Substation / converter station post-earthquake damage identification method and system

Abstract

The invention discloses a transformer substation / converter station post-earthquake damage identification method and system, and belongs to the technical field of electric power facilities. In order to solve the problems that existing post-earthquake electric power identification is not comprehensive enough and low in accuracy, the invention provides a post-earthquake damage identification method for a transformer substation / converter station, and the method comprises the steps: obtaining an appearance image, vibration displacement and a temperature image of in-station electrical equipment after an earthquake, and obtaining initial appearance image data, initial vibration displacement data and initial temperature image data; preprocessing the initial data to obtain intermediate image data, intermediate vibration feature vectors and intermediate thermal image data; and inputting the intermediate image data, the intermediate vibration feature vector and the intermediate thermal image data into a damage identification model for feature extraction and fusion, and outputting a damage identification result of the electrical equipment. By fusing the appearance image, the vibration data and the temperature image, the change of the internal state of the equipment can be captured, and more comprehensive damage detection is realized; and the detection accuracy is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of power facility safety monitoring, and more specifically, relates to a method and system for identifying post-earthquake damage in a transformer substation / converter station. Background Art

[0002] Earthquakes, as highly destructive natural disasters, pose a serious threat to the safety of equipment in converter and substation stations. Under the influence of earthquakes, these equipment undergoes complex dynamic responses, such as displacement, deformation, and even short-circuiting of transformer windings; cracking and damage of insulators, resulting in reduced insulation performance; and malfunction of circuit breaker operating mechanisms. Such equipment damage can not only cause power outages, disrupting the normal operation of society and life, but can also result in significant economic losses.

[0003] Traditional equipment damage identification methods rely primarily on manual inspections, which present numerous drawbacks. In large substations or converter stations, the equipment is vastly numerous and widely distributed. Manual inspections require significant manpower, material resources, and time, resulting in extremely low efficiency. Furthermore, manual inspections are susceptible to factors such as the inspector's expertise, work experience, and fatigue, making it difficult to comprehensively and accurately detect subtle equipment damage and prone to missed or false detections. Existing technologies combine image recognition with neural network models, but existing single-channel feature-based intelligent image systems for substation / converter station damage identification often fail to accurately determine the actual damage level. While image detection of equipment appearance excels at capturing external damage, it is insensitive to internal damage, early-stage hidden dangers, and functional abnormalities, and is susceptible to factors such as lighting. Equipment vibration monitoring is sensitive to structural changes and can reflect internal dynamic behavior, but it struggles to precisely locate the source of damage or distinguish between different types of damage, and is susceptible to interference from environmental vibrations. Equipment inspection is sensitive to electrical faults, mechanical friction overheating, and leakage and overcooling, and while it can indicate functional abnormalities and potential risks, it is insensitive to purely mechanical damage and is susceptible to factors such as ambient temperature.

[0004] Corresponding improvements have also been made to address the above-mentioned issues, such as Chinese patent application number CN202411009922.9, published on September 17, 2024. This patent discloses a method and system for analyzing the vulnerability and resilience of substation structures under strong earthquakes, specifically relating to the technical field of substation structure analysis; by dividing the substation structure and equipment into monitoring areas and collecting physical data before and after the strong earthquake in real time, pre-processing and clustering the data to identify anomalies, extracting and evaluating hidden damage characteristics, and finally dividing each monitoring area into high-impact and low-impact areas based on the degree of damage impact, and formulating corresponding risk assessment reports and maintenance plans. The shortcomings of this patent are: it relies on pre-divided areas, lacks dynamic adaptability, and the accuracy of damage identification needs to be improved.

[0005] Another example is Chinese patent application number CN202411690508.9, published on March 21, 2025. This patent discloses a method, system, and medium for intelligent image recognition of the operating status of traction substation equipment, including: obtaining initial image data of the operating status of the substation equipment; preprocessing the initial image data to obtain intermediate image data; obtaining actual labels of samples from the intermediate image data, and using the intermediate image data and the corresponding actual labels of the samples as a set of training samples to form a training sample set; constructing a state recognition model based on a deep learning neural network, and optimizing the weights and thresholds of the feature extraction unit and the recognition unit based on the particle swarm algorithm combined with the training sample set, and finally completing the training of the state recognition model; detecting the intermediate image data of the operating status of the substation equipment through the trained state recognition model. The shortcoming of this patent is that relying solely on images cannot accurately identify the internal damage of the equipment in the face of an earthquake scenario. Summary of the Invention

[0006] 1. Problems to be solved

[0007] To address the incompleteness and low accuracy of existing post-earthquake power identification methods, this paper provides a method and system for identifying damage in substations and converter stations. By integrating appearance images, vibration data, and temperature images, this method can capture changes in the internal state of equipment, enabling more comprehensive damage detection. This high level of accuracy enables rapid post-earthquake repairs and provides reliable technical support for post-earthquake emergency response in power systems.

[0008] 2. Technical solution

[0009] To solve the above problems, the present invention adopts the following technical solutions.

[0010] A method for identifying post-earthquake damage in a transformer substation / converter station comprises the following steps:

[0011] S1: Obtain post-earthquake appearance images, epicenter vibration displacement, and post-earthquake temperature images of electrical equipment in the station to obtain initial appearance image data, initial vibration displacement data, and initial temperature image data;

[0012] S2: Preprocessing the initial appearance image data, initial vibration displacement data, and initial temperature image data to obtain intermediate image data, intermediate vibration feature vectors, and intermediate thermal image data;

[0013] S3: Input the intermediate image data, intermediate vibration feature vector and intermediate thermal image data into the damage identification model for feature extraction and fusion, and output the damage identification results of the electrical equipment.

