Intelligent identification risk early warning method for distribution network non-power-off operation

Abstract

The application discloses a smart identification risk early warning method for distribution network live operation, and particularly relates to the technical field of live operation of distribution network, and comprises the following steps: high-definition cameras, positioning sensors and global satellite navigation system time service are arranged in the operation area to collect image and position information; a fusion sequence is generated by aligning with a unified time reference; a deep learning model is input to identify personnel protection, insulation shielding and boundary crossing, a dynamic risk index is calculated in combination with the relative potential of personnel, tools and live wires; the time service integrity is continuously monitored and self-checked; sound and light warnings are issued and safety interlocking is cut in when necessary; after the alarm is removed, the data is stored and incremental learning is performed for adaptive updating; the application fuses the collected data with a unified time reference, analyzes the protection, insulation shielding and boundary crossing with deep learning to generate a dynamic risk index, issues early warnings and interlocks, reduces electric shock, falling and misoperation; the time service integrity is monitored and self-checked, and the redundancy verification is performed to reduce false positives and false negatives; the recognition accuracy and power supply continuity are improved through incremental learning updating.

Description

Technical Field

[0001] This invention relates to the field of live-line operation and maintenance technology for distribution networks, and more specifically, to an intelligent risk identification and early warning method for live-line operation of distribution networks. Background Technology

[0002] Live-line work on power distribution networks is rapidly increasing in scenarios such as urban power distribution system upgrades, distributed power source integration, and fault repair with live connections. This work faces a combination of risks, including electric shock, step voltage, falls from heights, and accidental contact with conductors by tools. The on-site environment is highly dynamic, with factors such as sudden changes in lighting, partial obstruction, specular reflection, rain, fog, dust, and the involvement of multiple personnel and robotic arms all contributing to the challenges. This makes it difficult for single-channel video monitoring or manual visual inspection to detect dangerous signs in a timely and consistent manner. Current practices often rely on single-channel cameras or simple rule-based thresholds, lacking a unified quantitative representation of personnel protective posture, insulation integrity, and boundary crossing behavior. Risk warnings are delayed and prone to false alarms and missed alarms, failing to meet the dual requirements of safety and continuity in live-line work.

[0003] Although combined solutions involving multiple cameras, positioning, and attitude perception have emerged in recent years, and attempts have been made to use network or satellite time synchronization for time calibration, maintaining consistency across devices and links within the microsecond range over long periods remains challenging. Short-term satellite signal blockage, network jitter, and local time base drift introduce timestamp deviations, leading to misalignment between video frames and pose data, further distorting spatial relationship estimation and reducing the effectiveness of multi-source fusion. Existing recognition algorithms are mostly trained on offline data, lacking incremental learning and version verification processes for real-world operations, causing models to age gradually with seasonal and operational changes. Furthermore, the alarm link and safety interlocking are disconnected, making it impossible to reliably close the loop between risk quantification indicators and tiered handling strategies, hindering rapid and verifiable isolation control. Therefore, this invention proposes an intelligent risk identification and early warning method for uninterrupted power distribution network operations to address the aforementioned problems. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] The intelligent risk identification and early warning method for live-line work in power distribution networks includes the following steps:

[0006] Step 1: Deploy high-definition camera equipment, positioning sensor devices, and global satellite navigation system timing modules in the work area to simultaneously collect panoramic image data and location information of the work area;

[0007] Step 2: Using the unified time reference provided by the global satellite navigation system's timing module, align the panoramic image data with the location information to generate a fused data sequence;

[0008] Step 3: Input the fused data sequence into the deep learning model to analyze personnel protective posture, insulation shielding status and boundary crossing behavior, and calculate the dynamic risk index based on the relative spatial potential between personnel, tools and live wires;

[0009] Step 4: Continuously monitor the integrity of the time synchronization information. When time synchronization drift is detected and the alignment error exceeds the preset threshold, execute the self-calibration algorithm to restore time synchronization and start the redundancy verification process.

[0010] Step 5: When the dynamic risk index reaches the alarm threshold, output an audible and visual warning signal to the field terminal and dispatch platform, and automatically switch the working device to the safety interlock mode.

[0011] Step 6: After the audible and visual warning is lifted, the dynamic risk index and the corresponding work scenario data are stored in the safety database, and incremental learning is performed based on the stored data to adaptively update the deep learning model.

[0012] In a preferred embodiment, step one specifically includes the following:

[0013] Install high-definition camera equipment with automatic zoom and optical axis self-calibration functions, and use laser dot matrix to complete spatial calibration, with imaging distortion error controlled within 0.2 percent.

[0014] A three-axis inertial combined positioning sensor is deployed at the edge of the camera's field of view, and centimeter-level coordinate unification is achieved using differential positioning of a dual-antenna global satellite navigation system.

[0015] The timestamp of each frame of image is aligned with the timing module of the global satellite navigation system by hardware triggering signal to form an absolute time label and build a synchronous data stream.

[0016] The final image data and location information achieved a preset target level of coverage in three-dimensional space.

[0017] In a preferred embodiment, step two specifically includes the following:

[0018] The timing module of the dual-frequency global satellite navigation system is used to drive a distributed pulse trigger to measure the frame time error of each camera device in real time and output a microsecond-level correction vector.

[0019] The correction vector is dynamically injected into an adaptive synchronization algorithm that combines Kalman filtering and spline interpolation to perform continuous time axis alignment on panoramic image data and location information to generate an absolute timestamp.

[0020] Immediately after alignment, the image frame and corresponding position information are written into the multimodal fusion buffer, and a time integrity flag is attached by a hash function. Finally, the processed multimodal fusion data stream is obtained and written into the circular buffer.

[0021] In a preferred embodiment, a two-stage time calibration process is established between the distributed pulse trigger and the adaptive synchronization algorithm, specifically including:

[0022] By using timing pulses from the global satellite navigation system to perform coarse calibration, the time error of each camera device is reduced to within 20 microseconds;

[0023] Micro-calibration is performed using a combination of high-stability time base and inertial positioning data from the camera device to generate a nanosecond-level deviation matrix, which is then fed back to the adaptive synchronization algorithm in real time.

[0024] Following the hash mark in the buffer, an elliptic curve digital signature and a time integrity tag are appended to each frame to form a chain of traceable data blocks. Finally, the data stream that has been calibrated and signed at two levels is written to the circular buffer.

[0025] In a preferred embodiment, the processing steps for fusing data sequences into a deep learning model include:

[0026] The fused data sequence is rearranged into a fixed-length tensor set using temporal normalization.

[0027] A multi-branch convolutional network was used to extract key point features of the personnel skeleton, insulation occlusion contour features, and multi-scale texture features of the work boundary.

[0028] The three types of features are weighted and coupled through an attention gating mechanism, and the pose vector, occlusion mask, and out-of-bounds label are output in the self-attention decoding network.

