Substation early warning method and system based on data driving

Abstract

The invention provides a transformer substation early warning method based on data driving, relates to the field of image recognition, and is specifically applied to fault early warning of transformer cooling fin detection. The transformer substation early warning method based on data driving specifically comprises the steps that visible light image data and infrared image data of transformer cooling fins in a transformer substation are collected to construct a multi-mode superposition image; a cooling fin fault result is obtained based on multi-mode superposition image detection; performing a cooling fin transmittance experiment by using the colored smoke, and acquiring video data in the experiment process to obtain a video detection result; and combining the video data detection result and the visible light image data detection to obtain a cooling fin fault result, and judging whether to execute early warning or not. Through the above steps, the method not only overcomes the limitation of a single detection means, but also can remarkably improve the accuracy of fault early warning, and provides technical support for intelligent operation and maintenance of the transformer substation.

Description

Technical Field

[0001] The present invention relates to the field of image recognition, and particularly to a data-driven substation early warning method and system, which are specifically applied to the fault early warning of transformer radiator detection. Background Art

[0002] As a core hub device in the power system, the operation status of a substation is directly related to the safety and stability of the entire power network. As an important component of the substation, the transformer undertakes the key tasks of voltage conversion and power transmission, and its radiator is the core component to ensure the normal heat dissipation of the transformer and maintain the stable operation of the equipment. Once the radiator has operation failures, such as blockage, deformation, corrosion or local overheating, it will directly lead to a decrease in the heat dissipation efficiency of the transformer, thereby affecting the equipment life and operation stability, and may even cause large-scale power interruption accidents. Therefore, in order to ensure the safe and stable operation of the power grid, the fault early warning of the transformer radiator is of extremely high importance. The current operation and maintenance management of substations still mainly relies on manual inspections, supplemented by some monitoring devices. However, due to the long cycle and low accuracy of manual detection, and the single traditional equipment monitoring means, it is difficult to comprehensively cover the potential fault hidden dangers of the radiator. This current situation makes the intelligence and high efficiency of the substation early warning system an urgent problem to be solved. Especially in the field of radiator detection, there is an urgent need for a precise, efficient and reliable fault early warning means to fill the gap in the existing technology.

[0003] With the rapid development of modern technologies, especially the breakthroughs in artificial intelligence, image recognition and infrared spectral thermal imaging technologies, these technologies have been widely applied in many fields. For example, image recognition technology is widely used in the transportation field for vehicle detection and violation behavior recognition, and in the medical field for medical image analysis to assist doctors in diagnosing diseases; infrared spectral thermal imaging technology has been reliably applied in the fields of thermal anomaly detection of industrial equipment, building energy efficiency assessment and military reconnaissance. These technologies provide strong technical support for the intelligent development of various industries through efficient feature extraction and intelligent data analysis. However, in the field of early warning of the transformer heat dissipation capacity, the application of these technologies is still in the initial exploration stage, and a reliable detection and early warning system has not been formed yet. As a key component, the transformer radiator has a complex operation environment and diverse fault forms, and a single technical means is difficult to meet the detection requirements in complex scenarios. Therefore, how to organically combine these technologies and fully explore their potential in transformer radiator detection is an important direction to promote the development of intelligent operation and maintenance technologies for substations.

[0004] In view of the deficiencies of the prior art, the present application proposes a data-driven substation early warning method and system. Through the cross-verification mechanism of multi-modal data, the accuracy and reliability of transformer radiator fault detection are significantly improved. The use of multi-source data-driven not only overcomes the limitations of single detection means, but also can significantly improve the accuracy of fault early warning, providing technical support for the intelligent operation and maintenance of substations. Summary of the Invention

[0005] The present invention provides a data-driven substation early warning method, which specifically includes the following steps:

[0006] S1: Collect visible light image data and infrared image data of the transformer radiator in the substation to construct a multi-modal superimposed image;

[0007] S2: Based on the multi-modal superimposed image, detect the fault result of the radiator;

[0008] S3: Conduct a transmittance experiment on the radiator using colored smoke and collect video data during the experiment;

[0009] S4: Combine the detection result of the video data and the detection result of the visible light image data to obtain the fault result of the radiator, and judge whether to execute early warning.

