Intelligent safety management system for transformer substation and implementation management method

Abstract

The invention discloses an intelligent safety management system for a transformer substation and an implementation management method. The system comprises a remote intelligent patrol subsystem, an intelligent analysis subsystem and a client interaction subsystem. The remote intelligent inspection subsystem is deployed at a substation end and is composed of an inspection host, a multi-modal data acquisition device and a communication module. The intelligent analysis subsystem is deployed in a cloud server and comprises an image preprocessing module, an anomaly recognition module, an equipment classification module and a risk assessment module, the recognition precision is improved through multi-algorithm fusion, the image quality problem in a complex environment is solved through CycleGAN preprocessing, the anomaly recognition accuracy is improved through the combination of YOLOv5 and CycleGAN, and the risk assessment accuracy is improved. By combining with a random forest classifier, the equipment classification accuracy is improved; closed-loop management from data acquisition to decision support is intelligently realized in the whole process, so that the inspection efficiency is improved. According to the invention, by fusing multiple algorithms and the Internet of Things technology, the substation inspection efficiency and the safety management level are improved.

Description

Technical Field

[0001] This invention relates to the field of substation technology, specifically to a smart safety management system for substations and its implementation method. Background Technology

[0002] With the continuous expansion of the power grid, the number of substations has increased dramatically. Traditional operation and maintenance methods, primarily relying on manual inspections, suffer from low efficiency and delayed detection of safety hazards. Existing technologies, intelligent inspection systems based on single algorithms struggle to handle complex scenarios (such as image noise interference in high-altitude and desert regions) and lack dynamic monitoring capabilities throughout the equipment's entire lifecycle. Furthermore, substation equipment is diverse in type and similar in appearance, and traditional identification models achieve a classification accuracy of less than 80%, failing to meet the demands of refined management. Existing technologies exhibit the following shortcomings:

[0003] Weak image preprocessing capabilities: Images acquired in harsh environments suffer from problems such as blurriness and uneven lighting, which traditional filtering methods cannot effectively enhance;

[0004] Low accuracy in anomaly identification: The identification rate for minor equipment defects (such as insulator cracks) and personnel violations in complex backgrounds is less than 75%;

[0005] Poor system compatibility: The equipment layout of substations of different sizes varies greatly, and the existing system lacks flexible deployment options;

[0006] Insufficient decision support: It can only provide fault alarms, but cannot provide risk level assessments or operation and maintenance decision suggestions. Summary of the Invention

[0007] To address the aforementioned issues, this invention provides a smart safety management system for substations and a method for implementing that system. By integrating multiple algorithms with Internet of Things (IoT) technology, it aims to improve the efficiency of substation inspections and enhance safety management.

[0008] According to a first aspect of the present disclosure, a smart safety management system for substations is provided, including a remote intelligent inspection subsystem, an intelligent analysis subsystem, and a client interaction subsystem;

[0009] The remote intelligent inspection subsystem is deployed at the substation end and includes an inspection host, a multimodal data acquisition device, and a communication module. The inspection host is used for device access and control command issuance; the multimodal data acquisition device is used for dynamically acquiring images, videos, or audio; and the communication module uses the TCP / IP communication protocol for data transmission.

[0010] The intelligent analysis subsystem is deployed on a cloud server and includes an image preprocessing module, an anomaly recognition module, a device classification module, and a risk assessment module.

[0011] The client interaction subsystem adopts a B / S architecture and is used for data visualization, remote control and log management;

[0012] The image preprocessing module is based on the CycleGAN algorithm and includes a generator and a discriminator. The generator adopts an improved U-Net structure, which includes an encoding path and a decoding path containing multiple convolutional blocks. The encoding path and the decoding path are fused by skip connections to preserve image edges and texture details. The discriminator adopts a PatchGAN structure to determine the authenticity of local regions of the image.

[0013] The anomaly detection module is implemented based on the YOLOv5s network;

[0014] The equipment classification module uses a random forest classifier combined with a deep convolutional neural network to identify electrical equipment.

[0015] In some embodiments, the image preprocessing module includes two generators (G, F) and two discriminators (D). x D Y );

[0016] Generator G is used to map the source domain X to the target domain Y, and generator F is used to reverse map the target domain Y back to the source domain X, realizing bidirectional conversion between image domains;

[0017] The loss function of the image preprocessing module includes:

[0018] The cycle consistency loss is calculated using the following formula:

[0019] L cyc (G,F)=E x～X |F(G(x))-x||1+E y～Y |G(F(y))-y||1

[0020] Where x is the source domain image, y is the target domain image, and E is the expectation operator;

[0021] The formula for calculating the counter-loss is:

[0022] L adv (G,D Y ) = E y～Pdata(y) [logD Y (y)]+E x～Pdata(x) [log(1-D Y (G(x)))

[0023] L adv (F,D X ) = E x～Pdata(x) [logD X (x)]+Ey～Pdata(y) [log(1-D X (F(y)))

[0024] The total loss function is:

[0025] L(G,F,D X D Y ) = L adv (G,D Y )+L adv (F,D X )+λL cyc (G,F)

[0026] Where λ is the weight of the cycle consistency loss.

[0027] In some embodiments, the anomaly identification module includes a dataset construction unit, a model training unit, and a real-time detection unit; the dataset construction unit collects substation anomaly samples including various equipment failures and personnel violations; the model training unit, based on the YOLOv5s network, enhances the dataset images, performs adaptive anchor box calculation, sets hyperparameters, and trains and optimizes the network model based on the bounding box loss function; the real-time detection unit extracts images in real time and performs image preprocessing, feature extraction, feature fusion, and target prediction before outputting the results, wherein the feature fusion uses upsampling and downsampling to fuse features at different scales.

[0028] In some embodiments, the device classification module includes the following steps:

[0029] A training dataset was constructed using on-site collected data and the PowerImage database;

[0030] Preprocess the images in the training dataset;

[0031] Feature extraction is performed on the preprocessed training dataset, including using a pre-trained ResNet50 network, which is pre-trained on the ImageNet dataset.

[0032] The extracted features are filtered based on the variance inflation factor between features;

[0033] Random forest model training: Bootstrap sampling is used to generate several training subsets; parameters are constructed for each decision tree, including maximum depth, minimum number of split samples, and minimum number of leaf node samples, with the Gini index as the splitting criterion;

[0034] Feature random selection: Multiple features are randomly selected when each node splits;

[0035] Model Evaluation: Cross-validation was used to evaluate the model.

[0036] The feature importance scoring formula is as follows:

[0037]

[0038] Where T is the number of decision trees, OOB t (f) represents the out-of-bag error of the t-th tree. Let f be the out-of-bag error after feature f is randomly permuted.