[0014] The above technical solution provides direct information on the surface of electrical equipment through appearance images, mechanical conditions and dynamic characteristics through vibration displacement, and thermal conditions and potential thermal issues through temperature images. Data is collected simultaneously from both the exterior and interior of the equipment for subsequent analysis, avoiding the incomplete detection that previously occurred by only collecting external data and ignoring changes in internal conditions. Furthermore, by fusing multimodal data through a damage identification model, the location and extent of damage can be more accurately determined. This precise damage location capability provides more targeted guidance for post-earthquake repairs, improving repair efficiency and quality. Therefore, the entire substation / converter station post-earthquake damage identification method enables more comprehensive detection of electrical equipment and more accurate damage location capabilities, enabling rapid post-earthquake repairs, ensuring the rapid and stable operation of power equipment, and providing reliable technical support for post-earthquake emergency response in the power system.

[0015] Furthermore, the damage identification model in step S3 includes a first channel, a second channel, a third channel, a fusion layer and a fully connected layer;

[0016] The input of the first channel is the intermediate image data, and the output is the image feature subnet F img ;

[0017] The input of the second channel is the intermediate vibration feature vector, and the output is the vibration feature subnet F vir ;

[0018] The input of the third channel is the intermediate thermal image data, and the output is the thermal image feature subnet F ir ;

[0019] The input of the fusion layer is the image feature subnet F img , vibration feature subnet F vir and thermal image feature subnet F ir , the output is the modal weights of the three channels and the weighted fusion unit F fused ; Weighted fusion unit F fused for:

[0020] F fused =w img ·F img +w vib ·F vib +w ir ·F ir

[0021] Among them, w img is the fusion weight of the first channel; w vib is the fusion weight of the second channel; w ir is the fusion weight of the third channel;

[0022] The input of the fully connected layer is the weighted fusion unit F fused , the output is the damage identification result of the electrical equipment; the damage identification result of the electrical equipment includes the damage status of the electrical equipment and the first confidence of the damage identification model for the identification result.

[0023] By adopting the above technical solution, three channels are used in the damage identification model to independently process different data inputs, preventing noise contamination between different modes and ensuring the accuracy of feature extraction of each channel; and each channel can independently optimize the network structure without the need to reconstruct the entire model. The channel structure can be adjusted according to different equipment types, making the entire damage identification model highly maintainable and scalable; at the same time, the output result of the damage identification model introduces the concept of first confidence, so that low-confidence results can be introduced into manual review to reduce the risk of misjudgment; and it can realize self-reflection and continuous optimization of the damage identification model, thereby improving the recognition accuracy of the damage identification model; the entire damage identification model has high comprehensiveness, robustness and accuracy.

[0024] Furthermore, the first channel adopts a convolutional neural network, which includes five convolutional layers and two spatial attention modules. The two spatial attention modules are respectively embedded in the third and fifth convolutional layers, and the image feature subnet F is generated by generating a spatial weight mask. img ;

[0025] The second channel uses 1D-CNN to process the intermediate vibration feature vector to generate the vibration feature subnet F vir ;

[0026] The third channel uses a fully connected layer to process the intermediate thermal image data to generate a thermal image feature subnet F ir .

[0027] By adopting the above technical solution, different network structures are selected according to the characteristics of different data to realize feature extraction, thereby ensuring the accuracy of feature extraction to the greatest extent while taking into account the cost.

[0028] Furthermore, when the damage identification model is trained, the weight parameters of the first channel, the weight parameters of the second channel, the weight parameters of the third channel and the parameters of the fusion layer in the damage identification model are optimized using a particle swarm optimization algorithm.

[0029] Adopting the above technical solution, the particle swarm optimization algorithm is used to optimize the parameters of the damage identification model. This not only improves the model's recognition accuracy and training efficiency, but also enhances the model's multimodal fusion, adaptability, and generalization capabilities. This step enables the model to understand the status of electrical equipment more comprehensively and accurately, providing strong support for the intelligent operation and maintenance of substations or converter stations. Through the optimization of the PSO algorithm, the damage identification model can better adapt to the complex and changing substation environment, providing reliable protection for the safe and stable operation of equipment.

[0030] Furthermore, the method further includes step S4: combining the damage identification result of the electrical equipment with the real-time physical feature vector Input into the decision engine to obtain the damage type of the electrical equipment and the second confidence level of the decision engine for the damage type of the electrical equipment; is the temperature gradient, which is directly detected by the damage identification model; ΔE is the weekly growth rate of vibration energy, which is calculated according to the following formula:

[0031]

[0032] Among them, E current is the current vibration energy value, E last week It is the vibration energy value of the previous week.

[0033] By adopting the above technical solution and introducing a decision engine, the damage identification results of electrical equipment based on the damage identification model are integrated with the real-time feature vector that reflects the physical nature of the electrical equipment. This can effectively identify and suppress anomalies or errors in a single information source, achieve a more robust and reliable assessment of the health status of electrical equipment, and lay a solid foundation for accurate maintenance decisions. When the damage identification model has an abnormal output, the decision engine can use physical characteristics to effectively identify, suppress or mark the anomaly, prevent the transmission of erroneous diagnosis, greatly improve the availability and fault tolerance of the entire system, and ensure that valuable diagnostic information can still be provided under complex working conditions.

[0034] Furthermore, in step S4, the decision engine further includes determining the damage type of the electrical equipment, the second confidence level of the decision engine for the damage type of the electrical equipment, and the real-time physical feature vector. Generate early warning strategy:

[0035] Output "urgent" warning;

[0036] Output "high risk" warning;

[0037] ELSE outputs monitoring instructions.