[0010] The present invention provides a data-driven substation early warning system, which includes:

[0011] Image acquisition device: The image acquisition device collects visible light image data and infrared image data of the transformer radiator in the substation to construct a multi-modal superimposed image;

[0012] Multi-modal attention module: The multi-modal attention module detects the fault result of the radiator based on the multi-modal superimposed image;

[0013] Video data acquisition module: Conduct a transmittance experiment on the radiator using colored smoke and collect video data during the experiment;

[0014] Early warning module: The early warning module combines the detection result of the video data and the detection result of the visible light image data to obtain the fault result of the radiator, and judges whether to execute early warning.

[0015] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned data-driven substation early warning method.

[0016] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned data-driven substation early warning method.

[0017] Compared with the prior art, the present invention proposes a data-driven substation early warning method. The present invention constructs multi-modal images based on infrared images and visible light images, and detects blocked areas based on visible light image features and thermal infrared features. During the specific detection process, the structural features of the heat sinks that appear periodically are suppressed to improve the defect detection ability. At the same time, channel attention and position attention information are fused. In addition, the present invention also adopts a colored smoke leakage experiment combined with an optical flow detection algorithm to calculate the passing rate and flow velocity information of the heat sinks based on video information, and obtains the abnormal heat dissipation situation of the transformer in the substation based on the video detection result and the image detection result, overcoming the limitations of a single detection method. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0019] Figure 1 It is a flowchart for the present invention to perform data-driven substation early warning. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] The embodiments of the present application will be described in detail below with reference to the drawings.

[0021] The following specific examples illustrate the implementation manners of the present application. Those skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The present application can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts belong to the scope of protection of the present application.

[0022] It should be noted that the following description relates to various aspects of embodiments within the scope of the appended claims. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art should understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspects set forth herein can be used to implement an apparatus and / or practice a method. Additionally, this apparatus and / or method can be implemented using other structures and / or functionality in addition to one or more of the aspects set forth herein.

[0023] In addition, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the examples can be practiced without these specific details.

[0024] An embodiment of this specification proposes a data-driven substation early warning method, which specifically includes the following steps:

[0025] S1: Collect visible light image data and infrared image data of the transformer radiator in the substation to construct a multimodal superimposed image;

[0026] S2: Detect the radiator failure result based on the multimodal superimposed image;

[0027] S3: Conduct a transmittance experiment on the radiator using colored smoke and collect video data during the experiment;

[0028] S4: Combine the detection result of the video data and the visible light image data to detect the radiator failure result and determine whether to execute the early warning.

[0029] In order to realize the supervision and detection of the transformer radiator in the substation, the present invention designs a set of acquisition schemes based on visible light image data and infrared image data. During the acquisition process, a high-resolution industrial-grade visible light camera and a high-precision infrared thermal imager are selected as the main acquisition devices. The visible light camera is responsible for capturing the external physical state of the radiator, while the infrared thermal imager is used to obtain the temperature distribution on the surface of the radiator and measure abnormal phenomena through infrared thermal radiation. These two devices are respectively installed directly opposite the transformer radiator.

[0030] During the equipment installation phase, the visible light camera and the infrared thermal imager are installed side by side and fixed on a dedicated metal bracket. The height of the bracket is kept level with the center of the heat sink, which can not only avoid data distortion caused by shooting angle deviation but also ensure the spatial consistency of the two types of image data. This side-by-side installation method enables the visible light image and the infrared image to be collected within the same field of view. In addition, the bracket has high stability and is equipped with anti-seismic devices to adapt to the vibration factors that may occur in the complex operating environment of the substation. The outside of the equipment is equipped with a protective cover with an IP66 or higher protection level, which can prevent external factors such as dust from interfering with the equipment and can also cope with harsh climate conditions such as high temperature and high humidity. To achieve all-weather monitoring, the visible light camera also needs to be equipped with an infrared fill light to solve the problem of image acquisition at night or in low-light conditions.