[0039] In some embodiments, the risk assessment module includes the following steps:

[0040] Based on historical failure data, the equipment failure probability P is calculated using the Weibull distribution model. fault ;

[0041] The severity of the environment, S, is obtained by weighting scores based on various environmental parameters. env ;

[0042] The historical response time T is obtained based on the average response time of similar anomalies in recent times. response ;

[0043] The formula for calculating the risk index is: R = α·P fault +β·S env +γ·T response .

[0044] In some embodiments, the device failure probability P fault Calculated based on the equipment's service life t: P fault = 1 - exp(-(t / η) β ), where η is the characteristic life and β is the shape parameter; when the equipment has experienced a similar failure, P fault Multiply by a factor of 1 to 3.

[0045] In some embodiments, the environmental severity S env The weighted score includes temperature-weighted score, humidity-weighted score, and dust concentration-weighted score.

[0046] In some embodiments, the historical response time T response =min(1, actual response time / 60); If the response time of the most recent similar anomalies is not greater than the preset threshold, T response Multiply by a factor of 0.5 to 1.

[0047] In some embodiments, the multimodal data acquisition device includes any one or more of the following: a high-definition camera, a wheeled robot, a rail-mounted robot, a drone, and a voiceprint monitoring device that supports abnormal sounds.

[0048] According to a second aspect of the present disclosure, a method for implementing management using the substation intelligent safety management system as described above is provided, comprising the following steps:

[0049] Step 1: Deploy the remote intelligent inspection subsystem, including equipment installation, network configuration, and calibration;

[0050] Step 2: Image preprocessing, including inputting the original image, calculating sharpness using the Laplacian operator, and starting enhancement when the gradient mean is not greater than a preset threshold; calling the CycleGAN model to generate the enhanced image, and saving the original image and the enhanced image for comparison;

[0051] Step 3: Anomaly identification, including: loading the pre-trained YOLOv5 model weights and setting the confidence threshold; performing batch detection on the pre-processed images and outputting the anomaly category, location, and confidence level; generating an anomaly work order for the same anomaly detected in multiple consecutive frames;

[0052] Step 4: Equipment classification and risk assessment, including: cropping images of abnormal areas, inputting them into a random forest classifier to obtain equipment types; combining equipment types, operating parameters, and environmental data to calculate the risk index R, and triggering alarms based on the risk index.

[0053] This disclosure provides a substation intelligent safety management system and its implementation method, which addresses the pain points of traditional inspections through full-process optimization of image enhancement, intelligent recognition, equipment classification, and data visualization. The system includes a remote intelligent inspection subsystem, an intelligent analysis subsystem, and a client interaction subsystem. The remote intelligent inspection subsystem is deployed at the substation and consists of an inspection host, multimodal data acquisition equipment, and a communication module. This system improves recognition accuracy through multi-algorithm fusion; CycleGAN preprocessing addresses image quality issues in complex environments; and the combination of YOLOv5 and random forest improves anomaly recognition accuracy by 12.3% and equipment classification accuracy by 8.5%. Flexible deployment adapts to diverse scenarios, with two access methods meeting the needs of substations of different sizes; edge node type reduces data transmission volume by more than 60%. Full-process intelligent management achieves closed-loop management from data acquisition to decision support, increasing inspection efficiency by 300% and reducing maintenance costs by 40%. Dynamic risk assessment, combined with a risk index model of equipment status and environmental parameters, achieves a warning accuracy rate of over 90%.

[0054] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0055] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0056] Figure 1 This is a schematic diagram of the substation intelligent safety management system in an embodiment of the present invention;

[0057] Figure 2 This is a schematic diagram of the remote intelligent inspection subsystem structure in an embodiment of the present invention;

[0058] Figure 3 This is a schematic diagram of the intelligent analysis subsystem structure in an embodiment of the present invention;

[0059] Figure 4 This is a flowchart of the intelligent analysis subsystem in an embodiment of the present invention;

[0060] Figure 5 This is a comparison chart of the CycleGAN preprocessing effects in the embodiments of this application;

[0061] Figure 6 This is a flowchart illustrating the implementation of the electrical equipment identification module based on a random forest classifier in this application embodiment;

[0062] Figure 7 This is a flowchart of the client software architecture in an embodiment of this application;

[0063] Figure 8 This is a schematic diagram of the risk level visualization interface in the embodiments of this application;

[0064] Figure 9 This is a schematic diagram of the test report switching and selection interface in the embodiments of this application. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present invention are shown in the drawings, not the entire structure.

[0066] Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. The process can be terminated when its operation is complete, but may also have additional steps not included in the figures. The process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0067] This invention provides a smart safety management system for substations and a method for implementing that system. Specific embodiments are as follows:

[0068] A substation intelligent safety management system 100, such as Figure 1 As shown, it includes a remote intelligent inspection subsystem 110, an intelligent analysis subsystem 120, and a client interaction subsystem 130;

[0069] The remote intelligent inspection subsystem 110 is deployed at the substation station end, such as... Figure 2 As shown, the system includes a patrol host 111, a multimodal data acquisition device 112, and a communication module 113. The patrol host 111 is used for device access and control command issuance; the multimodal data acquisition device 112 is used to dynamically acquire images, videos, or audio; and the communication module 113 uses the TCP / IP communication protocol for data transmission.

[0070] The multimodal data acquisition device 112 includes any one or more of the following: high-definition camera, wheeled robot, rail-mounted robot, drone, and voiceprint monitoring device that supports abnormal sounds.

[0071] Specifically, the remote intelligent inspection subsystem 110 is deployed at the substation end and includes an inspection host 111, a multimodal data acquisition device 112, and a communication module 113. The inspection host 111 uses an industrial-grade server with an Intel Xeon W-1290 CPU, 64GB of memory, and a 2TB SSD storage capacity. It has the capability to operate 24 / 7 without interruption and supports up to 32 devices for access and control command issuance. Its core functions include device control, data caching, and time synchronization. Device control can issue inspection path adjustment commands to robots, drones, etc., with a response time of ≤500ms. The data cache can store 48 hours of data locally (based on 32 video streams) when the network is down, and automatically synchronize it to the cloud after the network is restored. Time synchronization is achieved through an NTP server to calibrate the time with each acquisition device, with an error of ≤1ms. The multimodal data acquisition device 112 includes a high-definition camera, mobile inspection equipment (wheeled robot, rail-mounted robot, drone, adaptable to complex indoor and outdoor terrain), and a voiceprint monitoring device.

[0072] The high-definition cameras include a 4K dome network camera with 20x optical zoom and infrared night vision distance ≥100 meters; a fixed network camera with a resolution of 1920×1080 and a frame rate of 25fps; a PTZ network camera with a horizontal rotation range of 0-360° and a vertical rotation range of -30° to 90°; a thermal imaging camera with a temperature measurement range of -20℃ to 150℃ and a temperature measurement accuracy of ±2℃; and an explosion-proof camera with an explosion-proof rating of Ex dⅡCT6 and a protection rating of IP68, enabling all-round, all-weather monitoring of the equipment.