[0038] Where d1, d2 are the elements in the equipment damage type D set, e1, e2 are the vibration energy limits, t1, t2 are the temperature gradient limits; c1 and c2

[0039] By adopting the above technical solution, the decision engine realizes risk quantification and grading through multi-condition joint judgment, dynamic threshold adaptation and closed-loop feedback, thereby achieving layered response and optimizing resource allocation; multimodal cross-validation reduces the risk of wrong decision-making; and parameter configurability adapts to the characteristics of different devices and has strong adaptability; the whole makes the operating cost low while taking into account both accuracy and efficiency.

[0040] Furthermore, after the early warning strategy is generated, different maintenance strategies are formulated according to the level of the early warning strategy:

[0041] When an "emergency" warning is issued, operations and maintenance personnel must complete the power outage and maintenance within 72 hours, immediately isolate the damaged electrical equipment, dismantle the equipment, and replace damaged parts. At the same time, they must conduct a comprehensive inspection of adjacent equipment of the same type, re-inspect after the inspection is complete, and restore power after confirming that the equipment is operating normally.

[0042] When a "high-risk" warning is issued, operations personnel will schedule a planned power outage for maintenance within seven days. This includes power outages for electrical equipment, repairs to damaged areas, and increased real-time monitoring frequency after repairs.

[0043] When the warning does not reach the "urgent" or "high-risk" standards, it enters the "monitoring" state by default and starts continuous monitoring without the need for immediate maintenance.

[0044] Furthermore, in step S1 , the appearance image and the vibration displacement are both acquired by using a camera; and the temperature image is acquired by using an infrared thermal imager.

[0045] Using the above technical solution, appearance images, vibration displacement and temperature images are all acquired visually, achieving non-contact measurement and avoiding interference problems caused by installing various sensors on the surface of electrical equipment. At the same time, cameras and infrared thermal imagers can be mounted on brackets, drones or robots without stopping or modifying equipment, enabling rapid installation. Moreover, visual methods can simultaneously acquire appearance, vibration and temperature data, which are aligned in time and space, helping to analyze the multi-parameter coupling relationship of the electrical equipment status.

[0046] Furthermore, the step S1 includes the following steps:

[0047] S11: Classify the electrical equipment in the station into three categories: electromagnetic equipment, switchgear, and reactive equipment;

[0048] S12: Determine the placement of cameras and infrared thermal imagers based on the categories of electrical equipment:

[0049] If it is electromagnetic equipment, the camera for collecting vibration displacement is placed at the base of the electrical equipment and the root of the bushing; the camera for collecting appearance images is placed on the outside of the bushing; the infrared thermal imager for collecting temperature images is placed in the winding area of ​​the electrical equipment and at the bushing joint;

[0050] If it is a switchgear, the camera for collecting vibration displacement is placed on the box of the operating mechanism of the electrical equipment; the camera for collecting appearance images is placed at the opening and closing positions of the electrical equipment; the infrared thermal imager for collecting temperature images is placed on the contact surface of the contact;

[0051] If it is a reactive device, the camera for collecting vibration displacement is placed on the mounting bracket of the electrical equipment; the camera for collecting appearance images is placed on the expander of the electrical equipment; and the infrared thermal imager for collecting temperature images is placed on the surface of the porcelain sleeve of the electrical equipment.

[0052] By adopting the above technical solution, the layout positions of cameras and infrared thermal imagers are determined according to the different categories of electrical equipment, which can achieve targeted data collection and ensure the accuracy of data collection.

[0053] A system using any of the above-mentioned methods for identifying post-earthquake damage in a substation / converter station comprises:

[0054] Data acquisition module: acquires the appearance image, vibration displacement and temperature image of the electrical equipment in the station after the earthquake, and obtains the initial appearance image data, initial vibration displacement data and initial temperature image data;

[0055] Preprocessing module: preprocesses the initial appearance image data, initial vibration displacement data and initial temperature image data to obtain intermediate image data, intermediate vibration feature vector and intermediate thermal image data respectively;

[0056] Identification module: Inputs the intermediate image data, intermediate vibration feature vectors and intermediate thermal image data into the damage identification model for feature extraction and fusion, and outputs the damage identification results of the electrical equipment.

[0057] By adopting the above technical solution, the system can capture changes in the internal state of the equipment, such as changes in vibration characteristics and abnormal heating, by fusing vibration data and temperature images, thereby achieving more comprehensive damage detection; at the same time, it can maintain stable recognition performance under a wider range of environmental conditions, improving the robustness and environmental adaptability of the method. Especially in the special scenario of post-earthquake, it can achieve rapid post-earthquake repair work and provide reliable technical support for the post-earthquake emergency response of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is the logical structure diagram of the decision engine in this application;

[0059] Figure 2 This is a schematic diagram of the system structure of this application;

[0060] Figure 3 This is a flowchart of the application. DETAILED DESCRIPTION

[0061] The present invention is further described below with reference to specific embodiments and accompanying drawings.