[0031] The core of data acquisition lies in the synchronous operation of the visible light camera and the infrared thermal imager. For this purpose, the system uses the Precision Time Protocol (PTP) to precisely synchronize the two devices, ensuring that the two types of image data are collected at the same time point, thereby achieving consistency in time and space. The collected data is transmitted to the monitoring center in real time through industrial Ethernet or fiber optic communication. In addition, in some scenarios with difficult wiring, a 5G communication module can also be used for wireless data transmission to improve the flexibility and adaptability of the system.

[0032] Preferably, in this solution, the device for collecting visible light images selects the Dahua DH-IPC-HFW7842H-Z industrial camera. This model has an ultra-high definition resolution (4K, 3840×2160 pixels), supports low-light operation, and can still capture clear images under an ambient light of as low as 0.002 Lux. At the same time, the camera supports a video frame rate of 30 frames per second, which can meet the dynamic monitoring video acquisition requirements of the heat sink. In addition, the camera has a wide dynamic range (120dB WDR), and can present clear image details even in areas with strong light reflection or shadows of the transformer.

[0033] The infrared thermal imager selects the FLIR A400 thermal imager. This device has a high-resolution thermal imaging ability of 640×480 pixels, with a minimum thermal sensitivity of ≤40mK and a temperature measurement range of -20°C to 650°C, which fully covers the operating temperature range of the transformer heat sink and can adapt to extreme situations of sudden temperature increase. The FLIR A400 supports data transmission through an Ethernet interface and can transmit temperature data to the control center in real time for remote analysis and processing.

[0034] Preferably, in terms of installation layout, considering the position and distribution characteristics of the radiators, the equipment needs to cover the monitoring areas of all the radiators of the transformer and avoid data acquisition blind spots. Therefore, a combined layout method of four visible light cameras and four infrared thermal imagers is adopted. Each group of cameras and thermal imagers are installed side by side on a bracket, and are evenly arranged around the transformer. The distance between each group of brackets and the transformer is about 3 meters, and the installation height is kept level with the middle of the radiator to ensure that the shooting angle is perpendicular to the surface of the radiator, reducing image distortion and temperature measurement errors. The specific installation bracket is a customized type 304 stainless steel shockproof bracket, the adjustable range of the bracket height is 2 meters to 4 meters, and the bottom is fixed with ground piles to ensure the structural stability and reliability. A rotary pan-tilt is installed on the bracket, which supports the angle adjustment of the visible light camera and the infrared thermal imager.

[0035] In the data processing stage, the collected visible light image data is first optimized through denoising processing and image enhancement technology to ensure the clarity and accuracy of the images. The infrared images are analyzed for thermal anomalies through a temperature calibration algorithm, and a temperature distribution map that is easy to observe is generated through pseudo-color rendering. Subsequently, the system uses image registration technology to spatially align the visible light images and the infrared images to generate a multi-modal superimposed image.

[0036] Based on the multi-modal superimposed image, the radiator fault results are detected. The fault results include radiator blockages caused by the obstruction of air flow due to dust, grease, and foreign object accumulation. The fault results are detected through a multi-modal attention model. The multi-modal attention model includes a feature extraction backbone network, and the feature extraction backbone network is composed of a dual-branch feature extraction module and a feature enhancement module. The dual-branch feature extraction module is used to independently extract the visible light feature map and the infrared image feature map in the multi-modal superimposed image; the dual-branch feature extraction module is composed of the first convolution and the first residual block in two identical ResNet-18 networks;

[0037] The feature enhancement module includes a horizontal stripe extraction module. The horizontal stripe extraction module is used to extract the horizontal stripe information of the radiator and suppress the horizontal stripe features in the visible light feature map. The horizontal stripe extraction module is defined as:

[0038]

[0039] where, F vision (x,y) and F ifr (x,y) are respectively the visible light feature map and the infrared image feature map obtained by the dual-branch feature extraction module, h filter is the weight function of the filter, and are respectively the horizontal stripe weight maps of the extracted visible light and infrared images, Kenele vticalis a two-dimensional directional gradient kernel used to extract edge features in the longitudinal direction. is a sine modulation function used to enhance the filter's response to periodic features in the vertical direction. f represents the frequency, which is the periodic frequency of the filter response, φ is the phase shift representing the starting position for adjusting the sine modulation, and ★ represents the element-wise dot product. and are the enhanced visible light feature map and the infrared image feature map respectively. is the visible light feature map after suppressing the horizontal stripes. Conv represents the convolution calculation, and the subscript represents the convolution kernel size.