[0029] In a preferred embodiment, the dynamic risk index calculation step includes:

[0030] A three-dimensional point cloud nearest neighbor search tree is used to perform Euclidean distance calculations on key points of the personnel skeleton, tool endpoints and all grid nodes of the live wires, extract the shortest distance and output the minimum safety gap vector.

[0031] The minimum safety gap vector and the attitude velocity vector are jointly mapped to the instantaneous energy potential field, and the potential field gradient is calculated using the finite difference method to form a cross-temporal tensor.

[0032] The cross-temporal tensor is input into the dual-gated cyclic unit model to simultaneously predict the probability of electric shock and the probability of falling, and the comprehensive risk score is obtained by weighted averaging.

[0033] The comprehensive risk score is normalized to a range of zero to one and then published as a dynamic risk index.

[0034] In a preferred embodiment, step four specifically includes the following:

[0035] A two-way time difference estimator is constructed by using the timing pulse of the global satellite navigation system and the local time base of the camera equipment to continuously output a synchronization residual sequence with a precision of hundreds of microseconds.

[0036] The root mean square value of the synchronization residual sequence is calculated within a fixed-length sliding window and compared with a preset threshold. When the root mean square value exceeds the preset threshold, the self-calibration algorithm is triggered. The local time base is corrected by polynomial regression and the correction vector is broadcast to reduce the alignment error back to the microsecond level.

[0037] After self-calibration is completed, a redundancy verification process is initiated, which compares the corrected timestamp with the distributed rubidium atomic clock signal bidirectionally, generates a time integrity hash, and writes the time integrity hash into a circular cache.

[0038] In a preferred embodiment, step five specifically includes the following:

[0039] The dynamic risk index is compared with two thresholds. Once the dynamic risk index exceeds the alarm threshold, a dual-frequency sound wave is triggered at the on-site terminal and a 120-lumen red pulse lamp is lit simultaneously to form a composite sound and light warning.

[0040] An encrypted alarm frame is sent to the scheduling platform through a preset time-sensitive network channel, and a heartbeat link is maintained for fifty milliseconds after the platform returns an acknowledgment signal;

[0041] Upon receiving the confirmation signal, the safety switching script is immediately executed to disconnect the power circuit of the working device, lock the hydraulic valve of the lifting arm, and deploy the insulation barrier, thereby switching the working device to the safety interlock mode.

[0042] During the safety interlock mode, the dynamic risk index is continuously recorded and compared with the reset threshold. When the dynamic risk index falls below the reset threshold, the interlock is automatically released and normal operation is restored. The reset threshold is set by multiplying the historical average risk index by the safety factor K.

[0043] The technical effects and advantages of this invention are as follows:

[0044] This invention uses high-definition camera equipment, positioning sensing devices, and a global satellite navigation system timing module to jointly collect panoramic image data and location information, and generates a fused data sequence under a unified time reference. The fused data sequence is then input into a deep learning model, which can simultaneously analyze personnel protective posture, insulation shielding status, and boundary crossing behavior, and calculate a dynamic risk index accordingly. This enables real-time risk identification and rapid alarm under uninterrupted power supply conditions, reducing the probability of electric shock, falls from heights, and misoperation, and improving operational safety and power supply continuity.

[0045] This invention sets up a continuous monitoring mechanism for the integrity of timing information. When timing drift is detected and the alignment error exceeds a preset threshold, a self-calibration algorithm is executed and a redundant verification process is initiated to ensure the reliability of the time alignment of the fused data sequence, reduce the risk of false alarms and missed alarms caused by time mismatch, make the judgment basis of the dynamic risk index stable and reliable, and ensure the accuracy and timeliness of the triggering of audible and visual warning signals and safety interlocking modes.

[0046] After the audible and visual warning is deactivated, this invention stores the dynamic risk index and corresponding work scenario data in a safety database. Based on the stored data, incremental learning is performed to adaptively update the deep learning model, forming a closed loop of data collection, risk assessment, and model update. This allows the model to be continuously optimized in the real work environment, gradually improving the ability to identify personnel protective posture, insulation shielding status, and boundary crossing behavior. As a result, higher risk identification accuracy and more stable linkage control effect can be achieved in long-term operation. Attached Figure Description

[0047] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0048] Figure 1 This is a schematic diagram of the intelligent risk identification and early warning method for uninterrupted power distribution network operations in this invention. Detailed Implementation

[0049] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0050] Reference Figure 1 The following examples were obtained:

[0051] Example 1: An intelligent risk identification and early warning method for live-line work in power distribution networks, comprising the following steps: Step 1: Deploy high-definition camera equipment, positioning sensing devices, and a global satellite navigation system timing module in the work area to simultaneously collect panoramic image data and location information of the work area; its significance lies in establishing a sensing base with centimeter-level positioning accuracy and microsecond-level unified time reference as its core, incorporating all moving and static targets such as personnel, tools, and wires into the same three-dimensional coordinate system, enabling subsequent algorithms to perform fine tracking of each object within the complete field of view; at the same time, the combination of high-resolution visual flow and high-stability time base can maintain the spatiotemporal continuity of key frames even in extreme scenarios such as rapid changes in illumination and complex background occlusion, providing a low-distortion, low-drift data source for risk feature extraction, and laying the foundation for subsequent multimodal fusion, time difference compensation, and reliability verification.

[0052] Step 2: Using the unified time reference provided by the global navigation satellite system's timing module, the panoramic image data and location information are aligned to generate a fused data sequence. The significance lies in using collaborative mechanisms such as distributed pulse triggering, Kalman filtering, and spline interpolation to reorganize asynchronously acquired multi-channel information onto the same microsecond-level time axis, constructing a continuous tensor sequence that retains fine-grained motion continuity while eliminating sensor jitter. Before writing this sequence to the circular buffer, an integrity hash is appended, giving each frame a traceable timestamp and an anti-tamper fingerprint. This ensures that the input stream received by the downstream deep learning model remains highly consistent in both the time and numerical domains, significantly improving the robustness of risk identification and the interpretability of results.

[0053] Step 3: Input the fused data sequence into the deep learning model to analyze personnel protective posture, insulation shielding status, and boundary crossing behavior. Calculate the dynamic risk index based on the relative spatial potential between personnel, tools, and live wires. The significance lies in utilizing an end-to-end architecture of a multi-branch convolutional backbone, a posture graph convolutional network, and a spatiotemporal attention decoder to output posture vectors, shielding masks, and boundary crossing labels in parallel during a single forward computation. These high-dimensional semantic features and continuous posture tensors are then mapped to an energy potential field model to assess the probability of electric shock and fall in real time. Finally, the results are quantified into a dynamic risk index ranging from zero to one floating-point value, providing a quantifiable and thresholdable safety indicator for subsequent decision-making and interlocking control.