[0073] The wheeled robot uses a tracked walking mechanism, has a climbing ability of ≥30°, a battery life of ≥8 hours, a maximum moving speed of 1.5m / s, and is equipped with a 360° panoramic camera and ultrasonic obstacle avoidance sensors with a detection distance of 0.1-5 meters.

[0074] The rail-mounted robot is compatible with standard C-shaped steel, model C100, with a running speed of 0.8m / s and an automatic charging function, with a charging time of ≤2 hours.

[0075] The drone is a quadcopter with a flight time of 30 minutes, a maximum flight altitude of 120 meters, and is equipped with a 1-inch CMOS camera.

[0076] The voiceprint monitoring device has a sampling rate of 44.1kHz and a frequency response of 20Hz-20kHz, and supports the identification of abnormal sounds, such as equipment noises and discharge sounds.

[0077] In the environmental and condition monitoring device, the acoustic fingerprint sensor is used to identify abnormal equipment noises, while temperature and humidity sensors, air pressure sensors, and dust concentration sensors are used to collect environmental parameters. Equipment status data such as voltage and current are obtained from the SCADA system via the IEC 61850 protocol. The communication module supports dual-link redundant transmission of 5G and fiber optics, employs a TCP / IP protocol stack, has a data transmission latency of ≤20ms, and uses AES-256 encryption algorithm to ensure data transmission security and prevent information leakage and tampering.

[0078] The communication module 113 supports 5G, Sub-6GHz frequency band, peak rate ≥1Gbps, dual-link redundant transmission with optical fiber (single-mode optical fiber, transmission rate 10Gbps), adopts TCP / IP protocol stack, data transmission latency ≤20ms, and has data encryption function (AES-256 encryption algorithm).

[0079] The remote intelligent inspection subsystem has two deployment methods: the edge node type is equipped with an NVIDIA Jetson AGXXavier edge computing module, which supports parallel processing of 8 video streams and is suitable for large substations of 220kV and above; the direct access type uses an embedded gateway, which only realizes data acquisition and uploading and is suitable for small and medium-sized substations of 110kV and below.

[0080] The intelligent analysis subsystem 120 is deployed on a cloud server, such as Figure 3 As shown, it includes an image preprocessing module 121, an anomaly recognition module 122, an equipment classification module 123, and a risk assessment module 124;

[0081] Specifically, such as Figure 4The flowchart of the intelligent analysis subsystem is shown. The intelligent analysis subsystem 120 is deployed on a cloud server. The GPU is NVIDIA A100 with 80GB of video memory and the CPU is Intel Xeon Platinum 8380. It includes an image preprocessing module, an anomaly recognition module, a device classification module and a risk assessment module. Each module is deployed through Docker containerization, supports horizontal scaling, and can process up to 100 video streams concurrently.

[0082] The client interaction subsystem 130 adopts a B / S architecture and is used for data visualization, remote control, and log management. Specifically, the client interaction subsystem 130 adopts a B / S architecture, is developed based on the Django 4.2 framework, and supports access from browsers such as Chrome 90+, Firefox 88+, and Edge 90+. It includes a data visualization module (using ECharts 5.4 to implement chart rendering), a remote control module (supporting WebSocket real-time communication), and a log management module (logs are stored in JSON format and retained for ≥3 years).

[0083] The image preprocessing module 121 is based on the CycleGAN algorithm and includes a generator and a discriminator. The generator adopts an improved U-Net structure, including an encoding path and a decoding path containing multiple convolutional blocks. Features from the encoding and decoding paths are fused through skip connections to preserve image edges and texture details. The discriminator adopts a PatchGAN structure to determine the realism of local image regions. Figure 5 The image shows a comparison of the preprocessing effects of CycleGAN.

[0084] The anomaly detection module 122 is implemented based on the YOLOv5s network;

[0085] The equipment classification module 123 uses a random forest classifier combined with a deep convolutional neural network to identify electrical equipment. For example... Figure 6 The flowchart shown is for the implementation of the electrical equipment identification module based on the random forest classifier.

[0086] Image preprocessing module 121 is built based on the CycleGAN algorithm, including two generators (G, F) and two discriminators (D). x D Y );

[0087] Generator G is used to map source domain X (low-quality substation images, including foggy, low-light, and noisy images) to target domain Y (sharpened images), and generator F is used to reverse map target domain Y back to source domain X, realizing bidirectional conversion between image domains;

[0088] Specifically, the generator adopts an improved U-Net structure. The encoding path contains 6 convolutional blocks, each consisting of two 3×3 convolutional layers, a batch normalization layer, and a LeakyReLU activation function with a stride of 2. The decoding path contains 6 deconvolutional blocks, each consisting of one 4×4 deconvolutional layer, a batch normalization layer, and a ReLU activation function with a stride of 2. By using skip connections, the features of the i-th layer of the encoding path and the (6-i)-th layer of the decoding path are fused to preserve image edges and texture details.

[0089] The discriminator adopts the PatchGAN structure, which contains 5 convolutional layers: the input is a 256×256×3 image, which is convolved to output feature maps of 128×128×64, 64×64×128, 32×32×256, 16×16×512, and 8×8×1 respectively. Finally, a 30×30 discriminant matrix is ​​output through Sigmoid activation to determine the authenticity of local regions in the image.

[0090] The loss function of the image preprocessing module 121 includes:

[0091] The cycle consistency loss is calculated using the following formula:

[0092] L cyc (G,F)=E x～X |F(G(x))-x||1+E y～Y |G(F(y))-y||1

[0093] Where x is the source domain image, y is the target domain image, and E is the expectation operator;

[0094] The formula for calculating the counter-loss is:

[0095]

[0096] The total loss function is:

[0097] L(G,F,D X D Y ) = L adv (G,D Y )+L adv (F,D X )+λL cyc (G,F)

[0098] Wherein, λ is the weight of the cycle consistency loss, and in the preferred embodiment, λ = 10.

[0099] The model training parameters included: batch size 1, number of iterations 200, optimizer Adam (β1 = 0.5, β2 = 0.999, initial learning rate 0.0002), and the learning rate linearly decayed to 0 after 100 iterations. After processing, the peak signal-to-noise ratio (PSNR) of the image increased from 20dB to ≥28dB, and the structural similarity index (SSIM) increased from 0.6 to ≥0.9.