[0040] Exemplarily, if the stripe interval of the heat sink in the longitudinal direction is about three pixels, then f = 1 / 3. When the phase shift is 0, the sine modulation function and the weight function of the filter are respectively:

[0041]

[0042] Before performing image detection, the present invention first suppresses the periodic horizontal stripe features in the visible light image. The horizontal stripes of the transformer heat sink are its basic structural features. Since they present a periodic repetition pattern in the image, if the preprocessing of these horizontal stripes is ignored, these stripe features may interfere with subsequent blockage detection. Specifically, when the neural network analyzes an image, it relies on significant features to determine whether there is a blockage. Due to their periodicity and regularity, the horizontal stripes are easily misinterpreted by the model as important features, thus masking the real blockage information. This may cause the neural network to be interfered by the stripe features during the training and detection processes, making the model unable to focus on the details truly related to blockage, such as dirt, dust accumulation, or corrosion. In addition, suppressing the horizontal stripes can significantly improve the contrast and background clarity of the image, making the abnormal features related to blockage more prominent and easier to be correctly identified by the model. When certain areas of the heat sink change in appearance due to blockage, such as becoming dull in color, locally blurred, or having abnormal structures, after suppressing the stripes, these changes will be more conspicuous, and the model can more efficiently extract these abnormal features and make judgments.

[0043] The multi-modal attention model further includes a feature fusion module, and the feature fusion module includes two attention weight generation branches.

[0044]

[0045] Among them, represent the first and second attention weight maps respectively,, F fused (x, y) represents the output feature map of the feature fusion module.

[0046] In the multi-modal feature fusion stage of the present invention, two attention weights are introduced to analyze and guide the feature fusion. Attention weight 1 focuses on the difference analysis between image channels, and can effectively mine the global feature relationships between different channels. Since the feature distributions and expression methods of different modal images are different, the channel attention mechanism can assign different importance weights to features, enabling the model to pay more attention to the modal features contributing to the final task during the fusion process, suppressing the interference of redundant or irrelevant features, and thus improving the efficiency and accuracy of feature fusion. At the same time, attention weight 2 focuses on the position information within the same channel, analyzes the feature distribution on the position information, and by capturing local details and spatial correlations, enables the fused features to more accurately retain the structural and detailed features in the image. Especially in tasks involving local differences (such as blocked areas, hot spots in images, etc.), this attention mechanism is particularly important. The combination of the two attention weights can achieve collaborative optimization between the global and local, and between channels and positions, making the feature fusion of multi-modal images more comprehensive and accurate. Specifically, the channel attention weight strengthens the complementarity of different modal features at the semantic level, while the position attention weight ensures the integrity and consistency of the structural features during the fusion process, thus fully exploiting the joint expression ability of multi-modal features. This dual attention mechanism not only improves the robustness of feature fusion, but also effectively avoids the limitations that may occur in a single attention mechanism, making the fused features perform better in subsequent tasks, and ultimately improving the detection accuracy and reliability of multi-modal tasks.

[0047] The multi-modal attention model further includes a region proposal network and a classification regression network. The output feature map of the feature fusion module is input into the region proposal network to obtain a feature map with feature proposal boxes, and the classification regression network receives the feature map with feature proposal boxes and outputs the defect position information.

[0048] In the detection of transformer radiator blockage of the present invention, the detection is carried out by constructing a multi-modal superimposed image of visible light and infrared images, significantly improving the detection accuracy and reliability. During the operation of the transformer radiator, if blockage occurs, it usually manifests as surface deposition of dust, oil stains or other pollutants, and these surface features are important bases for analyzing blockage. Visible light images can clearly capture the physical features on the surface of the radiator, such as the accumulation of dust, oil stain coverage or local corrosion and other appearance abnormalities, and these information intuitively reflect the surface state of the radiator, providing key clues for judging blockage. However, relying solely on visible light images for detection has certain limitations. For example, when the blockage occurs in the internal channel of the radiator, there may be no obvious abnormal features on the surface. In this case, relying solely on visible light images may lead to missed detection or misjudgment of the blocked area.