[0062] A method for identifying damage after an earthquake in a substation or converter station, such as Figure 3 As shown: The following steps are included:

[0063] S1: Obtain post-earthquake appearance images, epicenter vibration displacement, and post-earthquake temperature images of electrical equipment in the station to obtain initial appearance image data, initial vibration displacement data, and initial temperature image data;

[0064] In step S1, the appearance image can intuitively reflect the external features of the electrical equipment after the earthquake; the vibration displacement can intuitively reflect the vibration displacement of the electrical equipment during the ground vibration process; the temperature image can intuitively reflect the temperature abnormality area caused by internal faults or damage of the electrical equipment; the initial appearance image data focuses on the external features of the electrical equipment, that is, the obvious features visible to the naked eye; the initial vibration displacement data and initial temperature image data focus on the internal features of the electrical equipment, that is, the features that cannot be directly observed by the naked eye;

[0065] Furthermore, for a single electrical device, since the areas where it is prone to damage at the epicenter and the areas where temperature abnormalities are prone to occur are well known in the art, the acquisition equipment for collecting appearance images, vibration displacement and temperature images is placed at the key position corresponding to the electrical equipment; the key position is the position where each acquisition device is prone to damage when collecting corresponding data.

[0066] S2: Preprocessing the initial appearance image data, initial vibration displacement data, and initial temperature image data to obtain intermediate image data, intermediate vibration feature vectors, and intermediate thermal image data;

[0067] Specifically, in step S2, the initial appearance image data is subjected to Gaussian noise injection (σ=0.01-0.05), random rotation (±15), horizontal flipping, color space transformation (RGB→HSV to adjust saturation / degree), and PCA noise (λ=0.1); CycleGAN is used to generate virtual images under different illumination / viewing angles, and Gaussian filtering (kernel size 3×3, σ=1.5) is used for denoising, histogram equalization is used to enhance contrast, gamma correction (γ=0.8 or 1.2) is used to adapt to different lighting conditions, and affine transformation (translation ±5px, scaling 0.9-1.1 times) is used to simulate shooting errors. The preprocessing of the initial appearance image data is completed, and the intermediate image data is output;

[0068] A 50Hz Butterworth low-pass filter is used to remove environmental noise from the initial vibration displacement data, and the displacement value is mapped to the interval [-1, 1]. The time domain features of the vibration characteristic information are extracted. The time domain is converted to the frequency domain using a fast Fourier transform (FFT), and the frequency domain features of the vibration characteristic information are extracted. The vibration characteristic information is preprocessed and the intermediate vibration characteristic vector is output. The intermediate vibration characteristic vector includes the root mean square value, crest factor, kurtosis, fundamental frequency (50 / 60Hz), harmonic amplitude (2f, 3f), 1 / 3 octave spectrum energy (focusing on the 100-500Hz mechanical fault frequency band), and the preset fault characteristic frequency.

[0069] The initial temperature image data is corrected using a two-point correction method (Blackbody calibration, with the correction coefficient stored in the system) to eliminate detector noise and converted into actual temperature values ​​based on the surface emissivity of the electrical equipment (0.05-0.3 for metals and 0.8-0.95 for ceramics). Based on the region growing algorithm (seed point: temperature ≥ mean + 2σ region), the abnormal region is segmented in combination with Canny edge detection, and the maximum temperature T in the ROI is extracted. max , average temperature T avg , temperature standard deviation σ T , maximum temperature gradient ▽T max and the abnormal area A roi , output the intermediate thermal image data.

[0070] S3: Input the intermediate image data, intermediate vibration feature vectors, and intermediate thermal image data into the damage identification model for feature extraction and fusion, outputting a damage identification result for the electrical equipment. The damage identification result may include determining whether the electrical equipment is damaged, or determining the location of the damage based on the presence of damage. Therefore, the applicant intends to state that the damage identification result for the electrical equipment in this step can be determined based on actual site conditions or manually defined, and only requires training the damage identification model accordingly.

[0071] Specifically, the damage identification model extracts separate features from the intermediate image data, intermediate vibration feature vectors, and intermediate thermal image data, and then fuses the features to ultimately obtain damage identification results for electrical equipment. Separate feature extraction is performed on different types of data without interfering with each other, and there is no conflict between features, thus ensuring the accuracy of feature extraction. The damage identification results for electrical equipment integrate both the external and internal features of the electrical equipment, significantly improving the accuracy of the identification results.

[0072] Therefore, in this embodiment, direct information about the surface of electrical equipment is provided through appearance images, the mechanical state and dynamic characteristics of electrical equipment are provided through vibration displacement, and the thermal state and potential thermal problems of electrical equipment are provided through temperature images. Data is collected simultaneously from the outside and inside of the electrical equipment for subsequent analysis, avoiding the incomplete detection caused by only collecting external data of electrical equipment and ignoring changes in internal state. At the same time, by fusing multimodal data through the damage identification model, the location and scope of the damage can be more accurately determined. This precise damage location capability provides more targeted guidance for post-earthquake maintenance work, improving maintenance efficiency and quality. Therefore, the entire substation / converter station post-earthquake damage identification method can achieve more comprehensive detection of electrical equipment and have more accurate damage location capabilities, thereby achieving rapid post-earthquake repair work, ensuring the rapid and stable operation of power equipment, and providing reliable technical support for the post-earthquake emergency response of the power system.

[0073] In a specific embodiment, the damage identification model in step S3 includes a first channel, a second channel, a third channel, a fusion layer and a fully connected layer;

[0074] The input of the first channel is the intermediate image data, and the output is the image feature subnet F img ;

[0075] The input of the second channel is the intermediate vibration feature vector, and the output is the vibration feature subnet F vir ;

[0076] The input of the third channel is the intermediate thermal image data, and the output is the thermal image feature subnet F ir ;

[0077] The input of the fusion layer is the image feature subnet F img , vibration feature subnet F vir and thermal image feature subnet F ir , the output is the modal weights of the three channels and the weighted fusion unit F fused ; Weighted fusion unit F fused for:

[0078] F fused =w img ·F img +wvib ·F vib +w ir ·F ir

[0079] Among them, w img is the fusion weight of the first channel; w vib is the fusion weight of the second channel; w ir is the fusion weight of the third channel;

[0080] The input of the fully connected layer is the weighted fusion unit F fused , the output is the damage identification result of the electrical equipment; the damage identification result of the electrical equipment includes the damage status of the electrical equipment and the first confidence of the damage identification model for the identification result.