[0100] The anomaly detection module 122 is implemented based on the YOLOv5 algorithm and includes a dataset construction unit, a model training unit, and a real-time detection unit. The dataset construction unit collects substation anomaly samples, including various equipment faults and personnel violations. The model training unit is based on the YOLOv5s network and performs image enhancement, adaptive anchor box calculation, hyperparameter setting, and training and optimization of the network model based on the bounding box loss function. The real-time detection unit extracts images in real time and performs image preprocessing, feature extraction, feature fusion, and target prediction before outputting the results. The feature fusion uses upsampling and downsampling to fuse features at different scales.

[0101] Specifically, in the preferred embodiment, the substation anomaly samples collected by the dataset construction unit cover:

[0102] 32 types of equipment failures: insulator damage, surge arrester leakage, transformer oil leakage, switchgear overheating, cable joint overheating, abnormal noise from instrument transformers, capacitor bulging, circuit breaker failure to operate, poor contact of disconnecting switches, busbar icing, terminal block burning, battery leakage, low gas pressure in GIS equipment, voltage transformer open circuit, current transformer saturation, reactor vibration, fuse blowing, grounding grid corrosion, insulator contamination, bushing rupture, water accumulation in cable trenches, abnormal indicator lights in control cabinets, jamming of operating mechanisms, gas relay activation, excessively high oil temperature, abnormal oil level, SF6 gas leakage, insulator flashover, broken conductor strands, corrosion of hardware, tilting of towers, and foundation settlement.

[0103] Eight types of personnel violations: not wearing a safety helmet, not wearing insulated clothing, smoking, climbing over fences, working on live lines without supervision, entering restricted areas by mistake, operating in violation of regulations, and carrying fire sources;

[0104] The dataset contains 25,000 images (640×640 resolution), including 20,000 images in the training set and 5,000 images in the validation set. The images are labeled in VOC format using the labelimg tool. Each bounding box contains a category label (e.g., “blue_helmet”, “insulator_damage”) and coordinate information (xmin, ymin, xmax, ymax).

[0105] The model training unit is based on a YOLOv5s network (depth factor 0.33, width factor 0.5), and the following training strategy is adopted:

[0106] Data augmentation: Mosaic enhancement (random scaling, cropping, and stitching of 4 images), random horizontal flipping (probability 0.5), random angle rotation (-10° to 10°), HSV color dithering (saturation and brightness shift ±10%), and noise addition (Gaussian noise, variance 0.01).

[0107] Adaptive anchor box calculation: Based on the size of the labeled boxes in the training set, generate 9 anchor boxes ([10,13],[16,30],[33,23],[30,61],[62,45],[59,119],[116,90],[156,198],[373,326]);

[0108] Hyperparameter settings: epochs=300, batch_size=8, img_size=640, learning rate 0.001, warm-up strategy for the first 10 epochs, momentum 0.937, weight decay 0.0005;

[0109] Loss functions: CIoU_Loss was used for bounding box loss, BCEWithLogitsLoss for confidence loss, and BCEWithLogitsLoss for classification loss. During model training, the average accuracy on the validation set reached 88.5%, with an accuracy of 92.3% for "not wearing a safety helmet", 89.7% for "damaged insulator" and 87.6% for "oil stains on equipment".

[0110] The processing flow of the real-time detection unit is as follows:

[0111] Video frame extraction: Extract images from the RTSP stream at 25 frames per second, and resize to 640×640 using bilinear interpolation;

[0112] Image preprocessing: normalization (pixel value divided by 255), channel conversion (BGR to RGB);

[0113] Feature extraction: The 640×640×3 image was sliced ​​into 320×320×12 segments using the Focus structure, and multi-scale features (80×80, 40×40, 20×20) were extracted using the CSP1_X structure.

[0114] Feature fusion: An FPN+PAN architecture is adopted to fuse features at different scales through upsampling and downsampling;

[0115] Target prediction: The output layer predicts bounding boxes (x, y, w, h), confidence scores, and class probabilities, and uses DIOU-NMS (threshold 0.45) to filter redundant boxes;

[0116] Output results: Abnormal areas are marked, with green boxes indicating equipment failures and red boxes indicating personnel violations, along with confidence levels. Detection speed is ≥30fps.

[0117] The equipment classification module 123 uses a random forest classifier combined with a deep convolutional neural network (DCNN) to identify electrical equipment, including the following steps:

[0118] Training dataset construction:

[0119] Equipment categories: 3000 insulators, 2000 silicone tubes, 3000 breathers, 500 transformers, and 2000 power transmission poles. Image sources include on-site data collected by partner companies and the PowerImage database.

[0120] Image preprocessing: Crop the device area, remove the background, normalize the size, scale the long side to 256 pixels, scale the short side proportionally, fill empty parts with 0, convert to grayscale to a single-channel image, and apply Gaussian filtering (kernel size 3×3, σ=1.0).

[0121] Feature extraction stage:

[0122] A pre-trained ResNet50 network was used, pre-trained on the ImageNet dataset, and the last fully connected layer was removed, retaining the 2048-dimensional feature vector output by the average pooling layer.

[0123] Feature selection: Calculate the variance inflation factor (VIF) between features, remove highly redundant features with VIF > 10, and retain 15 key features (such as edge gradient, texture entropy, and shape moments).

[0124] Random Forest Model Training:

[0125] 1000 training subsets were generated using Bootstrap sampling (70% sampling rate);

[0126] The construction parameters for each decision tree are: maximum depth 20, minimum number of split samples 10, minimum number of leaf node samples 5, and the splitting criterion is the Gini index.

[0127] Feature random selection: 45 features are randomly selected when each node splits;

[0128] Model evaluation: Using 5-fold cross-validation, the average accuracy rate was 85.6%, with 88% accuracy rate for identifying power transmission poles and 86% accuracy rate for identifying transformers.

[0129] The formula for feature importance scoring is:

[0130]

[0131] Where T is the number of decision trees, OOB t (f) represents the out-of-bag error of the t-th tree. The out-of-bag error is denoted by feature f after random permutation. The classification process involves inputting a 256×256 single-channel image, preprocessing, extracting features using ResNet50, matching 15 feature subsets, random forest voting, and outputting the device type (insulator, transformer, etc.) and confidence score. A score ≥0.8 indicates valid identification, with an average accuracy of 85.6%.

[0132] Testing phase process:

[0133] Input the image to be recognized (256×256 single channel) and perform the same preprocessing steps as in the training phase;

[0134] Input the ResNet50 network to extract 2048-dimensional features and match them with 15 feature subsets selected during the training phase.

[0135] The features are input into the trained random forest model, and the majority class is selected by voting, with each tree casting one vote, and the device category is output.

[0136] Output the classification result and confidence level. A confidence level ≥ 0.8 is considered a valid identification.