[0049] On the other hand, the blockage of the transformer heat sink not only manifests as surface features but also directly affects its heat dissipation performance. When the internal channels of the heat sink are blocked, resulting in airflow obstruction, heat cannot be efficiently dissipated, and the temperature in the blocked area will rise significantly, forming local temperature anomalies. Infrared images, through thermal imaging technology, can accurately capture the temperature distribution characteristics of the heat sink, especially the hot spots caused by blockage. Compared with visible light images, infrared images directly reflect the thermal state changes of the heat sink and are particularly effective in detecting problems where there are no obvious external abnormalities but actual blockages inside. Therefore, infrared images can make up for the deficiencies of visible light images in detecting internal channel blockages, making the detection more comprehensive and accurate.

[0050] The present invention realizes a comprehensive analysis of the blockage of the transformer heat sink by fusing the multi-modal information of visible light images and infrared images, giving full play to the complementary advantages of the two modalities. Visible light images provide detailed information on surface physical characteristics, while infrared images provide dynamic characteristics of internal heat distribution. The combination of this multi-modal information can simultaneously focus on the external abnormalities and internal thermal state changes of the heat sink. During the fusion process, the present invention not only retains the key information of the two modalities but also improves the detection accuracy through the complementarity between the modalities. For the blocked area, visible light images can effectively identify direct blockage information such as surface dirt and corrosion, while infrared images provide strong evidence by capturing the local temperature rise in the blocked area. The superposition of these dual characteristics makes the detection results more robust and reliable.

[0051] In order to capture a video of colored smoke passing through the heat sink, the present invention designs an environmental layout and video acquisition scheme.

[0052] To ensure that the smoke is easily recognizable in the video, it is necessary to select colored smoke with high contrast and uniform diffusion. Preferably, the present invention uses industrial-grade red or blue smoke generators. In addition, the colored smoke has high visibility and low corrosiveness, making it suitable for electronic devices.

[0053] In the smoke release stage: A portable directional smoke generator is used to generate colored smoke, which can be achieved by using a cold smoke generator. To ensure that the smoke can be released at a stable flow rate, the present invention selects an industrial smoke machine with fan control and adjusts the output power to cover the entire heat sink inlet area. Specifically, place the smoke generator directly in front of the transformer heat sink to ensure that the released smoke can evenly cover all heat sink units, and use a deflector or smoke diffuser to guide the smoke flow to avoid excessive or insufficient local concentration.

[0054] Taking the height of the transformer radiator used in the experiment as 1 meter and the width as 2 meters as an example, a smoke generator with a horizontal diffusion angle of 120° is placed at a position 50 centimeters away from the radiator inlet, and a flow deflector is equipped to ensure that the smoke evenly covers the entire surface of the radiator. The video acquisition device can use an independent industrial camera for video shooting, or the visible light camera used to collect visible light image data in step S1 can be selected. The camera is equipped with an infrared filter to enhance the color contrast of the smoke. The camera is set outside the radiator to ensure that the entire radiator area is within the camera's field of view. During video shooting, the camera maintains a vertical distance of about 1 meter from the radiator to obtain a clear image of the smoke flow and avoid focal length problems caused by being too close. Preferably, an auxiliary light source (such as an LED panel light) is added to provide uniform illumination and avoid insufficient or uneven ambient light.

[0055] In a specific embodiment, industrial-grade blue smoke is used to detect the penetration of the radiator, and the collected video is recorded as: where I t (x, y) represents the t-th frame image, and T is the total number of frames.

