An intelligent diagnosis method for power transformation primary equipment
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LANGFANG POWER SUPPLY COMPANY STATE GRID JIBEI ELECTRIC POWER COMPANY
- Filing Date
- 2026-01-15
- Publication Date
- 2026-06-05
AI Technical Summary
In existing technologies, overheat detection of substation equipment relies on manual infrared spectrum analysis, which has the problems of consuming a lot of manpower, being easily affected by experience, having low diagnostic efficiency, and being difficult to manage large amounts of data.
A deep learning-based intelligent diagnostic model for substation equipment is adopted, which utilizes infrared thermal imaging spectra, environmental parameters, and equipment operating condition information to perform automated fault identification and diagnosis through a neural network model.
It enables intelligent diagnosis of power equipment, improves the accuracy and efficiency of fault identification, reduces manual intervention, and supports real-time monitoring and batch diagnosis of large-scale equipment.
Smart Images
Figure CN122159489A_ABST
Abstract
Description
Technical Field
[0001] This disclosure generally relates to the field of intelligent diagnostic technology for power equipment, and specifically to an intelligent diagnostic method for primary power equipment. Background Technology
[0002] With the continuous expansion of the power grid and the sustained increase in load levels, the operating pressure on substation power equipment has increased significantly, and equipment overheating has become one of the main hidden dangers threatening the safe operation of substations. Infrared thermal imaging, as a key means of live-line detection, can promptly detect abnormal heating points through visual monitoring of the equipment temperature field. It is one of the most widely used technologies in current maintenance and repair work, with the highest defect detection rate.
[0003] Under current technological conditions, maintenance personnel need to periodically carry portable infrared thermal imagers to substations to take thermal images of primary equipment, and then manually classify, calibrate, and calculate temperature differences from the captured infrared spectra. This temperature measurement process is not only labor-intensive but also susceptible to the influence of the operator's experience level, leading to the risk of missed or incorrect diagnoses. Furthermore, the sheer volume of infrared spectra obtained, coupled with a lack of centralized management technology, makes it difficult to support the increasingly heavy maintenance tasks of primary equipment, resulting in low efficiency in the maintenance of substation power equipment. Summary of the Invention
[0004] In view of the above-mentioned defects or deficiencies in the prior art, it is desirable to provide an intelligent diagnostic method for primary equipment of substations.
[0005] This invention provides an intelligent diagnostic method for primary equipment in a substation, comprising: Acquire the status information of the primary equipment of the substation, the status information including at least the infrared thermal imaging spectrum of the primary equipment of the substation, and the acquisition environment parameters and equipment operating condition information corresponding to the infrared thermal imaging spectrum; The state information is input into a pre-trained intelligent diagnostic model for substation equipment; wherein, the intelligent diagnostic model for substation equipment is a neural network model built based on a deep learning framework and trained using the historical state information of the primary substation equipment and corresponding fault condition labels; Based on the output of the intelligent diagnostic model for the substation equipment, determine whether the primary substation equipment has a fault and the corresponding fault type; If a fault occurs in the primary equipment of the substation, an alarm signal containing at least the fault type and equipment location information is generated and sent to the monitoring terminal.
[0006] According to the technical solution provided by the present invention, the pre-training process of the intelligent diagnostic model for substation equipment includes the following steps: Obtain the training dataset and validation dataset for training the intelligent diagnostic model of the substation equipment; The intelligent diagnostic model for the substation equipment is trained based on the training dataset. The training effect of the intelligent diagnostic model for substation equipment is evaluated using the validation dataset to obtain evaluation indicators. When the evaluation indicators meet the preset training requirements, the training of the intelligent diagnostic model for substation equipment is confirmed to be complete.
[0007] According to the technical solution provided by the present invention, obtaining the training dataset and validation dataset for training the intelligent diagnostic model of the substation equipment includes: Initial infrared training image sets corresponding to different types of primary substation equipment are collected, and each initial infrared training image set is filtered to obtain an infrared training image set corresponding to each type of primary substation equipment. All infrared training maps in each infrared training map set are filtered and preprocessed, and image processing is performed to obtain a binarized edge map corresponding to each infrared training map, so as to construct the final training map set. The training dataset and the validation dataset are constructed based on the training map set and the labels corresponding to each binarized edge map in the training map set.
[0008] According to the technical solution provided by the present invention, the step of filtering each of the initial infrared training spectrum sets to obtain an infrared training spectrum set corresponding to each type of primary substation equipment includes: The environmental acquisition conditions for each infrared training image in the initial infrared training image set are obtained; wherein, the environmental acquisition conditions include at least: meteorological conditions, light conditions, and electromagnetic radiation interference conditions; The meteorological conditions, light conditions, and electromagnetic radiation interference conditions are compared with preset standard environmental acquisition conditions. Infrared training spectra that meet the preset standard environmental acquisition conditions are retained, while infrared training spectra that do not meet the standard environmental acquisition conditions are deleted, so as to obtain the infrared training spectra set corresponding to each type of substation primary equipment.