[0137] The risk assessment module 124 is implemented by including the following steps:

[0138] Based on historical failure data, the equipment failure probability P is calculated using the Weibull distribution model. fault ;

[0139] The severity of the environment, S, is obtained by weighting scores based on various environmental parameters. env ;

[0140] The historical response time T is obtained based on the average response time of similar anomalies in recent times. response ;

[0141] The formula for calculating the risk index is: R = α·P fault +β·S env +γ·T response .

[0142] Equipment failure probability P fault Calculated based on the equipment's service life t: P fault = 1 - exp(-(t / η) β ), where η is the characteristic life and β is the shape parameter; when the equipment has experienced a similar failure, P fault Multiply by the positive correlation coefficient.

[0143] Environmental severity S envThe weighted score includes temperature-weighted score, humidity-weighted score, and dust concentration-weighted score.

[0144] Historical response time T response =min(1, actual response time / 60); If the response time of the most recent similar anomalies is not greater than the preset threshold, T response Multiply by the positive correlation coefficient.

[0145] The system may also include an energy efficiency analysis module, which calculates energy efficiency indicators based on equipment operating data and assesses the equipment's operating status. The formula is:

[0146]

[0147] Where P i U is the output power of the device. i I i For the operating voltage and current, t i Let n be the running time and n be the number of devices.

[0148] The remote intelligent inspection subsystem supports two access methods:

[0149] Edge node type:

[0150] Hardware configuration: Industrial control computer, CPU i7-12700E, memory 32GB, SSD 1TB, equipped with NVIDIA Jetson AGX Xavier edge computing module, GPU 512-core Volta, computing power 32TOPS;

[0151] Software features: A lightweight version of the locally deployed intelligent analysis subsystem (supports parallel processing of 8 video streams), with features such as automatic inspection path planning (based on the A* algorithm, path accuracy ±5cm), multi-device collaborative control (time synchronization error ≤100ms when robots and drones conduct collaborative inspections), and offline caching (data can be stored locally for 48 hours when the network is down).

[0152] Applicable scenarios: Large hub substations with voltage level ≥220kV, covering an area of ​​≥50,000 square meters, and with ≥500 pieces of equipment;

[0153] Direct access type:

[0154] Hardware configuration: Embedded gateway, supporting 4G / 5G communication modules;

[0155] Software functionality: It only performs data collection and uploading, uses the MQTT protocol, has QoS=1, does not have local analysis capabilities, and device control commands need to be sent from the cloud.

[0156] Response performance: The latency from receiving the instruction to executing the action is ≤500ms, and the data upload bandwidth is ≥2Mbps;

[0157] Applicable scenarios: Small and medium-sized substations with voltage level ≤110kV, floor area ≤10,000 square meters, and number of equipment ≤100 units.

[0158] The data visualization module includes:

[0159] Real-time monitoring panel:

[0160] Equipment operating parameters are displayed as follows: voltage (0-500kV, accuracy ±0.5%), current (0-2000A, accuracy ±1%), temperature (-40℃ to 150℃, refresh rate 1 time / second), and equipment status (running, stopped, faulty).

[0161] Abnormal alarm information: alarm type (equipment failure, personnel violation), time of occurrence (accurate to the second), location (latitude and longitude ±5m), processing status (unprocessed, processing, resolved);

[0162] Interface layout: It adopts a 1920×1080 resolution, with a device list tree on the left, a real-time video window on the right, and an alarm scroll bar at the bottom;

[0163] Historical trend analysis:

[0164] Time granularity: Supports time interval selection of 1 minute, 5 minutes, 1 hour, and 1 day;

[0165] Chart types: Line chart (parameter trend), bar chart (fault statistics), pie chart (equipment type percentage);

[0166] Data backtracking: Supports querying historical data for the past 30 days, which can be exported to Excel or PDF format, including data points, maximum value, minimum value, and average value;

[0167] Risk Level Map:

[0168] Visualization method: The substation plan map is drawn based on SVG, and the equipment locations are marked with actual coordinates;

[0169] Risk coding: Red (RGB 255,0,0) indicates high risk (risk index ≥ 0.7), yellow (RGB 255,255,0) indicates medium risk (0.3 ≤ risk index < 0.7), and green (RGB 0,255,0) indicates low risk (risk index < 0.3).

[0170] Interactive features: Clicking the device icon allows you to view detailed information (model, runtime, historical faults), and supports map zooming (zoom ratio 1:100 to 1:1000);

[0171] Alarm and early warning mechanisms:

[0172] Triggering conditions: Temperature exceeds 120% of the equipment's rated value, current exceeds 110% of the rated value, or personnel violation is detected;

[0173] Notification methods: client pop-up, SMS, and audible / visual alarms;

[0174] Tiered handling: Level 1 alarms (such as fire, electric shock) must be responded to within 5 minutes, and Level 2 alarms (such as equipment overheating) must be responded to within 30 minutes.

[0175] In a preferred embodiment, the workflow of the intelligent analysis subsystem includes:

[0176] S1. Multimodal data acquisition:

[0177] Image data: The camera collects a set of panoramic images every 30 minutes; the robot collects a close-up image of the equipment every 5 meters during inspection; and the drone collects an image of the power transmission line every 10 meters along the flight path.

[0178] Environmental data: temperature and humidity sensor (sampling interval 1 minute), air pressure sensor (sampling interval 5 minutes), dust concentration sensor (sampling interval 30 minutes);

[0179] Equipment status data: Voltage, current, and switch status are obtained from the substation SCADA system via the IEC 61850 protocol (sampling interval 1 second);

[0180] S2. Image Preprocessing:

[0181] Quality assessment: Calculate image sharpness (mean gradient magnitude), and activate CycleGAN enhancement when sharpness < 0.3;

[0182] Enhancement processing: A dehazing algorithm (atmospheric scattering model) is used for foggy images, histogram equalization is used for low-light images, and bilateral filtering is used for noisy images;

[0183] Output: Enhanced image (saved as JPEG format, 90% compression quality) and quality assessment report;

[0184] S3. Anomaly Detection:

[0185] Batch processing: 100 images are input into the YOLOv5 model in batches, and the detection time is ≤3 seconds per batch;

[0186] Results filtering: Remove detection boxes with a confidence level < 0.5, and mark anomalies detected three times consecutively at the same location as "confirmed anomalies";

[0187] Location: Combine device GPS coordinates with image pixel coordinates to calculate abnormal locations (error ≤ 1 meter);

[0188] S4. Equipment Classification:

[0189] The detected abnormal device area is cropped (the detection frame boundary is expanded by 20%) and then input into the device classification module;

[0190] Output device type and classification confidence level; when the confidence level is <0.8, mark it as "Pending Confirmation" and trigger manual review; S5. Risk Assessment:

[0191] Equipment failure probability P fault Based on historical fault data, the result was calculated using the Weibull distribution model (shape parameter 2.5, scale parameter 1000).