[0056] The collected video is subjected to image enhancement to obtain a differential image D t (x, y):

[0057] G t (x, y) = 0.114 · R t (x, y) + 0.587 · G t (x, y) + 0.299 · B t (x, y);

[0058] D t (x, y) = |G t (x, y) - B(x, y)|;

[0059] where R t (x, y), G t (x, y) and B t (x, y) are the R, G, and B channel images of I t (x, y) respectively, and B(x, y) is the reference background frame when there is no smoke passing through;

[0060] The blue region of the collected video is extracted and combined with the differential image to obtain a blue region mask M t (x, y):

[0061]

[0062] M t (x, y) = S t(x,y)·D t (x,y);

[0063] Among them, S t (x,y) represents a binarized mask image obtained according to a color threshold, and respectively represent the hue, saturation, and brightness channels after the t-th frame image is converted to the HSV space. S threshold and V threshold respectively represent the minimum saturation requirement and the minimum brightness requirement;

[0064] Calculate the passing rate and flow rate of the heat sink based on the blue area mask, and obtain the clogging situation of the heat sink.

[0065] Before calculating the passing rate and flow rate of the heat sink, divide the heat sink into a grid of N H *H W where each grid is defined as R i,j where i ∈ N H and j ∈ N W ;

[0066] Count the number of smoke pixels in the grid R i,j and calculate the passing rate

[0067]

[0068] where A i,j represents the total number of pixels in the grid R i,j ;

[0069] Use the dense optical flow method to calculate the average flow rate of the smoke in the grid:

[0070]

[0071] where u t (x,y) and v t (x,y) are the optical flow velocity components of the pixel point;

[0072] Based on the average passing rate and the average flow rate calculate the comprehensive clogging score, and determine whether there is a clogging situation in the grid area based on a preset clogging threshold. The comprehensive clogging score is defined as:

[0073]

[0074] where R th and V thThey represent the lowest passing rate and the lowest flow rate respectively. Considering that a small amount of smoke may not fully penetrate due to turbulence or diffusion, the lowest passing rate threshold is set to 0.8. If the value is lower than this, there may be a blockage. Considering that there may be certain fluctuations in the smoke flow rate, the lowest flow rate threshold is set to 5 px / frame. If the value is lower than this, there may be a blockage.

[0075] Combining the detection results of video data and the detection results of visible light image data to obtain the heat sink fault results, and determining whether to execute a warning.

[0076] Among them, the heat sink fault result obtained by detecting visible light image data is expressed as is the position coordinate box of the k-th blocked area, p k is the blockage probability of the k-th blocked area. The detection result of video data is expressed as represents the position coordinate of the grid R i,j ; for each heat sink fault result find all matching grids that intersect with , and the matching condition is: IoU is the intersection over union of two regions, which is used to judge the position matching degree; for each matching grid, combine the comprehensive blockage score and the blockage probability to judge the blockage situation, and execute a warning when the blockage situation exceeds the preset threshold.

[0077] In the blockage detection of transformer radiators, there are certain limitations in the methods that rely solely on images (visible light superimposed on infrared images) or solely on video smoke leakage detection, which may lead to false positives or false negatives. Each has its own advantages and disadvantages. Only by combining the two for mutual verification can a more accurate blockage detection result be achieved. First of all, the image-based detection method mainly relies on the analysis of the surface features of visible light images and infrared images. Visible light images can capture the physical appearance features of the radiator, such as whether there are problems like dirt, deformation, corrosion, or damage. Infrared images, on the other hand, reflect the heat dissipation performance through the temperature distribution, such as whether there are phenomena of local overheating or uneven heat dissipation. However, the core problem of the image detection method is that it can only judge the surface condition of the radiator and cannot directly reflect the smoothness of the internal air flow. The blockage problem of the radiator often occurs in the internal air flow channels, that is, between radiators or in the ventilation holes. If only the surface features are judged by image detection, false positives are very likely to occur. For example, the surface of a radiator may seem clean, but its internal channels may have been blocked by dust, foreign objects, or corrosion, resulting in air flow obstruction and thus affecting the heat dissipation performance. In this case, the image detection may draw a "normal" conclusion, but in fact, the heat dissipation performance of the radiator has seriously deteriorated. In addition, the accuracy of image detection is also affected by environmental conditions. Visible light images are easily interfered by the intensity of light, shadows, and shooting angles, resulting in the omission or misjudgment of key features. For example, in low-light conditions, the details of dirt or corrosion may not be clearly presented, while in strong light environments, the reflected light may cover up the abnormal features on the surface. Similarly, although infrared images can reflect the temperature distribution, they are easily interfered by external heat sources. For example, when there are other heat sources near the transformer, the infrared image may capture the thermal radiation signals of non-radiator components, leading to misjudgment of the temperature distribution. In addition, the infrared image has limited ability to distinguish heat dissipation abnormalities and may not be able to accurately capture some minor uneven heat dissipation phenomena. These problems may all lead to inaccurate image detection results.

