that, It includes a multimodal acquisition module, a time synchronization module, an edge processing module, a multi-channel communication module, a buffering and retransmission module, and a cloud diagnostic module. Each module works collaboratively through a data interaction link. The multimodal acquisition module is used to collect multimodal data of power equipment and its surrounding environment through various inspection devices. The multimodal data includes video data, image data, acoustic signal data, electrical parameter data, and environmental parameter data. The time synchronization module is used to uniformly synchronize the system clocks of each acquisition device in the multimodal acquisition module, and control the acquisition clock error of each device to not exceed a preset threshold, thereby ensuring the time synchronization of the acquired data. The edge processing module is deployed on the local data processing terminal and is used to perform preprocessing operations and lightweight model inference operations on the raw data collected by the multimodal acquisition module to generate structured abnormal events. The multi-channel communication module is configured with multiple types of communication links and is equipped with a link status detection unit, which is used to execute link selection strategy or link degradation strategy according to the link status detection result to achieve stable data transmission. The caching and retransmission module is used to store the original data and inference results when the communication link is abnormal, and to perform delayed retransmission in the order of timestamps after the link is restored. The cloud-based diagnostic module is deployed on the power operation and maintenance cloud server and is used to perform in-depth diagnostic operations on the uploaded data, outputting fault type, risk level and maintenance strategy.

2. The user-side power safety diagnostic system based on multi-channel communication according to claim 1, characterized in that, The multimodal acquisition module includes any one or more of the following: drone, clamp power meter, acoustic imager, infrared imager, thermometer and hygrometer, multimeter, clamp meter, power meter, transformer ratio tester, and VR glasses. The drone is used to collect image data, infrared thermal image data, and spatial positioning data of high-altitude or long-distance power equipment. The clamp-on power meter is used for non-contact acquisition of line current data, power factor data, and harmonic characteristic data; The acoustic imager is used to generate sound source imaging data through an array of microphones to achieve partial discharge localization. The infrared imager is used to collect surface temperature distribution data of power equipment and identify hot spots and overload conditions of the equipment. The thermometer and hygrometer are used to collect ambient temperature and humidity data. The multimeter is used to collect basic electrical quantity data and verify the continuity of the equipment and the correctness of the wiring; the clamp meter is used to perform high-frequency inspection current monitoring and collect high-frequency current data. The power meter is used to perform power quality analysis and output power quality related parameters. The transformer ratio tester is used to detect the winding ratio data and magnetic flux characteristic data of transformer equipment. The VR glasses are used to collect first-person perspective image data, eye movement information data, or voice data during manual inspections, and interact with the edge processing module through a local network.

3. The user-side power safety diagnostic system based on multi-channel communication according to claim 1, characterized in that, The time synchronization module achieves time alignment through the NTP protocol or hardware synchronization mechanism, and the preset threshold is set according to the requirements of multimodal data fusion analysis.

4. The user-side power safety diagnostic system based on multi-channel communication according to claim 1, characterized in that, The edge processing module includes a multimodal preprocessing unit, a lightweight inference unit, a cache and timestamp synchronization unit, and an abnormal event triggering unit; The multimodal preprocessing unit is used to perform image decoding, temperature matrix filtering, acoustic point cloud aggregation, data denoising and normalization on the raw data; The lightweight inference unit deploys a lightweight intelligent model to perform real-time inference on the preprocessed data and outputs structured inference results containing category labels, confidence levels, and anomaly levels. The cache and timestamp synchronization unit is used to cache the original data and inference results locally and align them with a unified timestamp. The abnormal event triggering unit is used to continuously monitor the confidence level output by the lightweight inference unit. When the confidence level reaches a preset threshold or the abnormal duration exceeds a preset duration, it sends an emergency upload command to the multi-channel communication module.

5. The user-side power safety diagnostic system based on multi-channel communication according to claim 4, characterized in that, The lightweight intelligent model includes any one of MobileNet, YOLO-nano, and TinyTransformer, and the real-time inference includes image target detection, acoustic anomaly recognition, temperature and humidity threshold exceeding judgment, and current waveform sudden change judgment.