[0054] Step 4: Continuously monitor the integrity of timing information. When timing drift is detected and the alignment error exceeds the preset threshold, execute the self-calibration algorithm to restore time synchronization and initiate the redundancy verification process. The significance of this is that the bidirectional time difference estimator continuously outputs the residual sequence and calculates the root mean square value within the sliding window, capturing the microsecond-level drift trend in real time. Once an anomaly occurs, the self-calibration algorithm immediately reconstructs the local time base and broadcasts the correction vector to pull the error back to a safe range, ensuring that the entire data link can quickly recover to usable synchronization accuracy under any abnormal scenario, avoiding erroneous data triggering error alarms or missed reports.

[0055] Step 5: When the dynamic risk index reaches the alarm threshold, an audible and visual warning signal is output to the field terminal and the dispatch platform, and the operating equipment is automatically switched to the safety interlock mode. The significance of this is that a dual-threshold decision strategy is used to distinguish between the warning and emergency shutdown levels, and after the high-level trigger, the audible and visual multi-mode prompt, encrypted alarm frame transmission, and hardware safety interlock execution are completed within hundreds of milliseconds, ensuring that the field personnel can immediately perceive and physically isolate the source of danger. At the same time, the heartbeat link is maintained through a time-sensitive network to ensure that the status of the dispatch center and the field controller is consistent, preventing secondary risks caused by accidental reset or loss of connection.

[0056] Step Six: After deactivating the audible and visual warning, store the dynamic risk index and corresponding operational scenario data in the safety database, and perform incremental learning based on the stored data to adaptively update the deep learning model. The significance of this is to construct a closed-loop data feedback mechanism, continuously injecting high-risk samples and edge cases from real-world operational scenarios into the training set. This utilizes an online incremental learning strategy to mitigate model aging and scenario drift issues, enabling risk identification accuracy to increase over time. Simultaneously, the safety database supports version management and difference auditing, providing reliable evidence for post-incident tracing, model rollback, and regulatory compliance checks, ensuring power supply continuity while achieving long-term self-evolution of the system.

[0057] Step one specifically includes the following: installing a high-definition camera with automatic zoom and optical axis self-calibration functions, and completing spatial calibration with a laser dot array, controlling the imaging distortion error to within 0.2 percent; deploying a three-axis inertial combined positioning sensor at the edge of the camera's field of view, and using dual-antenna global satellite navigation system differential positioning to achieve centimeter-level coordinate unification; aligning the timestamp of each frame of image with the global satellite navigation system timing module through a hardware trigger signal to form an absolute time tag and construct a synchronous data stream; finally, the coverage of the obtained image data and position information in three-dimensional space reaches the preset target level.

[0058] The system is equipped with a high-definition camera featuring automatic zoom and optical axis self-calibration. This high-definition camera refers to an industrial-grade camera with a resolution of at least 3840 x 2160 pixels and a frame rate of at least 60 frames per second. Internally, it features a stepper motor-driven zoom lens assembly and a twelve-zone phase-shift calibration unit to maintain a consistent focal plane and automatically eliminate lens center shift at different working distances. The automatic zoom and optical axis self-calibration functions drive the lens assembly displacement through a real-time closed-loop algorithm, reading the image sharpness evaluation function at ten-millisecond intervals and adjusting the focal length to maximize edge gradients. Simultaneously, the optical axis self-calibration module performs zero-point correction on the viewing axis deviation, ensuring that the target center drift does not exceed 0.02 degrees. The laser dot array uses a 940-nanometer wavelength infrared semiconductor emitter array, projecting 364 Gaussian points with a 100-m spacing, covering a 15-meter x 15-meter working surface. The spatial calibration process uses the Zhang Zhengyou method to obtain the extrinsic parameter matrix, combining it with the laser point coordinates to establish a camera imaging model. Ultimately, the imaging distortion error is compressed to within 0.2%, meaning the distortion displacement at the 2000-pixel edge of the image does not exceed four pixels. In the actual embodiment, a four-K line camera of model GX-UHD-60 was selected, paired with an F1.6 auto aperture zoom lens, which can meet the clear imaging requirements of shooting distances from ten meters to thirty meters.

[0059] A three-axis inertial positioning sensor is deployed at the edge of the camera's field of view. This sensor integrates a high-precision accelerometer, gyroscope, and magnetometer. The accelerometer has a range of ±16 times the gravitational acceleration, the gyroscope has a range of ±2000 degrees per second, and a sampling frequency of 200 Hz. Attitude calculation is achieved through Kalman filtering. The sensor is evenly distributed at the four corners of the camera's field of view, acquiring motion data and aligning it with the camera's time reference via synchronization pulses. The dual-antenna global navigation satellite system differential positioning uses an antenna pair with a baseline length of one meter. Differential observation equations are constructed by receiving ephemeris data from multiple systems, achieving a static accuracy of one centimeter horizontally and two centimeters vertically. In the actual embodiment, a dual-antenna multi-constellation receiver (model BD982) is used in conjunction with an omnidirectional helical antenna to periodically output centimeter-level coordinates, providing a unified reference frame for point cloud reconstruction and 3D registration.

[0060] The timestamp of each frame of image is aligned with the timing module of the Global Navigation Satellite System (GNSS) via a hardware trigger signal. This hardware trigger signal uses pulses per second and a 10 MHz constant-frequency pulse output by the GNSS, directly driving the camera's exposure start pin via a clock distribution circuit to achieve frame-level synchronization. After each exposure, the camera writes a timestamp to the metadata area of ​​the data stream. This timestamp is consistent with the common time encoding format of the timing module, with an accuracy better than 50 microseconds. The synchronized data stream is transmitted to the edge computing node via Gigabit Ethernet. The node has a circular buffer with a one-second depth to prevent frame loss due to sudden network jitter.

[0061] The final image data and location information achieved a 3D spatial coverage rate that met the preset target level. 3D spatial coverage is defined as the percentage of any point within a 15m x 15m x 10m work volume that is observed by at least two cameras at an angle greater than 60 degrees, with a joint positioning error of less than 3 centimeters. The system design target was a coverage rate of no less than 99.5%. In field testing, by adjusting the camera installation height to 8 meters and the azimuth angle interval to 45 degrees, and using a positioning sensor to correct the projection angle in real time, a spatial coverage rate of 99.7% was achieved, meeting the requirements of the dynamic risk identification algorithm for viewpoint redundancy and accuracy.