[0081] Specifically, in this embodiment, the main framework of the damage identification model is specifically explained, which mainly includes three channels, a fusion layer and a fully connected layer. By utilizing the three channels to independently process different data inputs, mutual contamination of noise between different modalities is prevented, and the accuracy of feature extraction of each channel is guaranteed; at the same time, the three channels can process corresponding tasks in parallel, greatly improving computing efficiency; and each channel can independently optimize the network structure without the need to reconstruct the entire model. The channel structure can be adjusted according to different equipment types, making the entire damage identification model highly maintainable and scalable.

[0082] What is more worth mentioning is that the fusion layer splices the high-level features extracted from the first channel, the second channel, and the third channel, and maps the input features into three scalar weights, which reflect the relative importance of each modality (image, vibration, thermal image) to the current damage identification; the attention mechanism is used for adaptive weighted fusion to improve the model's adaptability and noise suppression capabilities, thereby improving the accuracy of damage identification; at the same time, the output result introduces the first confidence level, so that manual review can be introduced for low-confidence results to reduce the risk of misjudgment; it can achieve self-reflection and continuous optimization of the damage identification model, thereby improving the recognition accuracy of the damage identification model; the entire damage identification model has high comprehensiveness, robustness and accuracy.

[0083] In a specific embodiment, the structures of the three channels are specifically exemplified:

[0084] The first channel adopts a convolutional neural network, which includes five convolutional layers and two spatial attention modules. The five convolutional layers are 3×3 kernels with a step size of 2. The two spatial attention modules are embedded in the third and fifth convolutional layers respectively, and a 256-dimensional image feature subnet F is generated by generating a spatial weight mask. img ;

[0085] The second channel uses 1D-CNN to process the intermediate vibration feature vector, and the 1D-CNN includes three convolutional layers with kernel size 5 and step size 1; generating a 128-dimensional vibration feature subnet F vir ;

[0086] The third channel uses two fully connected layers (256→128) to process the intermediate thermal image data and generate a 128-dimensional thermal image feature subnet F ir ;

[0087] The fusion layer inputs the image feature subnet F imgg , vibration feature subnet F vir , thermal image feature subnet F ir , through two layers of FC (256→64→3), the output modal weight w img 、w vib 、w ir , and calculate the weighted fusion unit F fused And output.

[0088] This embodiment selects different network structures to implement feature extraction according to the characteristics of different data, thereby ensuring the accuracy of feature extraction to the greatest extent while taking into account cost and computational efficiency.

[0089] In a specific embodiment, when the damage identification model is trained, the weight parameters of the first channel, the weight parameters of the second channel, the weight parameters of the third channel and the parameters of the fusion layer in the damage identification model are optimized using a particle swarm optimization algorithm.

[0090] Specifically, in this embodiment, the weight parameters of each channel in the model (including convolution kernel weights and fully connected layer weights) and the fusion layer parameters are jointly optimized. First, a range of values ​​is preset to reduce computational cost and improve computational efficiency. The convolution kernel weights (range [-1, 1]), fully connected layer weights (range [-0.5, 0.5]), and fusion weight parameters (initial value 0.33, optimization range [0, 1]) are preset for real-time optimization. All trainable parameters of the network (convolution kernel weights, fully connected layer weights, attention unit parameters) are mapped to particle position vectors. The number of particles N = 50, and the dimension of each particle position vector equals the total number of trainable parameters (e.g., CNN+1D-CNN+FC has approximately 100,000 parameters, which need to be reduced to a range that can be handled by PSO, such as selecting the first 1000 principal components through principal component analysis).

[0091] Each particle in the particle swarm is updated according to the improved speed-position update strategy. The speed update formula is: The position update formula is: in, is the velocity vector of the w-th dimension of the i-th particle in the t-th iteration, w∈[1,W], is the velocity vector of the w-th dimension of the i-th particle in the t-1-th iteration.

[0092] After initializing the particle swarm, it iteratively updates the position and terminates when |Fitnesst-Fitnesst-1|≤0.01 or the maximum number of iterations Tmax=200 is reached, and the optimal network parameters are output. The fitness function of PSO is defined as:

[0093] Where K is a set of high-risk damage categories, α=0.7, β=0.3; the high-risk damage categories include but are not limited to insulator breakage, winding deformation, transformer oil leakage, etc.;

[0094] Optionally, after outputting the optimal parameters, the entire network is fine-tuned through transfer learning to avoid falling into a local optimum. Preferably, when the warning level is misjudged three times in a row (e.g., there is actually no damage but the output is "high risk") or the system recognition accuracy drops by ≥5%, PSO optimization is automatically started. Preferably, new samples are collected in real time (the training set is updated every quarter, including the post-earthquake data of the current season) and manually annotated false positive cases (supplemented by feedback from operation and maintenance personnel) are used as the training set for the PSO algorithm. Preferably, incremental learning is used to update only the local network parameters related to the current sample (e.g., the channel weights of a specific equipment category) to avoid affecting the stability of the overall model. Preferably, the fusion layer weights are visualized (e.g., a heat map showing the contribution of each mode) to assist operation and maintenance personnel in understanding the damage judgment criteria and improving the credibility of decision-making.