[0078] On the other hand, smoke penetration detection is a dynamic detection method that directly reflects the air flow smoothness of the internal channels of the heat sink by observing the transmittance and flow rate changes of smoke in the heat sink grid. Compared with image detection, smoke penetration detection can more directly evaluate the internal blockage situation. Even if the surface of the heat sink seems in good condition, smoke detection may still find signs of blocked internal air flow. However, smoke penetration detection also has its drawbacks. First of all, it is easily interfered by environmental factors. For example, problems such as wind speed, air flow fluctuations, or uneven smoke diffusion may lead to deviations in the detection results. For example, the power problem of the generator for generating directional smoke results in a low smoke transmittance, thus misjudging some areas as blocked. Similarly, the unevenness of smoke diffusion may make some unblocked areas appear to have a lower transmittance, while some blocked areas may seem to have a faster flow rate due to smoke concentration. In addition, the accuracy of smoke detection is also affected by the stability of the smoke source. If the concentration or flow rate of the smoke source is unstable, it may cause fluctuations in the detection results, further increasing the risk of misjudgment. Therefore, although smoke detection can effectively complement the deficiencies of image detection in internal air flow assessment, it also has obvious limitations of its own.

[0079] Relying solely on image detection or solely on smoke detection may both have problems of false positives or false negatives. Image detection may miss internal blockages due to overemphasis on surface features, while smoke detection may misjudge unblocked areas as blocked due to environmental interference. Combining the two for mutual verification is the best way to improve the detection accuracy and robustness. By combining the surface feature analysis of image detection with the internal air flow assessment of smoke detection, a comprehensive judgment of the heat sink blockage situation can be achieved. For example, when the image detection results indicate that a certain area may have serious blockage, if the smoke detection results also support this conclusion, it can be more confidently judged that there is indeed a blockage problem in this area. On the contrary, if there is a conflict between the image detection and smoke detection results, such as the image detection indicates that an area is unblocked while the smoke detection indicates blockage, further analysis and verification can be carried out to avoid misjudgment that may be brought by a single method. Through the weighted fusion of the two detection results, the advantages and disadvantages of the two methods can be better balanced, and finally a more accurate and comprehensive detection conclusion can be obtained.

[0080] The present invention provides a data-driven substation early warning system, which includes:

[0081] Image acquisition device: The image acquisition device acquires visible light image data and infrared image data of the transformer heat sink in the substation to construct a multi-modal superimposed image;

[0082] Multi-modal attention module: The multi-modal attention module detects the heat sink fault result based on the multi-modal superimposed image;

[0083] Video data acquisition module: Conduct a heat sink transmittance experiment using colored smoke and acquire video data during the experiment process;

[0084] Early warning module: The early warning module combines the detection results of video data and visible light image data to obtain a heat sink fault result, and determines whether to execute an early warning.

[0085] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned data-driven substation early warning method.

[0086] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned data-driven substation early warning method.

[0087] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned method embodiments. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in this application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0088] In this specification, the same or similar parts among the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the embodiments described later, the description is relatively simple, and the relevant parts can be referred to the partial description of the foregoing embodiments.

[0089] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A data-driven substation early warning method, characterized in that: The method comprises the following steps: S1: Collect visible light image data and infrared image data of transformer heat sinks in substations to construct multi-modal superposition images; S2: Get the heat sink fault result based on multi-modal superposition image detection; S3: Use colored smoke to conduct a heat sink transmittance experiment, collect video data of the experimental process, and obtain video detection results; S4: Combining the video data detection result and the visible light image data detection result to obtain the heat sink fault result, and determining whether to execute the early warning.