[0062] Step two specifically includes the following: A distributed pulse trigger is driven by a dual-frequency global navigation satellite system (GNSS) timing module to measure the frame time error of each camera in real time and output a microsecond-level correction vector. The dual-frequency GNSS timing module simultaneously receives high-stability carrier signals from, for example, the US Global Positioning System (GPS) and the Chinese BeiDou Navigation Satellite System (BDS) in the L1 and L5 frequency bands, eliminating frequency drift caused by ionospheric delay and providing a 10 MHz constant-frequency reference for the clock network. The distributed pulse trigger consists of a field-programmable gate array (FPGA) and an adjustable delay line. Driven by the timing module with pulses per second, it generates an exposure start signal with a width of 3.7 milliseconds according to the frame period and acquires the time difference between the camera devices at the rising edge. The system reads the time difference sequence of each camera device in real time, calculates the frame-level cumulative error, and outputs a microsecond-level correction vector. The delay line is adjusted in steps to ensure that the time drift of the entire network does not exceed 25 microseconds. In the field embodiment, eight FX-60 industrial cameras were deployed. After correction by the pulse trigger, the synchronization deviation between the longest baselines was reduced from 300 microseconds to 22 microseconds.

[0063] The correction vector is dynamically injected into an adaptive synchronization algorithm combining Kalman filtering and spline interpolation to perform continuous time-axis alignment of panoramic image data and location information to generate an absolute timestamp. The Kalman filter uses the correction vector as an observation, fuses the time difference prediction value from the previous frame, and outputs a smooth synchronization deviation. The spline interpolator uses cubic B-splines to perform curve fitting on the discrete deviation sequence, calculating fine-tuning at any time point to resolve abrupt jumps caused by instantaneous satellite signal obstruction. The adaptive synchronization algorithm automatically adjusts the process noise covariance based on the filter residual. When the residual exceeds fifty microseconds, the gain is immediately reduced to enhance stability, and the response rate is increased when the residual is below twenty microseconds. The filter updates the state covariance matrix every ten frames to maintain the steady-state mean of the alignment error at fifteen microseconds. The absolute timestamp is represented in International Atomic Time format with an accuracy of one microsecond, and includes a frame number and camera equipment identification code to ensure cross-node traceability.

[0064] Immediately after alignment, the image frame and its corresponding position information are written to the multimodal fusion buffer, and a time integrity tag is appended using a hash function. Finally, the processed multimodal fusion data stream is obtained and written to the circular buffer. The multimodal fusion buffer, located in the shared memory of the edge computing nodes, is 256 megabytes in size and can hold three seconds of data. The buffer writing process first stores the 8-bit grayscale matrix of the image frame, then appends the 3D coordinates and pose quaternions calculated by differential positioning. The time integrity tag is calculated by concatenating the timestamp, frame number, and pose vector into a string using a SHA-256 hash function, outputting a 32-byte digest to prevent data tampering or replay. The processed multimodal fusion data stream is written to the circular buffer using a zero-copy method. The circular buffer has 1024 records and is overwritten according to the first-in-first-out principle, with a maximum latency of 50 milliseconds. The upper-layer deep learning inference process reads 16 frames in batches from the circular buffer as sliding window input, achieving millisecond-level non-blocking data supply.

[0065] A two-stage time calibration process is established between the distributed pulse trigger and the adaptive synchronization algorithm. Specifically, this includes: coarse calibration using timing pulses from the Global Navigation Satellite System (GNSS) to compress the time error of each camera to within 20 microseconds; and in the coarse calibration stage, a 1-hertz pulse signal output from the GNSS is used as a unified trigger source. The distributed pulse trigger has a built-in delay scan module that first measures the original time difference between the exposure start and the rising edge of the pulse for each camera, then uses a six-digit digital delay line to gradually compensate, with a compensation step size of one microsecond. After five iterations, the maximum time difference is converged from 350 microseconds to 18 microseconds. In the field example, eight 4K industrial cameras are deployed within a 20-meter baseline. After the coarse calibration, the mean synchronization residual is 13 microseconds, and the standard deviation is 3 microseconds, providing a low-deviation starting point for refined calibration.

[0066] Micro-calibration is performed using the high-stability time base and inertial positioning data integrated within the camera equipment, generating a nanosecond-level deviation matrix and feeding it back to the adaptive synchronization algorithm in real time. During the micro-calibration phase, relying on the camera's on-chip temperature-controlled crystal oscillator, the long-term drift rate of the time base is less than 10 to the power of -8. Combined with the angular velocity and linear acceleration output from the three-axis inertial positioning device, instantaneous platform attitude and jitter are collected for each frame. The adaptive synchronization algorithm uses the coarse calibration residual as the initial value, employs Kalman filtering to predict the next exposure trigger time, and then corrects the prediction using the slight optical path shift caused by attitude, obtaining the theoretical exposure timescale. The difference between the theoretical and actual timescales is used to construct a three-dimensional deviation matrix, with matrix elements in nanosecond units and a refresh cycle of 20 milliseconds. This matrix is ​​transmitted back in real time between the edge computing node and each camera via 100 Mbps Ethernet, maintaining the overall network synchronization error within the range of 80 nanoseconds.

[0067] Following the hash marking within the buffer, an elliptic curve digital signature and a time integrity tag are appended to each frame, forming a chain of traceable data blocks. Finally, the data stream, after two levels of calibration and signing, is written to a circular buffer. The time integrity tag field contains four items: absolute timestamp, camera number, frame sequence number, and synchronization deviation value, which are hashed using SHA-256 to obtain a 32-byte digest. The elliptic curve digital signature uses the national commercial curve P-256, with the private key stored in the trusted platform module; the signature result is 64 bytes. To ensure traceability, the hash of the current frame tag and the hash of the previous frame tag are XORed to generate a chaincode field, which is placed at the beginning of the data block, forming a unidirectional chain structure. The processed multimodal fusion data block is written to a circular buffer with a depth of 4,912 using zero-copy pointers. Even when the buffer is fully loaded, the latency does not exceed 35 milliseconds, which can meet the continuous throughput requirements of deep learning inference windows.

[0068] The processing steps for fusing data sequences into a deep learning model include:

[0069] Temporal normalization is used to rearrange the fused data sequence into a fixed-length tensor set. Temporal normalization refers to mapping heterogeneous segments on the continuous time axis into a sequence unit of uniform length of 256 frames through linear resampling and sliding window segmentation, while maintaining the integrity of the original frame rate information. Each unit contains three channels: an image grayscale matrix (4K resolution downsampled by a quarter to obtain 960 x 540 pixels), a 3D coordinate tensor (64 points including personnel keypoints and tool endpoints, each with 3D coordinates), and a pose quaternion sequence (one quaternion per frame). The tensor set is packaged into a batch size of 16 and entered into the data pipeline, and is transferred to the graphics card using zero-copy shared memory technology, with a total latency of no more than 52 milliseconds. In the specific implementation example, the TensorRT pipeline is used for sequence rearrangement. Its rearrangement operator maintains temporal smoothness by mirroring the incomplete window sequence and normalizes all channels to the zero-to-one range to ensure consistent network input scale.