[0192] Environmental severity S env Weighted scores for temperature, humidity, and dust concentration (weights of 0.4, 0.3, and 0.3 respectively), with a score range of 0-1;

[0193] Historical response time T response The average response time (in minutes) for similar anomalies over the past 3 months, normalized to 0-1;

[0194] The formula for calculating the risk index is:

[0195] R = α·P fault +β·S env +γ·T response

[0196] Among them, P fault S represents the probability of equipment failure. env To rate the severity of the environment, T response The historical response time is given by α, β, and γ, which are weighting coefficients with α = 0.5, β = 0.3, and γ = 0.2.

[0197] S6. Data Feedback:

[0198] Generate an inspection report, including a list of anomalies, risk levels, and handling recommendations, in PDF format;

[0199] The report is pushed to the client and simultaneously entered into the blockchain to ensure its immutability. Figure 7 The client software architecture flowchart shown is as follows: Figure 8 The diagram shows a visual interface illustrating the risk level.

[0200] Update the device health score, with a maximum score of 100. Each abnormality deducts 5-20 points. For example... Figure 9 The diagram shows the interface for switching and selecting test reports.

[0201] In the preferred embodiment, the specific values ​​and adjustment rules for each parameter in the risk index calculation formula are as follows:

[0202] Equipment failure probability P fault :

[0203] Calculated based on the equipment's service life t (years): P fault = 1 - exp(-(t / η) β ), where η is the characteristic life (15 years for transformers, 10 years for insulators), and β is the shape parameter, ranging from 1.8 to 2.2, adjusted according to the equipment type;

[0204] When the equipment has experienced a similar failure, P fault Multiply by a factor of 1.5;

[0205] Environmental severity S env :

[0206] Temperature score: 0.8 points when the temperature is >40℃ or <-20℃; 0.5 points when the temperature is between 25℃ and 40℃ or between -10℃ and -20℃; 0.2 points for all other cases.

[0207] Humidity score: 0.8 points when humidity > 90%; 0.5 points when humidity is between 60% and 90%; 0.2 points for all other conditions.

[0208] Dust concentration score: When the concentration is >10mg / m³ 3 At that time, a score of 0.8 was obtained; 5-10 mg / m² 3 In the first case, you get 0.5 points; in the other cases, you get 0.2 points.

[0209] S env = 0.4 × Temperature Score + 0.3 × Humidity Score + 0.3 × Dust Concentration Score;

[0210] Historical response time T response :

[0211] T response =min(1, actual response time / 60), where the actual response time is in minutes (if it exceeds 60 minutes, it is calculated as 60 minutes);

[0212] If the response time for the last three similar anomalies is less than 10 minutes, T response Multiply by a factor of 0.8;

[0213] Weighting coefficient adjustment:

[0214] α, β, and γ are adjusted annually based on the substation safety assessment results. When the annual accident rate is greater than 0.1 incidents per station, α increases by 0.1 (maximum 0.7) and β decreases by 0.1 (minimum 0.2).

[0215] Another embodiment is a method for implementing management using the substation intelligent safety management system described above, comprising the following steps:

[0216] Step 1: Deploy the remote intelligent inspection subsystem, including equipment installation, network configuration, and calibration;

[0217] Step 2: Image preprocessing, including inputting the original image, calculating sharpness using the Laplacian operator, and starting enhancement when the gradient mean is not greater than a preset threshold; calling the CycleGAN model to generate the enhanced image, and saving the original image and the enhanced image for comparison;

[0218] Step 3: Anomaly identification, including: loading the pre-trained YOLOv5 model weights and setting the confidence threshold; performing batch detection on the pre-processed images and outputting the anomaly category, location, and confidence level; generating an anomaly work order for the same anomaly detected in multiple consecutive frames;

[0219] Step 4: Equipment classification and risk assessment, including: cropping images of abnormal areas, inputting them into a random forest classifier to obtain equipment types; combining equipment types, operating parameters, and environmental data to calculate the risk index R, and triggering alarms based on the risk index.

[0220] In a preferred embodiment, the management method is implemented using the substation intelligent safety management system described above, including the following steps:

[0221] Step 1: Deploy the remote intelligent inspection subsystem:

[0222] Equipment installation: Cameras should be installed at a height of 3-5 meters, covering ≥90% of the equipment; robot tracks should be laid along the equipment with a spacing of ≤10 meters; drone take-off and landing points should be selected in open areas, avoiding high-voltage equipment;

[0223] Network configuration: Set up VLAN isolation, separate the data acquisition network from the office network, and configure firewall rules (allow only 5G or fiber optic port communication);

[0224] Calibration: Time synchronization of the camera (NTP server, error ≤1ms), and path calibration of the robot (error ≤5cm);

[0225] Step 2: Image preprocessing:

[0226] Input the original image, calculate the sharpness using the Laplacian operator, and start enhancement when the gradient mean is <50;

[0227] The CycleGAN model is invoked to generate a sharpened image, and the original and enhanced images are saved for comparison.

[0228] Perform quality checks on the enhanced image. If PSNR < 25dB, perform enhancement again (up to 3 times).

[0229] Step 3: Anomaly Detection

[0230] Load the pre-trained YOLOv5 model weights (best.pt) and set the confidence threshold to 0.6;

[0231] Perform batch detection on preprocessed images and output anomaly category, location, and confidence level;

[0232] For the same anomaly detected in 3 consecutive frames, generate an anomaly work order (including processing priority: high / medium / low);

[0233] Step 4: Equipment Classification and Risk Assessment

[0234] Cropping out images with abnormal regions and inputting them into a random forest classifier to obtain the device type;

[0235] Calculate the risk index R by combining equipment type, operating parameters, and environmental data;

[0236] When R ≥ 0.7, a Level 1 alarm is triggered; when 0.3 ≤ R < 0.7, a Level 2 alarm is triggered.

[0237] Step 5: Client Interaction:

[0238] Real-time display of alarm information and risk maps, and push of work orders to maintenance personnel's mobile APP;

[0239] After the maintenance personnel complete the task, they upload the results, and the system automatically updates the work order status.

[0240] A daily "Safety Operation Report" is generated, which includes indicators such as the number of anomalies, the handling rate, and risk trends.

[0241] Alternatively, the system can exist as a computer-readable storage medium, in which case the storage medium is an SSD (capacity ≥ 1TB, NVMe interface), formatted with the EXT4 file system, and the stored programs include:

[0242] Remote data acquisition module program: developed based on Python 3.9, using OpenCV 4.5 to acquire images, using the pyserial library to read sensor data, and supporting multi-threading (maximum 32 threads);

[0243] The intelligent analysis module program implements CycleGAN and YOLOv5 based on PyTorch 1.12, and uses scikit-learn 1.0 to implement random forest, supporting GPU acceleration (CUDA 11.6).