[0095] In a specific embodiment, the step S4 is also included: the damage identification result of the electrical equipment and the real-time physical feature vector are combined. Input into the decision engine to obtain the damage type of the electrical equipment and the second confidence level of the decision engine for the damage type of the electrical equipment; is the temperature gradient, which is directly detected by the third channel of the damage identification model; ΔE is the weekly growth rate of vibration energy, which is calculated according to the following formula:

[0096]

[0097] Among them, E current is the current vibration energy value, E last week is the vibration energy value of the previous week. The formula for the vibration energy value E is as follows:

[0098]

[0099] where X iis the value of the i-th vibration displacement sampling point (unit: pixel or physical displacement unit, such as mm); N is the number of sampling points (related to the signal duration and sampling frequency, such as when the sampling frequency is 100 Hz and the duration is 10 seconds, N = 1000);

[0100] Specifically, in this embodiment, the calculation logic of the initial damage type D and confidence C is based on the training of the historical fault data set (including 2000+ post-earthquake equipment samples), combined with the finite element simulation data (such as Abaqus simulation of transformer winding deformation and vibration characteristics) to calibrate the feature-damage mapping relationship.

[0101] In a specific embodiment, the setting of the decision engine can be based on the damage status of the equipment and the first confidence level; combined with the key physical signal characteristics in the intermediate vibration feature vector and / or intermediate thermal imaging data, the identified damage status, the second confidence level, and the equipment structure knowledge base and equipment damage law information base pre-stored in the system cloud, a comprehensive analysis is performed through the decision engine; the decision engine integrates the decision tree algorithm with the physics-based damage evolution rule reasoning to determine the damage mode of the equipment, evaluate the damage risk level, predict the development trend, and generate specific maintenance suggestions, warning levels and emergency response plans.

[0102] like Figure 1 As shown, specifically, in step S4, the decision engine further includes a second confidence level of the decision engine for the damage type of the electrical equipment and the real-time physical feature vector. Generate early warning strategy:

[0103] Output "urgent" warning;

[0104] Output "high risk" warning;

[0105] ELSE outputs monitoring instructions.

[0106] Where d1 and d2 are elements in the equipment damage type D set, e1 and e2 are vibration energy limits, t1 and t2 are temperature gradient limits, and c1 and c2 are initial confidence thresholds. It should be noted that the initial decision logic and threshold values ​​can be determined by referring to industry standards (such as DL / T 664-2016 Infrared Detection Procedure), equipment manufacturer manuals (such as ABB circuit breaker vibration thresholds), historical fault statistics (such as post-earthquake data from a 500kV substation), or finite element simulation results.

[0107] Of course, an early warning strategy can also be generated in another way: define an early warning value function G, the independent variables of which are damage type D, confidence C and real-time physical characteristic vector The warning value g1 can be calculated by function G, and the initial warning values ​​g1 and g2 can be set. If the warning value g is higher than g1, an "emergency" warning will be output. If the warning value g is higher than g2, a "high-risk" warning will be output. If it is not greater than the g1 or g2 threshold, a monitoring instruction will be output. This method is simple and feasible.

[0108] In a specific embodiment, after the early warning strategy is generated, different maintenance strategies are formulated according to the level of the early warning strategy:

[0109] When an "emergency" warning is issued, operations and maintenance personnel must complete a power outage and maintenance within 72 hours. They must immediately isolate the damaged electrical equipment, disassemble the equipment, and replace damaged components (such as broken insulators and deformed windings). They must also conduct a comprehensive inspection of adjacent equipment of the same type to identify maintenance orders with potential risks. After the maintenance is complete, multimodal data, such as vibration and infrared thermal imaging, must be re-inspected to confirm normal operation before restoring power.

[0110] When a "high-risk" warning is issued, operations and maintenance personnel will schedule a planned power outage for maintenance within seven days. During maintenance, electrical equipment will be prioritized for power outages and repairs will be performed on damaged parts. After maintenance, the frequency of real-time monitoring will be increased.

[0111] When the warning does not meet the "emergency" or "high-risk" standards (such as low-confidence anomalies or slight fluctuations in parameters), it enters the "monitoring" state by default and starts continuous monitoring without the need for immediate maintenance.

[0112] In one specific embodiment, the appearance image and vibration displacement in step S1 are both acquired using a camera; the temperature image is acquired using an infrared thermal imager. Of course, vibration displacement data can also be acquired using a vibration sensor, the specific method of acquisition being determined based on the circumstances. This application utilizes a camera to acquire vibration displacement data because both utilize a single visual system, providing a holistic and relatively simple operation.

[0113] It should be noted that the vibration displacement is acquired at the epicenter using a camera. Although the camera is located in the earthquake, the coupling between the camera and other equipment is not complicated, so the stress under the earthquake is not large. Therefore, the probability of direct damage and functional failure is very small. Therefore, the camera is used for acquisition. Of course, the displacement can also be measured visually, such as target visual monitoring, monocular ranging technology and other conventional technologies to acquire the vibration displacement data at the epicenter.

[0114] Specifically, the camera uses a high-definition camera with a frame rate of 30fps or higher, continuously capturing vibration video (duration ≥60 seconds) at the epicenter; external image acquisition (resolution ≥4K) is completed within 2 hours after the earthquake; the infrared thermal imager has an accuracy of ±2% or ±2°C, and scans the entire equipment within 30 minutes after the earthquake, targeting areas with abnormal temperatures (ROI marking is triggered when the temperature difference is ≥5°C); and post-earthquake external image acquisition of the equipment includes close-ups of key parts of the equipment (such as insulators, bushings, and contacts), with 3-5 images of each part taken from different angles.