[0070] A multi-branch convolutional network is employed to extract key features of the personnel skeleton, insulation occlusion contour features, and multi-scale texture features of the work boundary. The multi-branch convolutional network comprises three parallel convolutional backbones. Each backbone shares the first layer of convolutional weights during the encoding phase to extract low-level edges before entering its own deeper structure. The personnel skeleton key point branch uses a fusion module of 3D convolution and graph convolution, stacking eight layers of 3D convolution and four layers of spatial graph convolution to capture the spatiotemporal topological relationships of key points. The insulation occlusion contour branch employs a high-resolution network structure preservation approach, combined with a dilated convolutional module with dilation rates of 1, 2, 4, and 8, to detect the edges of insulation covers and shields. The work boundary texture branch uses a scale attention pyramid, inputting four scale feature maps and weighting them pixel-wise along the channel dimension to extract fine-grained textures such as wires, crossarms, and ground wires. The three branches output feature dimensions of 1024, 1288, and 512, which are then uniformly mapped to a 1500-dimensional feature tensor via a fusion and convergence layer. After AMP hybrid precision optimization, the network convergence speed achieved an average accuracy of 96.1% for personnel detection and 92.8% for insulation shielding segmentation crossover ratio on a self-built live-line work dataset after 80,000 training steps.

[0071] The three types of features are weighted and coupled using an attention gating mechanism, and the resulting pose vector, occlusion mask, and out-of-bounds label are output in the self-attention decoding network. The attention gating mechanism first calculates the cross-correlation matrix of the three-branch features in the channel dimension to obtain cross-branch attention weights, and then uses a gating unit in the spatial dimension to suppress background interference based on the static mask of the conductor, generating a fused feature map. The self-attention decoding network consists of six layers of multi-head self-attention decoders, each layer containing eight heads and a feedforward residual block. It takes 1500-dimensional fused features and positional encoding as input and outputs three types of results: a pose vector of length 64x4, representing the quaternion pose of each joint of the skeleton; an occlusion mask with the same resolution as the input image, restored to 960x540 pixels using bilinear upsampling; and an out-of-bounds label, a probability vector of length 3, corresponding to three states: safe, warning, and dangerous. The decoder tail uses Softmax normalization to obtain an interpretable probability output, which is written to the backend risk index calculation module within one millisecond. The network inference test on the NVIDIA RTX A4500 graphics card showed an average latency of 3.5 milliseconds, which meets the requirements for real-time live-line operation risk monitoring.

[0072] The deep learning model is a hybrid architecture of spatiotemporal convolution and Transformer. The encoding end uses convolution to extract local details and ensure computational efficiency, while the decoding end uses self-attention to capture long-distance dependencies and output multi-task prediction results. The model has 78 million parameters and uses the AdamW optimizer and cosine annealing learning rate scheduling. In mixed precision training mode, it occupies 4,200 megabytes of GPU memory.

[0073] The steps for calculating the dynamic risk index include:

[0074] A 3D point cloud nearest neighbor search tree is used to perform Euclidean distance calculations on all grid nodes of the personnel skeleton key points, tool endpoints, and live wires, extracting the shortest distance and outputting the minimum safety gap vector. The 3D point cloud nearest neighbor search tree is an index built based on a kd-tree structure. Node elements include spatial coordinates and category labels. There are 64 personnel skeleton key points, 12 tool endpoints, and 3200 live wire grid nodes. Each element uses a 32-bit floating-point number to record the 3D coordinates. The Euclidean distance calculation traverses to the leaf node level in a single query, with a time complexity of logarithmic order. A full-frame query takes 600 microseconds with parallel acceleration from an NVIDIA RTX A4500 graphics card. The algorithm calculates the 3D distance for each pair of dissimilar nodes and updates the minimum value in real time for the current frame. The minimum distance and the corresponding node number are written into a minimum safety gap vector of length two elements. The first element of the vector represents the numerical value, and the second element represents the object pair. In the example test, the worker's right wrist joint is closest to the wire, only 0.19 meters away. The vector content is a combination of 0.19 meters and labels 1, 2, 3, 4, 5, and 6.

[0075] The minimum safety gap vector and attitude velocity vector are jointly mapped to the instantaneous energy potential field, and the gradient of the potential field is calculated using the finite difference method to form a cross-temporal tensor. The attitude velocity vector is obtained by calculating the skeleton quaternion output from the previous deep learning model and the inter-frame displacement difference, with a length of 64 x 3 and a unit of meters per second. The instantaneous energy potential field is based on the Coulomb-like potential formula E = k x q₁ x q₂ divided by r², where q₁ and q₂ are replaced by the personnel potential coefficient and the conductor voltage coefficient, respectively, and r is the value of the minimum safety gap vector. The finite difference method uses first-order forward differencing in the time dimension and three-point central differencing in the spatial dimension to calculate the gradient of the potential field scalar field, resulting in a vector field of uniform size. The cross-temporal tensor accumulates over sixteen frames on the time axis, with a shape of 16 x 64 x 3 and a precision of 32 floating-point.

[0076] The cross-temporal tensor is input into the dual-gated recurrent unit (DRU) model to simultaneously predict the probabilities of electric shock and fall, and a weighted average is used to obtain the comprehensive risk score. The DRU model adds a pose attention gate and a distance suppression gate to the update and reset gates of the gated recurrent unit (GRU). The hidden layer dimension is 128, the sequence length is 16, and the loss function is a linear combination of binary cross-entropy and mean squared error. During training, 80,000 training segments are used with a batch size of 64 and a learning rate of 1 x 10^-3. The mean absolute error of electric shock probability is 0.039, and the mean absolute error of fall probability is 0.041 on the validation set. During inference, the model outputs the electric shock probability Pt and the fall probability Pf. The comprehensive risk score R is calculated using the formula R = 0.6 x Pt + 0.4 x Pf, where the weighting coefficients are determined based on the proportion of historical accidents.

[0077] The comprehensive risk score is normalized to a range of 0 to 1 and published as a dynamic risk index. The normalization process employs a Min-Max strategy, with a lower bound of 0.10 and an upper bound of 0.90, using piecewise linear extrapolation to prevent extreme value saturation. The dynamic risk index D is calculated using the formula D = (R minus lower bound) divided by the upper bound minus the lower bound, with the result limited to the range of 0 to 1. The system refreshes D every 100 milliseconds and pushes it to the risk interlocking module. When D exceeds 0.75, an early warning is triggered; when D exceeds 0.9, the interlock is cut off. A field experiment showed that the dynamic risk index detected an arm exceeding its limit 520 milliseconds in advance, triggering an audible and visual interlock to prevent direct contact with live wires.