[0009] According to the technical solution provided by the present invention, the step of preprocessing and edge processing all infrared training maps in each of the infrared training map sets to obtain a binarized edge map corresponding to each of the infrared training maps, so as to construct the final training map set, includes: A smoothing filter function is used to filter all infrared training images in the infrared training image set; Based on the filtered infrared training spectrum image, its horizontal and vertical gradients are calculated to obtain the gradient magnitude and gradient direction of each pixel in the corresponding image. For each pixel, its gradient magnitude is compared with the gradient magnitude of its neighboring pixels along its gradient direction, and each pixel is filtered according to a preset first magnitude threshold and a second magnitude threshold to obtain the binarized edge map corresponding to each image. Among them, all the binarized edge maps corresponding to each type of primary substation equipment constitute its final training graph set.
[0010] According to the technical solution provided by the present invention, the step of comparing the gradient magnitude of each pixel along its gradient direction with the gradient magnitude of adjacent pixels, and filtering each pixel according to a preset first magnitude threshold and a second magnitude threshold to obtain a binarized edge map corresponding to each image, includes: Obtain the gradient magnitude of each pixel along its gradient direction and compare it with that of its neighboring pixels. Based on the magnitude comparison results, retain the local maxima points in the image. Set the first amplitude threshold and the second amplitude threshold, wherein the first amplitude threshold is greater than the second amplitude threshold; Obtain the gradient magnitudes corresponding to all retained local maxima points and use them as the target gradient magnitude set; Pixels with a target gradient magnitude higher than a first magnitude threshold are marked as strong edge pixels; pixels with a target gradient magnitude lower than a second magnitude threshold are removed; and pixels with a target gradient magnitude between the first magnitude threshold and the second magnitude threshold are marked as weak edge pixels. For the weak edge pixel, if it is adjacent to or connected to any of the strong edge pixels, it is retained as an edge pixel; if it is not adjacent to or connected to any of the strong edge pixels, it is discarded. All retained edge pixels are assigned a value of 1, and the remaining pixels are assigned a value of 0, to obtain the binarized edge map corresponding to the filtered infrared training spectrum image.
[0011] According to the technical solution provided by the present invention, the step of constructing the training dataset and the validation dataset based on the training graph set and the labels corresponding to each binarized edge map in the training graph set includes: Acquire training map sets of different types of substation primary equipment and the labels corresponding to each binarized edge map in the training map set, wherein the labels include at least: original color bar information and its corresponding acquisition time, geographical coordinates, environmental parameters, operating condition information and fault condition labels; The binarized edge maps and their labels are divided into training datasets and validation datasets according to a preset ratio to construct the training dataset and validation dataset corresponding to each type of substation primary equipment.
[0012] According to the technical solution provided by the present invention, the training parameters of the intelligent diagnostic model for substation equipment include: 2 encoder layers, 4 multi-head attention heads, a dropout rate of 0.2, a batch size of 128, and a learning rate of 0.0001. The intelligent diagnostic model for substation equipment is based on a Transformer structure that includes a multi-head self-attention mechanism to extract correlation information between input features in parallel, and outputs the probability distribution of fault types of the primary substation equipment through a fully connected layer.
[0013] In summary, this technical solution specifically discloses an intelligent diagnostic method for primary substation equipment, comprising: acquiring the status information of the primary substation equipment, the status information including at least an infrared thermal imaging spectrum of the primary substation equipment, and acquisition environment parameters and equipment operating condition information corresponding to the infrared thermal imaging spectrum; inputting the status information into a pre-trained intelligent diagnostic model for the primary substation equipment; wherein, the intelligent diagnostic model for the primary substation equipment is a neural network model built based on a deep learning framework and trained using historical status information of the primary substation equipment and corresponding fault condition labels; determining whether the primary substation equipment has a fault and the corresponding fault type based on the output of the intelligent diagnostic model for the primary substation equipment; if the primary substation equipment has a fault, generating an alarm signal that includes at least the fault type and equipment location information, and sending the alarm signal to a monitoring terminal.
[0014] Beneficial Effects: This invention fundamentally transforms the traditional manual operation and maintenance model by deeply integrating infrared thermal imaging spectra with equipment operating conditions and environmental parameters, and automatically analyzing the pre-processed spectra using a Transformer-based deep learning model. This solution not only effectively isolates environmental and load interference and accurately identifies various defects in primary substation equipment, improving diagnostic accuracy and reliability, but also achieves full automation and intelligence from spectra analysis to fault early warning. It enables real-time monitoring and batch diagnosis of massive amounts of equipment, providing efficient and reliable technical support for improving the safe and stable operation of the power grid. Attached Figure Description
[0015] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.
[0016] Figure 1 This is a flowchart illustrating an intelligent diagnostic method for primary equipment in a substation.
[0017] Figure 2 This diagram illustrates the process of obtaining the training and validation datasets.
[0018] Figure 3 This is a flowchart illustrating the unfolding process of step A1.
[0019] Figure 4 This is a flowchart illustrating the process of step A2.
[0020] Figure 5 This is a flowchart illustrating the process of step A23.
[0021] Figure 6 This is a flowchart illustrating the process of step A3.
[0022] Figure 7 This is a schematic diagram of the core architecture of the multi-head self-attention mechanism.
[0023] Figure 8 This is a schematic diagram of the infrared spectrum of a primary substation device. Detailed Implementation
[0024] 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 merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.