[0244] Client-side interactive program: The front-end is developed based on Vue 3, and the back-end is developed based on Django 4.2. WebSocket is used to achieve real-time communication, and Redis is used to cache hot data (caching time is 5 minutes).

[0245] When the program is executed by the processor, it implements the steps of the method of claim 9 and has the following characteristics:

[0246] Fault tolerance: When a module crashes, it will automatically restart (restart time ≤ 30 seconds) and log the error.

[0247] Scalability: Supports plug-in development (reserves algorithm interfaces for integration with new detection models);

[0248] Compatibility: Compatible with operating systems such as Windows Server 2019, Ubuntu 20.04, and CentOS 8.

[0249] In addition to the modules described above, the system may also include other components. Some of these components are not related to the contents of this disclosure, so their illustrations and descriptions are omitted here and will not be repeated.

[0250] Here is a specific example and a verification example:

[0251] Taking the actual deployment and operation of a 220kV hub substation as an example, this paper provides a detailed explanation of the technical implementation of this system. The substation covers an area of ​​approximately 60,000 square meters and contains more than 500 electrical devices, including transformers, switchgear, and insulators. The surrounding environment is complex, with climatic characteristics such as high temperatures and frequent fog. Traditional manual inspections have an average annual delay of more than 48 hours in detecting faults, making intelligent upgrades urgently needed.

[0252] During the system deployment phase, the remote intelligent inspection subsystem is configured as an edge node: the inspection host uses an Intel Xeon W-1290 industrial-grade server, equipped with 64GB of memory and 2TB of SSD storage, and connects to 8 high-definition cameras, 2 wheeled robots, and 1 quadcopter drone via an industrial switch. The thermal imaging cameras are deployed in the transformer area, covering a temperature range of -20℃ to 150℃ with an accuracy controlled within ±2℃, acquiring one frame of infrared image time series every 30 minutes; the wheeled robots use a tracked walking mechanism, with a climbing ability of up to 35° and a flight time of 9 hours, acquiring close-up images of the equipment every 5 meters along a preset track, and simultaneously recording ultrasonic obstacle avoidance data; the drone is equipped with a 1-inch CMOS camera, flying along the power transmission line three times a day, taking one image of the line every 10 meters, with a maximum flight altitude controlled at 100 meters to ensure safety. The communication module adopts a dual-link redundancy design of 5G and fiber optic. The 5G module operates in the Sub-6GHz band with a measured peak rate of 1.2Gbps, while the fiber optic link has a transmission rate of 10Gbps. The infrared image, device status and other data are encrypted and transmitted using the AES-256 encryption algorithm. The data transmission latency is consistently within 15ms after testing. The local cache capacity supports 48 hours of full data storage when the network is down.

[0253] The intelligent analysis subsystem is deployed on an NVIDIA A100 server in the cloud, using Docker containerization to deploy its four main modules. The image preprocessing module, designed for the foggy and low-light environment of substations, employs the CycleGAN algorithm for image enhancement: the generator G receives low-quality images from foggy conditions, extracts features through a 6-layer coded convolutional block (3×3 kernel, stride 2), reconstructs a clear image (target domain Y) through a 6-layer decoded deconvolutional block (4×4 kernel, stride 2), and a skip connection fuses the features from the 3rd coded and 3rd decoded layers, preserving minute details such as insulator cracks; the discriminator D... x The PatchGAN structure is adopted, which outputs a 30×30 discriminant matrix for a 256×256 input image. The model is optimized by adversarial loss and cycle consistency loss (λ=10). In actual tests, the PSNR of foggy images improved from 21dB to 29dB and the SSIM improved from 0.58 to 0.92 after processing, meeting the requirements for subsequent recognition.

[0254] The anomaly detection module is trained on a YOLOv5s model. The dataset contains 25,000 substation images, covering 32 types of equipment faults and 8 types of personnel violations. During training, Mosaic enhancement technology is used to randomly scale and stitch together four different equipment images, combined with HSV color dithering to simulate lighting changes, adaptively generating nine anchor boxes to fit different target sizes. After model deployment, preprocessed images are sampled at 25 frames per second. A Focus structure is used to slice the 640×640×3 image into 320×320×12 feature maps. Multi-scale features of 80×80, 40×40, and 20×20 are extracted using a CSP1_X structure, and finally fused and output through an FPN+PAN architecture. In actual testing, the accuracy rate for "switchgear overheating" reached 91.2%, and the accuracy rate for "not wearing a safety helmet" reached 92.3%, with a single frame processing time of 0.03 seconds, meeting real-time requirements.

[0255] The equipment classification module employs a ResNet50 + Random Forest combination: after cropping images of abnormal regions, a pre-trained ResNet50 network extracts 2048-dimensional features, calculates the variance inflation factor between features, and selects 15 key features (VIF≤8), including edge gradient and texture entropy. The Random Forest model constructs 1000 decision trees, each with a maximum depth of 20 and a minimum of 5 samples per leaf node, generating a training subset through Bootstrap sampling. In the testing phase, 500 equipment images were classified, achieving recognition accuracies of 84%, 86%, 85%, 85%, and 88% for five types of equipment, including insulators and transformers, respectively. The effective recognition rate for confidence levels ≥0.8 reached 92%.

[0256] The risk assessment process employs a three-level index model: equipment failure probability P fault Based on the Weibull distribution calculation, for example, a 110kV insulator that has been in operation for 7 years has a characteristic life of 10 years, a shape parameter of 2.0, and P... fault =0.343; Environmental severity S env Based on real-time monitoring data, the score for high temperature (42℃) and high humidity (90%) in summer is 0.4×0.8+0.3×0.8+0.3×0.5=0.71; historical response time T response The average of 30 minutes over the past 3 months was taken and normalized to 0.5. The final risk index R = 0.5 × 0.343 + 0.3 × 0.71 + 0.2 × 0.5 = 0.487, which is judged as medium risk, triggering a level 2 alarm.

[0257] The client-side interaction subsystem implements a B / S architecture using the Django 4.2 framework. The data visualization module uses ECharts 5.4 to draw the real-time monitoring panel. The device list tree on the left is categorized by voltage level, and the video window on the right simultaneously displays four real-time video feeds. A scroll bar at the bottom displays alarm information from the past 24 hours. The risk level map uses SVG to draw a substation plan view, with high-risk areas (R≥0.7) highlighted in red and flashing. Clicking the transformer icon allows viewing its operating parameter curves and historical fault records. Maintenance personnel can adjust the robot's inspection path via the remote control module, with a command response time of ≤0.8 seconds. The log management module stores operation records in JSON format and retains them for 3 years.