2. A data-driven early warning method for substations according to claim 1, characterized in that: During the data processing stage, the collected visible light image data is first optimized through denoising and image enhancement technology to ensure the clarity and accuracy of the image; the infrared image is analyzed for thermal anomalies through a temperature calibration algorithm, and an easy-to-observe temperature distribution map is generated through pseudo-color rendering. Subsequently, the visible light image and infrared image are spatially aligned using image registration technology to generate a multimodal overlay image.

3. A data-driven early warning method for substations according to claim 1, characterized in that: Based on multimodal superposition image detection, a heat sink fault result is obtained, wherein the fault result includes a heat sink blockage caused by accumulation of dust, grease, and foreign matter, which results in air flow obstruction. The fault result is obtained through multimodal attention model detection.

4. The data-driven early warning method for substations according to claim 3 is characterized in that: The multimodal attention model includes a feature extraction backbone network, which is composed of a dual-branch feature extraction module and a feature enhancement module. The dual-branch feature extraction module is used to independently extract visible light feature maps and infrared image feature maps in multimodal superposition images; the dual-branch feature extraction module is composed of the first convolution and the first residual block in two ResNet-18 networks with the same structure.

5. The data-driven early warning method for substations according to claim 4 is characterized in that: The feature enhancement module includes a horizontal stripe extraction module, which is used to extract horizontal stripe information of the heat sink and suppress the horizontal stripe features in the visible light feature map. The horizontal stripe extraction module is defined as: Among them, F vision (x,y) and F ifr (x, y) are the visible light feature map and infrared image feature map obtained by the dual-branch feature extraction module, respectively. filter is the weight function of the filter, and are the horizontal stripe weight maps of the extracted visible light and infrared images, respectively. vetical is a two-dimensional directional gradient kernel, used to extract edge features in the longitudinal direction. is a sinusoidal modulation function used to enhance the filter's response to vertical periodic features, f represents frequency, which indicates the periodic frequency of the filter response, φ is the phase shift, which indicates the adjustment of the starting position of the sinusoidal modulation, and ★ represents the element-by-element dot product; and They are the enhanced visible light feature map and infrared image feature map, To suppress the visible light feature map behind the horizontal stripes, Conv represents convolution calculation, and the subscript represents the convolution kernel size.

6. A data-driven early warning method for substations according to claim 5, characterized in that: The multimodal attention model also includes a feature fusion module, which includes two attention weight generation branches. in, Represent the first and second attention weight maps, respectively, F fused (x,y) represents the output feature map of the feature fusion module.

7. The data-driven early warning method for substations according to claim 1 is characterized in that: Industrial-grade blue smoke is used to detect the leakage of the heat sink, and the collected video is recorded as follows: Among them, I t (x, y) represents the t-th frame image, T is the total number of frames; Perform image enhancement on the captured video to obtain the background-removed difference image D t (x, y): Extract the blue area of ​​the captured video and combine it with the difference image to obtain the blue area mask M t (x,y): M t (x,y)=S t (x,y)·D t (x,y); Among them, S t (x,y) represents the binary mask image obtained according to the color threshold. and They represent the hue, saturation and brightness channels of the t-th frame image after conversion to the HSV space, respectively. threshold and V threshold Respectively represent the minimum saturation requirement and the minimum brightness requirement; The heat sink pass rate and flow rate are calculated based on the blue area mask, and the heat sink blockage situation is obtained.

8. A data-driven substation early warning system, characterized in that: The system is used to execute a data-driven substation early warning method as described in any one of claims 1 to 7, and the system includes: Image acquisition device: The image acquisition device acquires visible light image data and infrared image data of the transformer heat sink in the substation to construct a multi-modal superposition image; Multimodal attention module: The multimodal attention module obtains heat sink fault results based on multimodal superposition image detection; Video data acquisition module: Use colored smoke to conduct heat sink transmittance experiments and collect video data of the experimental process; Early warning module: The early warning module combines the video data detection results and the visible light image data detection to obtain the heat sink failure result and determine whether to execute the early warning.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, a data-driven substation early warning method as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements a data-driven substation early warning method according to any one of claims 1 to 7.

Images