[0025] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0026] Example 1 To make the technical solutions of the embodiments of the present invention clearer and easier to understand, the application background of the embodiments of the present invention will be introduced below.
[0027] With the continuous expansion of the power grid and the sustained increase in load levels, the operating pressure on substation power equipment has increased significantly, and equipment overheating has become one of the main hidden dangers threatening the safe operation of substations.
[0028] Infrared thermal imaging, a key method for live-line testing, can promptly detect abnormal heat points through visual monitoring of the equipment's temperature field. It is one of the most widely used and highest-detection-rate technologies in current maintenance and repair work. Under current technological conditions, maintenance personnel need to periodically carry portable infrared thermal imagers to substations to take thermal images of primary equipment, and then manually classify, calibrate, and calculate temperature differences from the captured infrared spectra. This temperature measurement process is not only labor-intensive but also susceptible to the influence of the operator's experience level, leading to the risk of missed or incorrect diagnostic results.
[0029] Especially for voltage-heated equipment with even lower temperature difference requirements and more difficult defect identification, manual judgment is labor-intensive and its accuracy is hard to guarantee. Furthermore, the sheer volume of infrared spectra measured lacks centralized management technology, making it difficult to support the increasingly heavy primary equipment maintenance tasks. During peak summer load periods, maintenance personnel need to frequently travel between substations to monitor equipment overheating, resulting in low work efficiency.
[0030] In view of this, the present invention proposes an intelligent diagnostic method for primary substation equipment, comprising: acquiring the status information of the primary substation equipment, the status information including at least an infrared thermal imaging spectrum of the primary substation equipment, and acquisition environment parameters and equipment operating condition information corresponding to the infrared thermal imaging spectrum; inputting the status information into a pre-trained intelligent diagnostic model for the primary substation equipment; wherein, the intelligent diagnostic model for the primary substation equipment is a neural network model built based on a deep learning framework and trained using historical status information of the primary substation equipment and corresponding fault condition labels; determining whether the primary substation equipment has a fault and the corresponding fault type based on the output of the intelligent diagnostic model for the primary substation equipment; if the primary substation equipment has a fault, generating an alarm signal that includes at least the fault type and equipment location information, and sending the alarm signal to a monitoring terminal.
[0031] Please refer to Figure 1 The flowchart shown in this embodiment illustrates an intelligent diagnostic method for primary substation equipment, including: S100. Obtain the status information of the primary equipment of the substation. The status information includes at least the infrared thermal imaging spectrum of the primary equipment of the substation, as well as the acquisition environment parameters and equipment operating condition information corresponding to the infrared thermal imaging spectrum. The types of primary equipment in a substation can include circuit breakers, disconnect switches, current transformers, voltage transformers, etc., without any special restrictions.
[0032] The infrared thermal imaging spectrum in the status information can be acquired using fixed or patrol-type infrared thermal imagers. This data reflects the temperature distribution field on the surface of the corresponding equipment. Environmental parameters can be acquired using various sensors, such as temperature sensors, humidity sensors, and wind speed sensors; no specific limitations are imposed here. Equipment operating condition information can include equipment operating status, equipment load, and equipment operating environment. For example, equipment operating status can include whether the equipment is energized, whether it is under load, or whether it is under maintenance; equipment load can include equipment load rate and equipment load type; and equipment operating environment can include ambient temperature, ambient humidity, and ambient wind speed; again, no specific limitations are imposed here.
[0033] S200. Input the status information into the pre-trained intelligent diagnostic model of the substation equipment; wherein, the intelligent diagnostic model of the substation equipment is a neural network model built on a deep learning framework and trained using the historical status information of the primary substation equipment and the corresponding fault condition labels. The aforementioned collected state information serves as input to the intelligent diagnostic model for substation equipment. Based on this state information, the model ultimately outputs the final fault type. This intelligent diagnostic model is a neural network model built upon a deep learning framework. It utilizes a Transformer structure incorporating a multi-head self-attention mechanism to extract correlation information between input features in parallel. Through a fully connected layer, it outputs the probability distribution of fault types for the primary substation equipment. This model has undergone extensive offline training using massive amounts of historical state information from the primary substation equipment and expert-confirmed fault labels, possessing the intelligent reasoning ability to map multi-source input state information into fault probabilities. Therefore, the intelligent diagnostic model for substation equipment proposed in this invention possesses strong feature learning and contextual understanding capabilities, is suitable for complex and ever-changing equipment state identification tasks, and exhibits stable training, fast convergence, and accurate diagnosis.
[0034] Specifically, in this embodiment of the invention, the training parameters of the intelligent diagnostic model for the power equipment include: 2 encoder layers, 4 multi-head attention heads, a dropout rate of 0.2, a batch size of 128, and a learning rate of 0.0001. It should be noted that the specific training parameters are not specifically limited here.
[0035] S300. Based on the output of the intelligent diagnostic model for power equipment, determine whether there is a fault in the primary equipment of the power equipment and the corresponding fault type. S400 If a fault exists in the primary equipment of the substation, an alarm signal containing at least the fault type and equipment location information is generated and sent to the monitoring terminal.