[0078] Step four specifically includes the following: A two-way time difference estimator is constructed using the timing pulses from the Global Navigation Satellite System (GNSS) and the local time base of the camera equipment, continuously outputting a synchronization residual sequence with a precision of hundreds of microseconds. At the hardware level, the two-way time difference estimator uses a 1Hz pulse and a 10MHz constant-frequency pulse as a reference. The rising edge of the timing pulse and the frame trigger signal generated by the camera's internal thermostatic crystal oscillator are simultaneously sent to a pair of time-to-digital converters (TD-SCDMAs). The converters have a resolution of 50 picoseconds. The forward channel records the one-way deviation from the satellite to the local station, and the return channel records the one-way deviation from the local station to the satellite. The average of the two data streams yields the round-trip residual. The system writes the residual sequence at a period of two milliseconds, with single-point errors controlled between 90 and 110 microseconds, generating 500 records per second. Each record includes a timestamp, channel identifier, and error value. In the field example, eight NVIDIA Jetson AGX Orin main control boards collect the residual sequence via Gigabit Ethernet, with an average bandwidth usage of 1.2 megabits per second.

[0079] Within a fixed-length sliding window, the root mean square (RMS) value of the synchronization residual sequence is calculated and compared with a preset threshold. When the RMS value exceeds the preset threshold, a self-calibration algorithm is triggered. This algorithm corrects the local time base using polynomial regression and broadcasts a correction vector to reduce the alignment error to the microsecond level. The sliding window length is 128 samples, equivalent to 256 milliseconds of historical data. The RMS value is calculated in real time within the window. The preset threshold is set to 60 microseconds, determined based on the three-standard-deviation principle, which can cover 98.75 percentile random jitter.

[0080] When the RMS exceeds the threshold, the self-calibration algorithm immediately calls the third-order polynomial regression model:

[0081] Δt=a·n 3 +b·n 2The formula +c·n+d is used to fit the residual of the most recent 256 milliseconds, outputting a correction vector for the next 16 frames. This correction vector is injected into the camera's phase controller via a delay-locked loop, fine-tuning the local time base frequency by 1 x 10^-7, reducing the maximum alignment error from 150 microseconds to 8 microseconds within 200 milliseconds. The correction result is synchronized to all edge nodes via multicast messages, ensuring that the distributed data streams return to microsecond-level synchronization.

[0082] After self-calibration, a redundant verification process is initiated. The corrected timestamp is bidirectionally compared with the distributed rubidium atomic clock signal to generate a time integrity hash, which is then written to a circular buffer. The distributed rubidium atomic clock, mounted on the top of the work vehicle, has a short-term stability better than 1 x 10^-12 and broadcasts a 10 MHz reference clock to all nodes. Edge nodes compare the corrected timestamp with the atomic clock timestamp, using a latitudinal weighted residual evaluation method to calculate the difference. Verification is considered successful if the two-minute moving average remains within five microseconds. After successful verification, the node calls the SecureHash256 function to hash the timestamp, node ID, and residual value, obtaining a 32-byte time integrity digest. This digest is concatenated with the original data frame header to form a 128-byte data block. The data block is written to a 1000-line circular buffer. The circular buffer uses a single-producer, multi-consumer lock free queue structure with a maximum write latency of 20 microseconds, ensuring that subsequent deep learning inference can read the double-verified data stream without blocking within one millisecond.

[0083] Step 5 specifically includes the following: Perform a dual-threshold comparison on the dynamic risk index. Once the dynamic risk index exceeds the alarm threshold, a dual-frequency sound wave is triggered at the on-site terminal, and a 120-lumen red pulse lamp is simultaneously lit to form a composite sound and light warning. The dual-threshold comparison uses the early warning threshold Twarn and the forced interlocking threshold Tlock, which are typically set to 0.75 and 0.95, respectively. The sound wave is emitted simultaneously by two electromagnetic buzzers, with frequencies alternating between 2,500 Hz and 4,000 Hz, a cycle of 120 milliseconds, and a sound pressure level of 90 Bel. The red pulse lamp uses six one-watt light-emitting diodes, driven by constant current, to pulse and flash in the 520 to 630 nanometer band, with a flashing period of 120 milliseconds, a duty cycle of 50%, and a luminous flux of 120 lumens, which can be visually identified by operators from 30 meters away.

[0084] Encrypted alarm frames are sent to the scheduling platform via a pre-defined time-sensitive network channel, and a heartbeat link is maintained for 50 milliseconds after the platform returns an acknowledgment signal. The time-sensitive network channel is based on IEEE 802.1AS precise synchronization and IEEE 802.1Qbv time slot scheduling, with a gigabit Ethernet fiber optic physical layer and AES 256-bit block encryption. The first alarm packet is 256 bytes in size, including an event timestamp, node number, risk index, and signature digest. The packet header priority is marked with 6 or 5 to ensure priority forwarding during network congestion. After receiving the packet, the scheduling platform returns an acknowledgment frame (ACK) with a field length of 64 bytes. The edge node then sends a heartbeat packet every 50 milliseconds to continuously monitor link connectivity and control consistency.

[0085] Upon receiving the confirmation signal, the safety switching script immediately runs, disconnecting the power circuit of the working device, locking the hydraulic valve of the lifting arm, and deploying the insulation barrier, thereby switching the working device to the safety interlock mode. The safety switching script runs with the highest priority in the embedded real-time operating system, with a total execution time of less than 200 milliseconds. The power circuit cuts off the three-phase 400V power supply through a solid-state relay, with a relay switching time of 3.5 milliseconds. The hydraulic valve locking command controls the proportional valve driver via the CAN bus, stopping the lifting arm at its current position with an error of less than 2 centimeters. The insulation barrier adopts a spring reset structure and deploys within 20 milliseconds after the electromagnetic lock is released, covering the potential contact path between personnel and wires.

[0086] During the safety interlock mode, the dynamic risk index is continuously recorded and compared with the reset threshold. When the dynamic risk index falls below the reset threshold, the interlock is automatically released and normal operation is restored. The reset threshold is set by multiplying the historical risk index average by the safety factor K. The historical risk index average is the moving average of the data over the last 30 minutes. The safety factor K ranges from 0.5 to 0.7, with a typical value of 0.62. When the risk index is detected to be below the reset threshold for 30 consecutive seconds and the trend is a negative slope, the release process is triggered. The release process is executed in reverse order: first, the insulation barrier is retracted, then the hydraulic valve is unlocked, and then the power circuit is restored. The entire reset sequence is completed in two seconds. After completion, a status zeroing report is sent to the scheduling platform through the time-sensitive network for continuous auditing.

[0087] A more detailed explanation of the dual thresholds is as follows: Dynamic Risk Index D: Continuously ranges from 0 to 1, with values ​​closer to 1 indicating higher risk. The system updates D every 100 milliseconds. Early Warning Threshold Twarn: 0.75 (configurable). Used for alerting and corrective actions. Forced Interlock Threshold Tlock: 0.95 (configurable). Used to immediately initiate hardware security interlocks.