6. The user-side power safety diagnostic system based on multi-channel communication according to claim 1, characterized in that, The multi-channel communication module includes multiple types of communication links, including 4G links, WiFi modules, Bluetooth modules, Ethernet wired interfaces, and Type-C wired interfaces. The link status detection unit is used to periodically monitor the signal strength, network latency, packet loss rate, and available bandwidth of each communication link; The link selection strategy is executed based on a preset link priority, which is Ethernet > Type-C > WiFi > Dual 4G > BT; The link degradation strategy is triggered when the main link meets the preset threshold conditions. The preset threshold conditions include at least one of the following: RSSI≤-90dBm, packet loss rate≥3%, upload latency≥300ms, TCP handshake failures≥2 consecutive times, and uplink bandwidth<minimum video encoding requirements. The 4G link is used for video streaming or high-priority data transmission during normal inspection processes; The WiFi module is used to enable wireless data interaction between the VR glasses and the local data processing terminal. The local data processing terminal acts as a WiFi hotspot, and the VR glasses act as a client connecting to the hotspot. The Bluetooth module is used for low-power data uploading from various acquisition terminals. The local data processing terminal acts as the master device, and multiple sensors or detection instruments act as slave devices, uploading the acquisition results in broadcast or short connection mode. The Type-C wired interface is used for high-speed wired download of video and image data after the drone inspection is completed.

7. The user-side power safety diagnostic system based on multi-channel communication according to claim 1, characterized in that, The cloud-based diagnostic module includes a high-precision defect identification unit, a multimodal power feature fusion unit, a fault knowledge graph unit, a power fault mode matching unit, an operational risk quantification unit, and a maintenance strategy feedback unit. The high-precision defect identification unit is equipped with a Transformer or visual attention network to accurately locate and identify equipment defects in image data, infrared thermal imaging data and acoustic data. The multimodal power feature fusion unit uses a device-level feature fusion method to integrate different types of collected data at the device level and mine the correlation between the data. The fault knowledge graph unit stores typical fault types of power equipment, industry standard safety thresholds, the relationship between operating load and fault evolution, and environmental factor influence models, which are used for tracing and explaining the causes of faults. The power fault mode matching unit is used to compare the fused feature data with the preset fault mode and output the specific fault type and cause chain. The operational risk quantification unit is used to comprehensively assess the failure risk in four levels, from L1 to L4, based on the frequency of failure occurrence, equipment type, load level, environmental factors, and historical degradation data. The maintenance strategy feedback unit is used to output maintenance suggestions based on the deep diagnostic results and to send control commands to the multimodal acquisition module to adjust the data acquisition parameters or acquisition method.

8. The user-side power safety diagnostic system based on multi-channel communication according to claim 7, characterized in that, Level L1 indicates continued operation is possible, and periodic reviews are recommended; Level L2 indicates potential accident risks, and maintenance is recommended; Level L3 indicates a serious fault, requiring immediate shutdown or replacement; Level L4 indicates a threat to safe power supply, requiring emergency handling and preservation of evidence.

9. The user-side power safety diagnostic system based on multi-channel communication according to claim 1, characterized in that, The system is applicable to inspection and safety hazard diagnosis scenarios of substations, transmission lines, power cabinets, distribution ring network cabinets, main transformers and switchgear, and can be implemented in the form of single-machine integration or split deployment.

10. A method for diagnosing user-side power safety hazards based on multi-channel communication, characterized in that, include: S1: Multimodal data acquisition and time synchronization: Through various inspection devices of the multimodal acquisition module, video data, image data, acoustic signal data, electrical parameter data and environmental parameter data of power equipment and its surrounding environment are collected; the time synchronization module performs a unified timestamp alignment operation on the data collected by each acquisition device to control the timestamp error of the output data of each acquisition device to not exceed the preset threshold. S2: Edge Preprocessing and Lightweight Inference: The edge processing module performs preprocessing operations and lightweight model inference operations on the collected raw data to generate structured anomaly events containing category labels, confidence levels, and anomaly levels. S3: Multi-channel data transmission: The multi-channel communication module monitors the operating status of various types of communication links through the link status detection unit, and executes link selection strategy or link degradation strategy according to the detection results, so as to stably transmit structured abnormal events and related raw data to the cloud diagnostic module. S4: Data caching and delayed retransmission: When the communication link is abnormal, the original data and inference results are stored through the caching and retransmission module; after the communication link is restored to normal, the delayed retransmission operation is automatically executed in the order of timestamps, and the cached data is uploaded to the cloud diagnostic module. S5: Cloud-based in-depth diagnostics and results output: The cloud-based diagnostic module performs in-depth diagnostic operations on the uploaded data and outputs the fault type, risk level, and maintenance strategy of the power equipment.

Images