[0036] Based on the trained intelligent diagnostic model for substation equipment, the model is applied to the online overheating fault diagnosis of primary equipment in substations. In practical application, the collected status information is processed and then input into the trained intelligent diagnostic model. Finally, the intelligent diagnostic model outputs the diagnostic results: if a fault occurs, the alarm signal consisting of the determined fault type of the primary equipment, the corresponding data such as the acquisition time, geographical coordinates (equipment location information), and environmental parameters is transmitted back to the monitoring terminal; if no fault occurs, no alarm signal is sent to the monitoring terminal. In this way, the intelligent fault diagnosis of the primary equipment is completed.
[0037] It is evident that the pre-trained intelligent diagnostic model for substation equipment can not only determine whether a fault exists, but also provide specific fault types and corresponding data information (including equipment ID, geographical coordinates, fault type, occurrence time, associated infrared thermal imaging spectrum, etc.) to generate structured alarm signals. These signals are then pushed to the main monitoring center or the mobile terminals of maintenance personnel via a dedicated power network or secure wireless channel. This provides maintenance personnel with clear inspection directions, helps implement differentiated and precise maintenance strategies, avoids blind power outages for inspection, and improves the scientific nature of maintenance decisions.
[0038] The following pre-training process of the intelligent diagnostic model for power equipment serves as the foundation for model learning. In this embodiment of the invention, a training dataset is used to optimize model parameters, while an independent validation dataset is used to evaluate model performance, prevent overfitting, and ensure that the model has good generalization ability for unknown data.
[0039] Specifically, the pre-training process of the intelligent diagnostic model for power equipment includes the following steps: Step 1: Obtain the training dataset and validation dataset for training the intelligent diagnostic model of substation equipment; Step 2: Train the intelligent diagnostic model for substation equipment based on the training dataset; Step 3: Use the validation dataset to evaluate the training effect of the intelligent diagnostic model for substation equipment, obtain evaluation indicators, and when the evaluation indicators meet the preset training requirements, confirm that the intelligent diagnostic model for substation equipment has completed training.
[0040] As explained above, standardized dataset partitioning is fundamental to the machine learning process, ensuring the scientific rigor of the training and the reliability of the final model performance. Specifically, using the training dataset, optimization methods such as backpropagation are employed to iteratively update the model's weights and bias parameters, minimizing the loss function between the predicted output and the actual fault labels. The training effectiveness of the intelligent diagnostic model for substation equipment is evaluated using a validation dataset. Common evaluation metrics include accuracy, precision, recall, F1 score, or mean absolute error (MAE). When the metrics on the validation set remain stable and reach the preset target (e.g., accuracy > 98%), training is stopped, and the final model parameters are saved. This ensures the robustness and practicality of the deployed model.
[0041] In a preferred embodiment, see Figure 2 and Figure 3 Obtain the training dataset and validation dataset for training the intelligent diagnostic model of substation equipment, including: Step A1: Collect initial infrared training image sets corresponding to different types of primary substation equipment, filter each initial infrared training image set to obtain the infrared training image set corresponding to each type of primary substation equipment; Furthermore, in order to fundamentally eliminate invalid samples dominated by environmental noise and ensure the purity of the training dataset so that the model learns the real thermal features of the equipment rather than environmental interference artifacts, the process of "screening each initial infrared training map set to obtain the infrared training map set corresponding to each type of substation primary equipment" includes: Step A11: Obtain the environmental acquisition conditions for each infrared training image in the initial infrared training image set; wherein, the environmental acquisition conditions include at least: meteorological conditions, light conditions, and electromagnetic radiation interference conditions. Step A12: Compare the meteorological conditions, light conditions, and electromagnetic radiation interference conditions with the preset standard environmental acquisition conditions. Retain the infrared training spectra that meet the preset standard environmental acquisition conditions and delete the infrared training spectra that do not meet the standard environmental acquisition conditions to obtain the infrared training spectra set corresponding to each type of substation primary equipment.
[0042] In this embodiment of the invention, infrared training images of each primary substation device are collected in advance using a fixed or patrol-type infrared thermal imager. For each type of primary substation device, an infrared training image set consisting of multiple infrared training images is obtained through periodic collection.
[0043] The selection of infrared training images is mainly based on the environmental acquisition conditions during the acquisition process, as detailed below: Condition 1: Meteorological conditions refer to whether the images were taken under weather conditions such as thunder, rain, fog, and snow.
[0044] Condition 2: Light conditions include whether the images were taken outdoors under direct or reflected sunlight, or indoors under direct artificial light.
[0045] Condition 3: Electromagnetic radiation interference conditions include whether the images were taken under conditions where there are strong electromagnetic fields, heat radiation sources, and human body heat sources nearby, such as welding, starting of large motors, etc.
[0046] Different infrared training spectra are compared with preset standard environmental acquisition conditions (such as humidity ≤85%, no severe weather, no direct light / strong reflection, no strong electromagnetic, thermal radiation sources and human body heat sources) according to the above conditions 1-3. Only the infrared training spectra that fully meet the preset standard environmental acquisition conditions are retained. In this way, a high-quality infrared training spectra set for each type of substation primary equipment can be constructed.