[0088] Triggering rules are preset (with time retention and trend judgment). Early warning trigger: When D ≥ 0.75 and remains continuously for ≥ 200 milliseconds (at least 2 sampling periods), an audible and visual early warning is triggered. Interlock trigger (either of two conditions must be met to trigger): When D ≥ 0.95 and remains continuously for ≥ 100 milliseconds; or when the rising slope dD / dt ≥ 0.3 / second and D ≥ 0.9 (to prevent missed detection of "sudden surges"), an immediate interlock is triggered.

[0089] Alarm cancellation and interlock disengagement: Alarm cancellation: When D≤0.70 and remains continuously for ≥10 seconds, the audible and visual alarm will automatically stop (creating a hysteresis of 0.05 to avoid frequent jitter). Interlock disengagement: After interlocking, the following rules will be followed: Calculate the moving average of the risk index DP30min over the most recent 30 minutes, set the safety factor K=0.62 (configurable), and obtain the reset threshold Dreset=K×DP30min; when D≤Dreset and remains continuously for ≥30 seconds, the interlock will be disengaged in the predetermined order (first retract the insulation barrier → unlock the hydraulic valve → restore the power circuit).

[0090] Two specific scenario examples: Example 1: From early warning to interlocking, t=14:32:10.000, D=0.68 (safe); t=14:32:10.100, D=0.78; t=14:32:10.300, D is still ≥0.75, lasting for 200ms → triggering audible and visual warning; t=14:32:10.420, D=0.93, dD / dt≈0.75 / second; t=14:32:10.450, D=0.96 (≥0.95 and maintained for 100ms) → triggering interlocking: cutting off the power circuit, locking the hydraulic valve, and deploying the insulation barrier; after interlocking, the monitoring stage begins, and the timing for disinterlocking is determined according to Dreset=0.62×DP30min.

[0091] Example 2: Warning without interlocking: t=09:15:00.000, D=0.77; t=09:15:00.250, D is still ≥0.75, lasting for 250ms → warning triggered; t=09:15:03.000, D drops to 0.66 (<0.70), lasting for 12 seconds → warning automatically deactivated; the value never reaches 0.95, nor does it meet the condition of "rapid rise + ≥0.9" → no interlocking, event record is entered into the database for incremental learning.

[0092] Step Six: After deactivating the audible and visual warning, the dynamic risk index and corresponding work scenario data are stored in the safety database. Incremental learning is then performed based on the stored data to adaptively update the deep learning model. The safety database is a dedicated data storage cluster for long-term storage of risk-related data, comprising structured tables and object storage. Structured table fields include absolute timestamps, dynamic risk indexes, work scenario numbers, personnel skeleton key point coordinates, tool endpoint coordinates, pose quaternions, occlusion mask summaries, out-of-bounds labels, root mean square values ​​of synchronization residuals, device numbers, and signature digests. Object storage stores image frames and point cloud slices and indexes them using time-integrity hashes. The write process performs initial caching and deduplication on edge computing nodes, then submits in batches via encrypted channels, with a batch size of 160 records and a submission cycle of one second. Work scenario data refers to a time window data set related to a continuous work process, with a window length of 64 frames, covering 20 seconds before and after the audible and visual warning. Window numbers are used for easy retrieval and playback.

[0093] Incremental learning refers to a training strategy that uses new data to update parameters in small steps without discarding the capabilities of the existing model. The training data sampling rules include three types of samples: window samples with a dynamic risk index higher than the higher of the two thresholds; disputed samples with out-of-bounds labels and inconsistent interlocking results; and critical samples with the root mean square value of the synchronization residual close to the threshold. The three types of samples are organized into training batches in a ratio of 4:3:3, with each batch containing 128 windows. Sample preprocessing steps include keypoint trajectory smoothing, occlusion mask hole repair, point cloud sparsification to 12,000 points per frame, and unitization constraints on pose quaternions to ensure numerical stability. The annotation source adopts a human-machine collaborative mode. First, an automatic label generator provides initial labels, and then the review terminal revises the skeleton, boundaries, and occlusion areas on the keyframes and writes back the version number.

[0094] The deep learning model refers to the end-to-end network defined earlier, which combines multi-branch convolutional layers with spatiotemporal attention. It includes branches for the human skeleton, occlusion contour, and job boundary texture. After fusion, it outputs a pose vector, occlusion mask, and out-of-bounds label via a self-attention decoding network. Adaptive updates employ a two-stage strategy: the first stage fine-tunes only the parameters of the fusion and decoding layers, using a learning rate of 1 x 10^-4 and iterating for 2000 steps; the second stage introduces a lightweight adapter module and a normalization layer bias term for joint fine-tuning, using a learning rate of 5 x 10^-5 and iterating for 1000 steps. To prevent forgetting, a knowledge distillation loss is introduced. The new model output is matched with the previous version's output after temperature scaling in the pose vector, occlusion mask, and out-of-bounds label paths. The distillation temperature is 2, and the distillation weight is 0.35. The total loss function is a weighted sum of pose loss, mask loss, out-of-bounds loss, distillation loss, and a regularization term, with weights set at 0.3, 0.3, 0.2, 0.15, and 0.05, respectively.

[0095] After the parameters are updated, a two-step acceptance process is performed: offline verification and online shadow inference. Offline verification uses a fixed verification set and a subset of newly added data from the past seven days to calculate key indicators, including the mean attitude angle error, occlusion mask intersection-union ratio, out-of-bounds label accuracy, dynamic risk index, and correlation coefficient of interlocking events. Online shadow inference runs the old and new models in parallel on the on-site inference node. The new model only records output and does not participate in actual alarms or interlocks, continuously observing for 24 hours and recording alarm distribution, latency, and resource usage. After the shadow phase ends, the difference distribution of output between the two versions and the difference in threshold trigger counts are calculated. The difference statistics are written to the audit table and a version switch report is generated.

[0096] Model version management employs semantic numbering and immutable release packages. Each release package includes parameter snapshots, operator configurations, input specifications, and threshold recommendations. Timestamps and signature digests are recorded for each release and rollback. The deployment strategy uses a canary distribution approach, initially enabling a new version on a single node, observing it for two hours, and then expanding to all nodes. If any node experiences an input specification mismatch or inference latency exceeding a preset limit, it automatically reverts to the previous version. Threshold recommendations include warning thresholds, interlocking thresholds, and reset threshold safety factor ranges. Combined with an online calibration process, threshold calibration is completed within ten minutes of activation and fixed to session storage.

[0097] Data security and compliance controls include three mechanisms: access control lists restricting read and write permissions; time integrity hashing and elliptic curve digital signature binding records; and retention period and de-identification strategy configuration. Image frames are anonymized before being included in the training set. Anonymization includes masking facial regions and identity markers. The de-identification mask is stored separately from the original frame and referenced by a pointer. The default retention period is ninety days, after which the images are automatically archived to offline media, and a hash digest and metadata are retained in the index table.