[0258] During the six-month trial operation of the system, a total of 120,000 frames of infrared image time series were collected, 32,000 low-quality images were processed, 46 equipment failures were identified, including 12 minor defects, and 18 personnel violations were identified. The average failure detection time was shortened to 1.2 hours, the inspection efficiency was improved by 300%, and the operation and maintenance cost was reduced by 42%, which verified the effectiveness and reliability of the system in complex substation environments.

[0259] In this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a step or method that comprises a list of elements includes not only those elements but also other elements not expressly listed or inherent to such a step or method.

[0260] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A smart safety management system for substations, characterized in that, It includes a remote intelligent inspection subsystem, an intelligent analysis subsystem, and a client interaction subsystem; The remote intelligent inspection subsystem is deployed at the substation end and includes an inspection host, a multimodal data acquisition device, and a communication module. The inspection host is used for device access and control command issuance; the multimodal data acquisition device is used for dynamically acquiring images, videos, or audio; and the communication module uses the TCP / IP communication protocol for data transmission. The intelligent analysis subsystem is deployed on a cloud server and includes an image preprocessing module, an anomaly recognition module, a device classification module, and a risk assessment module. The client interaction subsystem adopts a B / S architecture and is used for data visualization, remote control and log management; The image preprocessing module is based on the CycleGAN algorithm and includes a generator and a discriminator. The generator adopts an improved U-Net structure, which includes an encoding path and a decoding path containing multiple convolutional blocks. The encoding path and the decoding path are fused by skip connections to preserve image edges and texture details. The discriminator adopts a PatchGAN structure to determine the authenticity of local regions of the image. The anomaly detection module is implemented based on the YOLOv5s network; The equipment classification module uses a random forest classifier combined with a deep convolutional neural network to identify electrical equipment.

2. The intelligent safety management system for substations according to claim 1, characterized in that, The image preprocessing module includes two generators (G, F) and two discriminators (D). x D γ ); Generator G is used to map the source domain X to the target domain Y, and generator F is used to reverse map the target domain Y back to the source domain X, realizing bidirectional conversion between image domains; The loss function of the image preprocessing module includes: The cycle consistency loss is calculated using the following formula: L cyc (G,F)=E x～X |F(G(x))-x||1+E y～Y |G(F(y))-y||1 Where x is the source domain image, y is the target domain image, and E is the expectation operator; The formula for calculating the counter-loss is: L adv (G,D Y )=E y～Pdata(y) [logD Y (y)]+E x～Pdata(x) [log(1-D Y (G(x))) L adv (F,D X )=E x～Pdata(x) [logD X (x)]+E y～Pdata(y) [log(1-D X (F(y))) The total loss function is: L(G,F,D X ,D Y )=L adv (G,D Y )+L adv (F,D X )+λL cyc (G,F) Where λ is the weight of the cycle consistency loss.

3. The intelligent safety management system for substations according to claim 1, characterized in that, The anomaly identification module includes a dataset construction unit, a model training unit, and a real-time detection unit. The dataset construction unit collects substation anomaly samples, including various equipment failures and personnel violations. The model training unit, based on the YOLOv5s network, enhances the dataset images, performs adaptive anchor box calculation, sets hyperparameters, and trains and optimizes the network model based on the bounding box loss function. The real-time detection unit extracts images in real time and performs image preprocessing, feature extraction, feature fusion, and target prediction before outputting the results. The feature fusion uses upsampling and downsampling to fuse features at different scales.

4. The intelligent safety management system for substations according to claim 1, characterized in that, The equipment classification module includes the following steps: A training dataset was constructed using on-site collected data and the PowerImage database; Preprocess the images in the training dataset; Feature extraction is performed on the preprocessed training dataset, including using a pre-trained ResNet50 network, which is pre-trained on the ImageNet dataset. The extracted features are filtered based on the variance inflation factor between features; Random forest model training: Bootstrap sampling is used to generate several training subsets; parameters are constructed for each decision tree, including maximum depth, minimum number of split samples, and minimum number of leaf node samples, with the Gini index as the splitting criterion; Feature random selection: Multiple features are randomly selected when each node splits; Model evaluation: Cross-validation was used to evaluate the model. The feature importance scoring formula is as follows: Where T is the number of decision trees, OOB t (f) represents the out-of-bag error of the t-th tree. Let f be the out-of-bag error after feature f is randomly permuted.

5. The intelligent safety management system for substations according to claim 1, characterized in that, The risk assessment module includes the following steps: Based on historical failure data, the equipment failure probability P is calculated using the Weibull distribution model. fault ; The severity of the environment, S, is obtained by weighting scores based on various environmental parameters. env ; The historical response time T is obtained based on the average response time of similar anomalies in recent times. response ; The formula for calculating the risk index is: R = α·P fault +β·S env +γ·T response .

6. The intelligent safety management system for substations according to claim 5, characterized in that, Equipment failure probability P fault Calculated based on the equipment's service life t: P fault = 1 - exp(-(t / η) β ), where η is the characteristic life and β is the shape parameter; when the equipment has experienced a similar failure, P fault Multiply by a factor of 1 to 3.

7. The intelligent safety management system for substations according to claim 5, characterized in that, Environmental severity S env The weighted score includes temperature-weighted score, humidity-weighted score, and dust concentration-weighted score.

8. The intelligent safety management system for substations according to claim 5, characterized in that, Historical response time T response =min(1, actual response time / 60); If the response time of the most recent similar anomalies is not greater than the preset threshold, T response Multiply by a factor of 0.5 to 1.

9. The intelligent safety management system for substations according to claim 1, characterized in that, The multimodal data acquisition equipment includes any one or more of the following: high-definition camera, wheeled robot, rail-mounted robot, drone, and voiceprint monitoring device that supports abnormal sounds.

10. A method for management using the intelligent safety management system for substations as described in claims 1-7, characterized in that, Includes the following steps: Step 1: Deploy the remote intelligent inspection subsystem, including equipment installation, network configuration, and calibration; Step 2: Image preprocessing, including inputting the original image, calculating sharpness using the Laplacian operator, and starting enhancement when the mean gradient is not greater than a preset threshold; The CycleGAN model is called to generate enhanced images, and the original and enhanced images are saved for comparison. Step 3: Anomaly identification, including: loading the pre-trained YOLOv5 model weights and setting the confidence threshold; performing batch detection on the pre-processed images and outputting the anomaly category, location, and confidence level; generating an anomaly work order for the same anomaly detected in multiple consecutive frames; Step 4: Equipment classification and risk assessment, including: cropping images of abnormal areas, inputting them into a random forest classifier to obtain equipment types; combining equipment types, operating parameters, and environmental data to calculate the risk index R, and triggering alarms based on the risk index.

Images