[0047] Step A2: Perform filtering preprocessing and image processing on all infrared training maps in each infrared training map set to obtain the binarized edge map corresponding to each infrared training map, so as to construct the final training map set. Specifically, see Figure 4 Step A2 above includes the following processing steps: Step A21: Use a smoothing filter function to filter all infrared training images in the infrared training image set; Preprocessing is performed on the infrared training images in each infrared training image set to improve the distinguishability of edges and hot spots. The preprocessing methods include: first, converting the image from the original format to RGB three channels while retaining the original color bar information; then, smoothing the image using a Gaussian blur algorithm; generating the weights corresponding to the convolution kernel through a smoothing filter function; and finally, performing convolution calculation on the image to filter out random noise and subtle interference.
[0048] Furthermore, the formula for calculating the smoothing filter function is as follows: (1) Formula (1); in, It is a smoothing filter function; The size of the convolution kernel. , These represent the length and width of the input infrared training map, respectively.
[0049] Step A22: Based on the filtered infrared training spectrum image, calculate its gradient in the horizontal and vertical directions to obtain the gradient magnitude and gradient direction of each pixel in the corresponding image; Step A23: For each pixel, compare its gradient magnitude with the gradient magnitude of its neighboring pixels along its gradient direction, and filter each pixel according to the preset first magnitude threshold and second magnitude threshold to obtain the binarized edge map corresponding to each image. Among them, all the binarized edge maps corresponding to each type of primary substation equipment constitute its final training graph set.
[0050] Combining steps A22 and A23 above, after obtaining the filtered infrared training spectrum image, the gradient magnitude and gradient direction can be calculated based on the pixels displayed on the image. By comparing the gradient magnitude along the gradient direction, the goal of retaining only local maxima and suppressing non-edge gradient values to zero can be achieved.
[0051] Further, see Figure 5 Step A23 specifically includes the following steps: Step A231: Obtain the gradient magnitude of each pixel along its gradient direction and compare it with that of its neighboring pixels. Based on the magnitude comparison results, retain the local maxima points in the image. Step A232: Set a first amplitude threshold and a second amplitude threshold, wherein the first amplitude threshold is greater than the second amplitude threshold; Step A233: Obtain the gradient magnitudes corresponding to all retained local maxima points and use them as the target gradient magnitude set; Step A234: Mark the pixels in the target gradient magnitude set that are higher than the first magnitude threshold as strong edge pixels, remove the pixels in the target gradient magnitude set that are lower than the second magnitude threshold, and mark the pixels in the target gradient magnitude set that are between the first magnitude threshold and the second magnitude threshold as weak edge pixels. Step A235: For weak edge pixels, if they are adjacent to or connected to any strong edge pixel, they are retained as edge pixels; if they are not adjacent to or connected to any strong edge pixel, they are discarded. Step A236: Assign a value of 1 to all retained edge pixels and a value of 0 to the remaining pixels to obtain the binarized edge map corresponding to the filtered infrared training map.
[0052] In practical applications, the above process can be summarized as follows: First, non-maximum suppression is performed on the gradient magnitude of pixels in each infrared training spectrum to retain local maxima; then, based on the preset first magnitude threshold and second magnitude threshold, double threshold hysteresis processing and edge connection are performed on the pixels after non-maximum suppression to generate a binarized edge map.
[0053] Specifically, non-maximum suppression processing is applied to the gradient magnitude of pixels in each infrared training spectrum: The first step is to use classic image gradient operators (such as Sobel operator, Prewitt operator, etc.) to perform convolution operation on the filtered infrared training spectrum, and calculate the gradient value of each pixel in the horizontal direction (x-axis) and vertical direction (y-axis). These two gradient values can be denoted as horizontal gradient and vertical gradient, respectively, to reflect the gray level change rate of the image at that pixel along the x-axis and y-axis.
[0054] The second step is to calculate the gradient magnitude and gradient direction of each pixel based on its horizontal and vertical gradients. The gradient magnitude represents the strength of the edge at that pixel; the larger the magnitude, the higher the probability that the pixel belongs to an edge. The gradient direction is perpendicular to the possible edge direction at that pixel, providing a basis for determining the local direction of the edge in the next step.
[0055] The third step is to compare the gradient magnitude of each pixel in the image with the gradient magnitude of its two adjacent pixels along the gradient direction calculated in the previous step within a discrete neighborhood (usually the two adjacent pixels closest to the gradient direction, such as the four main directions of 0°, 45°, 90°, and 135°). Only if the gradient magnitude of the pixel is a local maximum along the gradient direction is its value retained. Therefore, the definition of a local maximum point is the pixel with the largest gradient magnitude in the discrete neighborhood; otherwise, its gradient magnitude is set to zero.
[0056] Next, double threshold hysteresis processing and edge connection are performed on the pixels after non-maximum suppression: The first step is to set two gradient magnitude thresholds, namely the first magnitude threshold and the second magnitude threshold, and satisfy the condition that the first magnitude threshold is greater than the second magnitude threshold. The two magnitude thresholds are used as the high threshold and the low threshold, respectively, to filter all pixels retained after non-maximum suppression.
[0057] The second step involves marking pixels with gradient magnitudes higher than the first magnitude threshold as strong edge pixels along the gradient direction, removing pixels with gradient magnitudes lower than the second magnitude threshold, and marking pixels with gradient magnitudes between the first and second magnitude thresholds as weak edge pixels.