[0115] Equipment epicenter vibration data collection includes the camera using the optical flow method to calculate the vibration displacement of key parts (pixel-level accuracy), with a sampling frequency of 100Hz, to obtain the XYZ three-axis displacement sequence;

[0116] The equipment's post-earthquake thermal imaging data collection conditions require an ambient temperature of 25±5°C, relative humidity ≤70%, and avoid direct sunlight; the thermal imager must be at a fixed distance from the equipment to ensure a pixel resolution ≥0.1mrad.

[0117] Optionally, the high-definition camera and infrared thermal imager are synchronized via GPS (accuracy ≤ 1ms) to ensure the consistency of timestamps of vibration, image, and thermal imaging data, facilitating spatiotemporal alignment analysis.

[0118] In this embodiment, appearance images, vibration displacement, and temperature images are all acquired visually, enabling non-contact measurement and avoiding interference problems caused by installing various sensors on the surface of electrical equipment. At the same time, cameras and infrared thermal imagers can be mounted on brackets, drones, or robots, without the need to shut down or modify equipment, enabling rapid installation. Furthermore, visual methods can simultaneously acquire appearance, vibration, and temperature data, which are aligned in time and space, facilitating analysis of the multi-parameter coupling relationship of the electrical equipment status.

[0119] In a specific embodiment, step S1 includes the following steps:

[0120] S11: Classify the electrical equipment in the station into three categories: electromagnetic equipment, switchgear, and reactive equipment;

[0121] S12: Determine the placement of cameras and infrared thermal imagers based on the categories of electrical equipment:

[0122] If it is electromagnetic equipment, the camera for collecting vibration displacement is placed at the base of the electrical equipment and the root of the bushing; the camera for collecting appearance images is placed on the outside of the bushing; the infrared thermal imager for collecting temperature images is placed in the winding area of ​​the electrical equipment and at the bushing joint;

[0123] If it is a switchgear, the camera for collecting vibration displacement is placed on the box of the operating mechanism of the electrical equipment; the camera for collecting appearance images is placed at the opening and closing positions of the electrical equipment; the infrared thermal imager for collecting temperature images is placed on the contact surface of the contact;

[0124] If it is a reactive device, the camera for collecting vibration displacement is placed on the mounting bracket of the electrical equipment; the camera for collecting appearance images is placed on the expander of the electrical equipment; and the infrared thermal imager for collecting temperature images is placed on the surface of the porcelain sleeve of the electrical equipment.

[0125] By adopting the above technical solution, the layout positions of cameras and infrared thermal imagers are determined according to the different categories of electrical equipment, which can achieve targeted data collection and ensure the accuracy of data collection.

[0126] In one embodiment, Figure 2 As shown, a system using the above-mentioned method for identifying post-earthquake damage in a substation / converter station includes:

[0127] Data acquisition module: acquires the appearance image, vibration displacement and temperature image of the electrical equipment in the station after the earthquake, and obtains the initial appearance image data, initial vibration displacement data and initial temperature image data;

[0128] Preprocessing module: Preprocesses the initial appearance image data, initial vibration displacement data, and initial temperature image data to obtain intermediate image data, intermediate vibration feature vectors, and intermediate thermal image data. It is worth noting that the preprocessing module is deployed on an edge server (such as NVIDIA Jetson), where data cleaning and feature extraction are performed on-site. Only intermediate features are uploaded to the cloud, reducing communication bandwidth pressure.

[0129] Identification module: Inputs the intermediate image data, intermediate vibration feature vectors and intermediate thermal image data into the damage identification model for feature extraction and fusion, and outputs the damage identification results of the electrical equipment.

[0130] The system in this embodiment can capture changes in the internal state of the equipment, such as changes in vibration characteristics and abnormal heating, by fusing vibration data and temperature images, thereby achieving more comprehensive damage detection; at the same time, it can maintain stable recognition performance under a wider range of environmental conditions, improving the robustness and environmental adaptability of the method; the entire system realizes closed-loop management from data acquisition equipment deployment to damage decision-making, significantly improving the accuracy and efficiency of post-earthquake equipment damage identification, providing core technical support for earthquake disaster relief in the power system, enabling rapid post-earthquake repair work, and providing reliable technical guarantee for the post-earthquake emergency response of the power system.

[0131] The examples described in the present invention are merely descriptions of the preferred embodiments of the present invention and are not intended to limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various modifications and improvements made to the technical solutions of the present invention by engineers and technicians in this field should fall within the scope of protection of the present invention.

Claims

1. A method for identifying post-earthquake damage in a substation / converter station, characterized by: The following steps are involved: S1: Obtain post-earthquake appearance images, epicenter vibration displacement, and post-earthquake temperature images of electrical equipment in the station to obtain initial appearance image data, initial vibration displacement data, and initial temperature image data; S2: Preprocessing the initial appearance image data, initial vibration displacement data, and initial temperature image data to obtain intermediate image data, intermediate vibration feature vectors, and intermediate thermal image data; S3: Input the intermediate image data, intermediate vibration feature vector and intermediate thermal image data into the damage identification model for feature extraction and fusion, and output the damage identification results of the electrical equipment.