[0098] In this embodiment, 14,000 windows of operational scenario data were collected over a month. After screening, 4,200 windows of samples were selected for incremental learning. The training time was 30 minutes. The resource environment consisted of one edge server, one NVIDIA RTX A4500 graphics processor, and 32 gigabytes of video memory occupying 4,200 megabytes. The shadow inference stage ran for 24 hours, recording output differences distributed between 0 and 0.06. After the version switch report was generated, grayscale expansion and version fixation were completed. The entire process was executed at a fixed cycle within the data loop after the audible and visual warnings were lifted. The cycle parameter is configurable at one hour.

[0099] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0100] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0101] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0102] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0103] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent risk identification and early warning for live-line work in distribution networks, characterized in that, Includes the following steps: Step 1: Deploy high-definition camera equipment, positioning sensor devices, and global satellite navigation system timing modules in the work area to simultaneously collect panoramic image data and location information of the work area; Step 2: Using the unified time reference provided by the global satellite navigation system's timing module, align the panoramic image data with the location information to generate a fused data sequence; Step 3: Input the fused data sequence into the deep learning model to analyze personnel protective posture, insulation shielding status and boundary crossing behavior, and calculate the dynamic risk index based on the relative spatial potential between personnel, tools and live wires; Step 4: Continuously monitor the integrity of the time synchronization information. When time synchronization drift is detected and the alignment error exceeds the preset threshold, execute the self-calibration algorithm to restore time synchronization and start the redundancy verification process. Step 5: When the dynamic risk index reaches the alarm threshold, output an audible and visual warning signal to the field terminal and dispatch platform, and automatically switch the working device to the safety interlock mode. Step 6: After the audible and visual warning is lifted, the dynamic risk index and the corresponding work scenario data are stored in the safety database, and incremental learning is performed based on the stored data to adaptively update the deep learning model. The steps for calculating the dynamic risk index include: A three-dimensional point cloud nearest neighbor search tree is used to perform Euclidean distance calculations on key points of the personnel skeleton, tool endpoints and all grid nodes of the live wires, extract the shortest distance and output the minimum safety gap vector. The minimum safety gap vector and the attitude velocity vector are jointly mapped to the instantaneous energy potential field, and the potential field gradient is calculated using the finite difference method to form a cross-temporal tensor. The cross-temporal tensor is input into the dual-gated cyclic unit model to simultaneously predict the probability of electric shock and the probability of falling, and the comprehensive risk score is obtained by weighted averaging. The comprehensive risk score is normalized to a range of zero to one and then published as a dynamic risk index. Step four specifically includes the following: A two-way time difference estimator is constructed by using the timing pulse of the global satellite navigation system and the local time base of the camera equipment to continuously output a synchronization residual sequence with a precision of hundreds of microseconds. The root mean square value of the synchronization residual sequence is calculated within a fixed-length sliding window and compared with a preset threshold. When the root mean square value exceeds the preset threshold, the self-calibration algorithm is triggered. The local time base is corrected by polynomial regression and the correction vector is broadcast to reduce the alignment error back to the microsecond level. After self-calibration is completed, a redundancy verification process is initiated, which compares the corrected timestamp with the distributed rubidium atomic clock signal bidirectionally, generates a time integrity hash, and writes the time integrity hash into a circular cache.

2. The intelligent risk identification and early warning method for live-line work in distribution networks according to claim 1, characterized in that, Step one specifically includes the following: Install high-definition camera equipment with automatic zoom and optical axis self-calibration functions, and use laser dot matrix to complete spatial calibration, with imaging distortion error controlled within 0.2 percent. A three-axis inertial combined positioning sensor is deployed at the edge of the camera's field of view, and centimeter-level coordinate unification is achieved using differential positioning of a dual-antenna global satellite navigation system. The timestamp of each frame of image is aligned with the timing module of the global satellite navigation system by hardware triggering signal to form an absolute time label and build a synchronous data stream. The final image data and location information achieved a preset target level of coverage in three-dimensional space.

3. The intelligent risk identification and early warning method for live-line work in distribution networks according to claim 1, characterized in that, Step two specifically includes the following: The timing module of the dual-frequency global satellite navigation system is used to drive a distributed pulse trigger to measure the frame time error of each camera device in real time and output a microsecond-level correction vector. The correction vector is dynamically injected into an adaptive synchronization algorithm that combines Kalman filtering and spline interpolation to perform continuous time axis alignment on panoramic image data and location information to generate an absolute timestamp. Immediately after alignment, the image frame and corresponding position information are written into the multimodal fusion buffer, and a time integrity flag is attached by a hash function. Finally, the processed multimodal fusion data stream is obtained and written into the circular buffer.

4. The intelligent risk identification and early warning method for live-line work in distribution networks according to claim 2, characterized in that, A two-stage time calibration process is established between the distributed pulse trigger and the adaptive synchronization algorithm, specifically including: By using timing pulses from the global satellite navigation system to perform coarse calibration, the time error of each camera device is reduced to within 20 microseconds; Micro-calibration is performed using a combination of high-stability time base and inertial positioning data from the camera device to generate a nanosecond-level deviation matrix, which is then fed back to the adaptive synchronization algorithm in real time. Following the hash mark in the buffer, an elliptic curve digital signature and a time integrity tag are appended to each frame to form a chain of traceable data blocks. Finally, the data stream that has been calibrated and signed at two levels is written to the circular buffer.

5. The intelligent risk identification and early warning method for live-line work in distribution networks according to claim 4, characterized in that, The processing steps for fusing data sequences into a deep learning model include: The fused data sequence is rearranged into a fixed-length tensor set using temporal normalization. A multi-branch convolutional network was used to extract key point features of the personnel skeleton, insulation occlusion contour features, and multi-scale texture features of the work boundary. The three types of features are weighted and coupled through an attention gating mechanism, and the pose vector, occlusion mask, and out-of-bounds label are output in the self-attention decoding network.

6. The intelligent risk identification and early warning method for live-line work in distribution networks according to claim 5, characterized in that, Step five specifically includes the following: The dynamic risk index is compared with two thresholds. Once the dynamic risk index exceeds the alarm threshold, a dual-frequency sound wave is triggered at the on-site terminal and a 120-lumen red pulse lamp is lit simultaneously to form a composite sound and light warning. An encrypted alarm frame is sent to the scheduling platform through a preset time-sensitive network channel, and a heartbeat link is maintained for fifty milliseconds after the platform returns an acknowledgment signal; Upon receiving the confirmation signal, the safety switching script is immediately executed to disconnect the power circuit of the working device, lock the hydraulic valve of the lifting arm, and deploy the insulation barrier, thereby switching the working device to the safety interlock mode. During the safety interlock mode, the dynamic risk index is continuously recorded and compared with the reset threshold. When the dynamic risk index falls below the reset threshold, the interlock is automatically released and normal operation is restored. The reset threshold is set by multiplying the historical average risk index by the safety factor K.

Images