[0058] The third step is to perform edge connectivity on weak edge pixels: each weak edge pixel is examined, and it is only accepted as a valid edge pixel if it is connected to at least one strong edge pixel in its neighborhood; otherwise, it is discarded. This ensures that only weak edges that form a continuous contour with strong edges are retained, effectively filling edge gaps while suppressing isolated noise points.
[0059] Finally, all pixels ultimately determined as valid edges (strong edge pixels and retained weak edge pixels) are assigned a value of 1 (white), while all other pixels in the image are assigned a value of 0 (black), thus generating a clear, coherent binary edge map containing only the device outline information. Simultaneously, all binary edge maps obtained after processing the infrared training images of all primary substation equipment of the same type (such as circuit breakers, disconnectors, etc.) through the above process are compiled to form a standardized training map set for training the intelligent diagnostic model of the substation equipment.
[0060] As can be seen, by employing the combined strategy of nonmaximum suppression and double-threshold hysteresis, this step can robustly extract true, complete, single-pixel-width device edges from noisy gradient information. This significantly improves the quality of input features for subsequent neural network models and effectively enhances the accuracy of fault identification.
[0061] Step A3: Based on the training graph set and the labels corresponding to each binarized edge map in the training graph set, construct the training dataset and the validation dataset.
[0062] See Figure 6 This step connects data preprocessing and model training, aiming to integrate standardized image data with multi-dimensional auxiliary information to construct a structured dataset suitable for training and evaluating neural network models within a deep learning framework. It is evident that by systematically constructing a high-quality, multimodal labeled dataset, rich and accurate learning samples are provided for subsequent neural network models. Specifically, this step includes the following processes: Step A31: Obtain training map sets for different types of substation primary equipment and the labels corresponding to each binarized edge map in the training map set. The labels include at least: original color bar information and its corresponding acquisition time, geographical coordinates, environmental parameters, operating condition information, and fault condition labels. Explanation of the label: Tag 1: Retain the color bar extracted from the raw data of the infrared thermal imager. This color bar defines the mapping relationship between color and absolute temperature, and serves as the benchmark for subsequent quantitative temperature analysis (such as calculating absolute temperature values, temperature differences between different parts of the same equipment, or temperature differences between the same part at different times), ensuring that the diagnosis moves from qualitative to quantitative.
[0063] Tag 2: Acquisition time is the precise timestamp for recording the image, used to correlate historical data and analyze the development trend of defects (such as the temperature rise rate of thermal defects).
[0064] Tag 3: Geographic coordinates record the specific location of equipment within the substation (such as latitude and longitude or station code), enabling precise fault location and supporting location-based statistical analysis.
[0065] Tag 4: Environmental parameters include ambient temperature, relative humidity, wind speed, and light intensity during recording. These parameters are crucial for calibrating the impact of the environment on the device's surface temperature, helping the model distinguish between heat generated by the device itself and heat generated by the environment.
[0066] Tag 5: Operating status information refers to the real-time electrical quantities of the equipment at the time of acquisition, such as three-phase current, voltage, active / reactive power, and load rate, which are synchronously obtained from the station's monitoring system. This is the core basis for determining whether the overheating is caused by overload or other operating conditions, or by defects in the equipment itself (such as poor contact or insulation deterioration).
[0067] Tag 6: Fault Status Tag is an authoritative diagnostic conclusion given by domain experts based on a comprehensive analysis of this inspection (i.e., combining different detection methods). This tag is the standard answer for model learning, and its form can be a specific fault type (such as "disconnector contact overheating", "abnormal bushing dielectric loss") or a more general state classification (such as "normal", "current-induced heating defect", "voltage-induced heating defect"). No special limitations are made here.
[0068] Based on this, the constructed integrated labeling system of image features, environment, operating conditions, and fault conditions creates conditions for deep learning models to learn cross-modal association rules. The model can not only identify image patterns but also learn the deeper logic of what image features appear in what environments and operating conditions, corresponding to what fault types, thus enhancing the reliability of diagnostic results and the practicality of the system.
[0069] Step A32: Divide each binarized edge map and its label into a training dataset and a validation dataset according to a preset ratio, so as to construct the training dataset and validation dataset corresponding to each type of substation primary equipment.
[0070] In practical applications, all samples (binarized edge maps and their labeled data packets) corresponding to a specific type of primary transformer equipment (such as a circuit breaker) are randomly and stratified into a predetermined ratio (e.g., 8:2 or 7:3). The larger proportion (e.g., 80%) is used as the training dataset to directly adjust the model parameters, while the smaller proportion (e.g., 20%) is used as the validation dataset to independently evaluate the model performance during training.
[0071] It should be noted that, to ensure the fairness of the partitioning, especially when the number of samples in different fault categories is unbalanced, a stratified sampling strategy can be adopted. This means that the proportion of samples in each fault category in the training and validation sets is basically consistent with the overall proportion in the original dataset, preventing the model from underlearning a few categories due to partitioning bias. At the same time, it is also necessary to ensure that a data sample can only be assigned to one of the training or validation sets, ensuring that the two sets are completely independent and there is no data leakage.
[0072] In a preferred embodiment, the training parameters of the intelligent diagnostic model for power equipment include: 2 encoder layers, 4 multi-head attention heads, a dropout rate of 0.2, a batch size of 128, and a learning rate of 0.0001. The intelligent diagnostic model for substation equipment is based on a Transformer structure that includes a multi-head self-attention mechanism to extract the correlation information between input features in parallel and output the probability distribution of fault types of primary substation equipment through a fully connected layer.