2. The method for identifying post-earthquake damage in a substation / converter station according to claim 1, characterized in that: The damage recognition model in step S3 includes a first channel, a second channel, a third channel, a fusion layer and a fully connected layer; The input of the first channel is the intermediate image data, and the output is the image feature subnet F img ; The input of the second channel is the intermediate vibration feature vector, and the output is the vibration feature subnet F vir ; The input of the third channel is the intermediate thermal image data, and the output is the thermal image feature subnet F ir ; The input of the fusion layer is the image feature subnet F img , vibration feature subnet F vir and thermal image feature subnet F ir , the output is the modal weights of the three channels and the weighted fusion unit F fused ; Weighted fusion unit F fused for: F fused =w img ·F img +w vib ·F vib +w ir ·F ir Among them, w img is the fusion weight of the first channel; w vib is the fusion weight of the second channel; w ir is the fusion weight of the third channel; The input of the fully connected layer is the weighted fusion unit F fused , the output is the damage identification result of the electrical equipment; the damage identification result of the electrical equipment includes the damage status of the electrical equipment and the first confidence of the damage identification model for the identification result.

3. The method for identifying post-earthquake damage in a substation / converter station according to claim 2, characterized in that: The first channel adopts a convolutional neural network, which includes five convolutional layers and two spatial attention modules. The two spatial attention modules are embedded in the third and fifth convolutional layers respectively, and the image feature subnet F is generated by generating a spatial weight mask. img ; The second channel uses 1D-CNN to process the intermediate vibration feature vector to generate the vibration feature subnet F vir ; The third channel uses a fully connected layer to process the intermediate thermal image data to generate a thermal image feature subnet F ir .

4. A method for identifying post-earthquake damage in a substation / converter station according to claim 2 or 3, characterized in that: When the damage identification model is trained, the weight parameters of the first channel, the weight parameters of the second channel, the weight parameters of the third channel and the parameters of the fusion layer in the damage identification model are optimized using a particle swarm optimization algorithm.

5. The method for identifying post-earthquake damage in a substation / converter station according to claim 1, characterized in that: The step S4 is also included: combining the damage identification result of the electrical equipment and the real-time physical feature vector Input into the decision engine to obtain the damage type of the electrical equipment and the second confidence level of the decision engine for the damage type of the electrical equipment; is the temperature gradient, which is directly detected by the damage identification model; ΔE is the weekly growth rate of vibration energy, which is calculated according to the following formula: Among them, E current is the current vibration energy value, E lastweek It is the vibration energy value of the previous week.

6. The method for identifying post-earthquake damage in a substation / converter station according to claim 5, characterized in that: In step S4, the decision engine further includes determining the damage type of the electrical equipment, the second confidence level of the decision engine for the damage type of the electrical equipment, and the real-time physical feature vector. Generate early warning strategy: Output "urgent" warning; Output "high risk" warning; ELSE outputs monitoring instructions. Where d1 and d2 are elements in the equipment damage type D set, e1 and e2 are vibration energy limits, t1 and t2 are temperature gradient limits; c1 and c2 are initial confidence thresholds.

7. The method for identifying post-earthquake damage in a substation / converter station according to claim 6, characterized in that: After the early warning strategy is generated, different maintenance strategies are formulated according to the level of the early warning strategy: When an "emergency" warning is issued, operations and maintenance personnel must complete a power outage and maintenance within 72 hours. They must immediately isolate the damaged electrical equipment, dismantle the equipment, and replace damaged components. They must also conduct a comprehensive inspection of adjacent equipment of the same type. After the inspection is complete, they must re-inspect and restore power to normal operation. When a "high-risk" warning is issued, operations personnel will schedule a planned power outage for maintenance within seven days. This includes power outages for electrical equipment, repairs to damaged areas, and increased real-time monitoring frequency after repairs. When the warning does not reach the "urgent" or "high-risk" standards, it enters the "monitoring" state by default and starts continuous monitoring without the need for immediate maintenance.

8. The method for identifying post-earthquake damage in a substation / converter station according to claim 1, characterized in that: In step S1 , the appearance image and the vibration displacement are both acquired by a camera; and the temperature image is acquired by an infrared thermal imager.

9. The method for identifying post-earthquake damage in a substation / converter station according to claim 8, characterized in that: The step S1 comprises the following steps: S11: Classify the electrical equipment in the station into three categories: electromagnetic equipment, switchgear, and reactive equipment; S12: Determine the placement of cameras and infrared thermal imagers based on the categories of electrical equipment: If it is electromagnetic equipment, the camera for collecting vibration displacement is placed at the base of the electrical equipment and the root of the casing; The camera for collecting appearance images is arranged on the outer surface of the casing; Infrared thermal imagers for collecting temperature images are placed in the winding area and bushing joints of electrical equipment; If it is a switchgear, the camera for collecting vibration displacement is placed on the box of the operating mechanism of the electrical equipment; the camera for collecting appearance images is placed at the opening and closing positions of the electrical equipment; the infrared thermal imager for collecting temperature images is placed on the contact surface of the contact; If it is a reactive device, the camera for collecting vibration displacement is placed on the mounting bracket of the electrical equipment; the camera for collecting appearance images is placed on the expander of the electrical equipment; and the infrared thermal imager for collecting temperature images is placed on the surface of the porcelain sleeve of the electrical equipment.

10. A system using the method for identifying post-earthquake damage in a substation / converter station according to any one of claims 1 to 9, characterized in that: include: Data acquisition module: acquires the appearance image, vibration displacement and temperature image of the electrical equipment in the station after the earthquake, and obtains the initial appearance image data, initial vibration displacement data and initial temperature image data; Preprocessing module: preprocesses the initial appearance image data, initial vibration displacement data and initial temperature image data to obtain intermediate image data, intermediate vibration feature vector and intermediate thermal image data respectively; Identification module: Inputs the intermediate image data, intermediate vibration feature vectors and intermediate thermal image data into the damage identification model for feature extraction and fusion, and outputs the damage identification results of the electrical equipment.

Images