[0073] Specifically, see Figure 7 The training dataset is fed into the intelligent diagnostic model for substation equipment for training. The intelligent diagnostic model for substation equipment is mainly composed of a Transformer model and an output fully connected layer. The key structure of the Transformer model is the multi-head self-attention mechanism layer. The multi-head self-attention mechanism layer is calculated as follows: Formula (2) and Formula (3): Formula (2); Formula (3); Based on the above formulas (2) and (3), the intelligent diagnostic model for substation equipment first performs position encoding on the input sequence, and then performs three sets of learnable linear transformations on the encoded sequence to obtain the query matrix. Key matrix Value matrix And based on the dot product of self-attention and the equal distribution of attention heads, we obtain Then calculate to get The group attention module allows the model to extract attention separately. , and The relationship between them.
[0074] In the above formula, For the activation function of the self-attention mechanism layer; This represents the parameters that need to be trained in the self-attention mechanism layer; Indicates the length of the vector; AV is an abbreviation for Attention; For the first i The output of each attention head.
[0075] The attention values are merged by column, corresponding to the following formula (4): Formula (4); in, This is a multi-head self-attention output.
[0076] Finally, after transformation by the fully connected layer, the specific calculation is shown in the following formula (5): Formula (5); in, Indicates the first The output of each neuron Indicates the first The input of each neuron, and These represent the weights and biases, respectively. It represents the probability of belonging to a certain type of defect.
[0077] Finally, the training effect of the intelligent diagnostic model for substation equipment is evaluated using the validation dataset. When the evaluation index (mean absolute error) calculated from the validation dataset is less than the set value, it is determined that the intelligent diagnostic model for substation equipment meets the preset training requirements, that is, the current intelligent diagnostic model for substation equipment has completed training and the model parameters are saved. The formula for calculating the mean absolute error is as follows: Formula (6): Formula (6); in, Mean absolute error; Indicates the number of batches. Indicates the first The absolute accuracy of predictions for each batch.
[0078] Therefore, after the intelligent diagnostic model for power equipment has been trained, it can be applied to the online overheating fault diagnosis of primary power equipment in substations; by real-time acquisition of infrared thermal images of primary power equipment (see...). Figure 8 The system records the collection time, geographical coordinates, environmental parameters, and operating condition information. The input data for online fault diagnosis first needs to undergo the same data preprocessing operation as the training data in the previous steps, and then it is input into the trained intelligent diagnostic model of the substation equipment to obtain the diagnostic results. If a fault occurs, the alarm signal consisting of the fault type of the primary substation equipment and the corresponding data information such as the collection time, geographical coordinates (equipment location information), and environmental parameters is transmitted back to the monitoring terminal. If it is determined that no fault has occurred, no alarm signal is sent to the monitoring terminal. Based on this, the intelligent fault diagnosis of the primary substation equipment is completed.
[0079] As can be seen, to address the problems of large data volume, disordered recording, low efficiency of manual diagnosis, and difficulty in co-analyzing with electrical quantity data in existing infrared spectral analysis, this invention proposes an intelligent diagnostic method for primary equipment in substations, with the following main effects: Effect 1: Deeply linking infrared data with equipment files, operating conditions, and environmental parameters enables collaborative analysis of infrared data and electrical quantity data, facilitating condition assessment and fault diagnosis, and solving problems such as low infrared defect detection rate in field applications, especially the strong concealment and difficulty in detecting voltage-induced thermal defects.
[0080] Effect 2: Gaussian blur filtering is applied to the acquired images to suppress noise and extract the equipment contours, effectively improving the accuracy of equipment recognition; in addition, the intelligent diagnostic model for substation equipment integrates the multi-head self-attention mechanism of the Transformer model to improve the model's global relationship processing capability, thereby improving the effectiveness of feature processing.
[0081] Effect 3: It can perform large-scale, automated diagnosis of infrared spectra of different equipment types. By introducing deep learning into infrared spectra analysis, it can automatically complete target recognition, feature extraction and fault diagnosis, reduce the probability of missed or false judgments in the case of large amounts of data processing, save time and labor costs in processing spectra and producing equipment fault analysis reports, and reduce the workload of grassroots staff.
[0082] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to the specific combination of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this invention.
Claims
1. A method for intelligent diagnosis of primary equipment in a substation, characterized in that, include: Acquire the status information of the primary equipment of the substation, the status information including at least the infrared thermal imaging spectrum of the primary equipment of the substation, and the acquisition environment parameters and equipment operating condition information corresponding to the infrared thermal imaging spectrum; The state information is input into a pre-trained intelligent diagnostic model for substation equipment; wherein, the intelligent diagnostic model for substation equipment is a neural network model built based on a deep learning framework and trained using the historical state information of the primary substation equipment and corresponding fault condition labels; Based on the output of the intelligent diagnostic model for the power equipment, determine whether the primary equipment of the power equipment has a fault and the corresponding fault type; If a fault occurs in the primary equipment of the substation, an alarm signal containing at least the fault type and equipment location information is generated and sent to the monitoring terminal.
2. The intelligent diagnostic method for primary equipment in a substation according to claim 1, characterized in that, The pre-training process of the intelligent diagnostic model for power equipment includes the following steps: Obtain the training dataset and validation dataset for training the intelligent diagnostic model of the substation equipment; The intelligent diagnostic model for the substation equipment is trained based on the training dataset. The training effect of the intelligent diagnostic model for substation equipment is evaluated using the validation dataset to obtain evaluation indicators. When the evaluation indicators meet the preset training requirements, the training of the intelligent diagnostic model for substation equipment is confirmed to be complete.
3. The intelligent diagnostic method for primary equipment in a substation according to claim 2, characterized in that, The step of obtaining the training dataset and validation dataset for training the intelligent diagnostic model of the substation equipment includes: Initial infrared training image sets corresponding to different types of primary substation equipment are collected, and each initial infrared training image set is filtered to obtain an infrared training image set corresponding to each type of primary substation equipment. All infrared training maps in each infrared training map set are filtered and preprocessed, and image processing is performed to obtain a binarized edge map corresponding to each infrared training map, so as to construct the final training map set. The training dataset and the validation dataset are constructed based on the training graph set and the labels corresponding to each binarized edge map in the training graph set.
4. The intelligent diagnostic method for primary equipment in a substation according to claim 3, characterized in that, The step of filtering each of the initial infrared training image sets to obtain the infrared training image set corresponding to each type of primary substation equipment includes: The environmental acquisition conditions for each infrared training image in the initial infrared training image set are obtained; wherein, the environmental acquisition conditions include at least: meteorological conditions, light conditions, and electromagnetic radiation interference conditions; The meteorological conditions, light conditions, and electromagnetic radiation interference conditions are compared with preset standard environmental acquisition conditions. Infrared training spectra that meet the preset standard environmental acquisition conditions are retained, while infrared training spectra that do not meet the standard environmental acquisition conditions are deleted, so as to obtain the infrared training spectra set corresponding to each type of substation primary equipment.
5. The intelligent diagnostic method for primary equipment in a substation according to claim 3, characterized in that, The step of preprocessing and edge processing all infrared training maps within each infrared training map set to obtain a binarized edge map corresponding to each infrared training map includes: A smoothing filter function is used to filter all infrared training images in the infrared training image set. Based on the filtered infrared training spectrum image, its horizontal and vertical gradients are calculated to obtain the gradient magnitude and gradient direction of each pixel in the corresponding image. For each pixel, its gradient magnitude is compared with the gradient magnitude of its neighboring pixels along its gradient direction, and each pixel is filtered according to a preset first magnitude threshold and a second magnitude threshold to obtain the binarized edge map corresponding to each image. Among them, all the binarized edge maps corresponding to each type of primary substation equipment constitute its final training graph set.
6. The intelligent diagnostic method for primary equipment in a substation according to claim 5, characterized in that, The step of comparing the gradient magnitude of each pixel along its gradient direction with the gradient magnitude of its neighboring pixels, and filtering each pixel according to a preset first magnitude threshold and a second magnitude threshold to obtain a binarized edge map corresponding to each image, includes: Obtain the gradient magnitude of each pixel along its gradient direction and compare it with that of its neighboring pixels. Based on the magnitude comparison results, retain the local maxima points in the image. Set the first amplitude threshold and the second amplitude threshold, wherein the first amplitude threshold is greater than the second amplitude threshold; Obtain the gradient magnitudes corresponding to all retained local maxima points and use them as the target gradient magnitude set; Pixels with a target gradient magnitude higher than a first magnitude threshold are marked as strong edge pixels; pixels with a target gradient magnitude lower than a second magnitude threshold are removed; and pixels with a target gradient magnitude between the first magnitude threshold and the second magnitude threshold are marked as weak edge pixels. For the weak edge pixel, if it is adjacent to or connected to any of the strong edge pixels, it is retained as an edge pixel; if it is not adjacent to or connected to any of the strong edge pixels, it is discarded. All retained edge pixels are assigned a value of 1, and the remaining pixels are assigned a value of 0, to obtain the binarized edge map corresponding to the filtered infrared training map.
7. The intelligent diagnostic method for primary equipment in a substation according to claim 3, characterized in that, The training dataset and the validation dataset are constructed based on the training graph set and the labels corresponding to each binarized edge map within the training graph set, including: Acquire training map sets of different types of substation primary equipment and the labels corresponding to each binarized edge map in the training map set, wherein the labels include at least: original color bar information and its corresponding acquisition time, geographical coordinates, environmental parameters, operating condition information and fault condition labels; The binarized edge maps and their labels are divided into training datasets and validation datasets according to a preset ratio to construct the training dataset and validation dataset corresponding to each type of substation primary equipment.
8. The intelligent diagnostic method for primary equipment in a substation according to claim 1, characterized in that, The training parameters of the intelligent diagnostic model for the substation equipment include: 2 encoder layers, 4 multi-head attention heads, a dropout rate of 0.2, a batch size of 128, and a learning rate of 0.0001. The intelligent diagnostic model for substation equipment is based on a Transformer structure that includes a multi-head self-attention mechanism to extract correlation information between input features in parallel, and outputs the probability distribution of fault types of the primary substation equipment through a fully connected layer.