A fault monitoring method and device for a network cabinet

Abstract

The application discloses a kind of failure monitoring method and device for network cabinet, it is related to the technical field of image processing.Acquire the visible light image and infrared image of target area, obtain infrared feature image by feature enhancement to infrared image;Enhance the color image obtained by filtering the visible light image, obtain feature label by feature extraction to enhanced color image by target model;Feature label is added to infrared feature image to obtain final enhanced image, input final enhanced image into fault classification model to obtain fault result;According to fault result, carry out fault early warning.Infrared and visible light image are fused, and the fault heat feature and appearance details are strengthened.Infrared image is highlighted abnormal heating after feature extraction, visible light image is filtered and enhanced texture information, after bimodal fusion, input classification model accurately identifies overheating, damage and other faults, and triggers automatic early warning, improves the accuracy and efficiency of network cabinet monitoring.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to a fault monitoring method and device for network cabinets. Background Technology

[0002] With the rapid development of data centers and communication networks, the stable operation of network cabinets has become crucial to ensuring the reliability of information systems. Traditional temperature fault monitoring mainly relies on single-point detection methods such as infrared thermometers, which suffer from problems such as incomplete coverage and insufficient real-time performance.

[0003] Publication No. CN120558409A discloses a data center power and environment monitoring system based on multi-sensor fusion; an infrared thermal imaging camera for acquiring infrared image information from the sides of server racks within the data center; a temperature analysis unit for calculating the average temperature of each rectangular area on the side of the server rack based on the infrared image information, thereby obtaining the temporal and spatial distribution of rectangular areas in the server rack exhibiting heat dissipation anomalies; and a control center for adjusting the sampling frequency of each temperature sensor based on the temporal and spatial distribution of rectangular areas in the server rack exhibiting heat dissipation anomalies. After the aforementioned server racks have been operating for a period of time, the enclosed indoor environment causes heat accumulation, leading to a gradual decrease in the infrared image features and a deterioration in feature extraction effectiveness.

[0004] In existing network cabinet fault monitoring, the fault identification capability is limited because a single image modality cannot simultaneously capture both the details of the equipment's appearance and the characteristics of its thermal distribution: visible light images cannot effectively detect abnormal internal heating, while infrared images lack semantic information about the equipment's status and texture, affecting the complete extraction and accurate classification of fault features. Summary of the Invention

[0005] The purpose of this invention is to solve the problem mentioned in the background art that a single image modality cannot take into account both appearance details and thermal distribution characteristics, resulting in incomplete fault identification and limited accuracy. Therefore, this invention proposes a fault monitoring method and device for network cabinets.

[0006] A first aspect of this invention provides a fault monitoring method for network cabinets, the method comprising: Acquire visible light and infrared images of the target area, and perform feature enhancement on the infrared image to obtain an infrared feature image; The visible light image is filtered and enhanced to obtain an enhanced color image, and feature labels are obtained by extracting features from the enhanced color image using a target model; The feature labels are added to the infrared feature image to obtain the final enhanced image, and the final enhanced image is input into the fault classification model to obtain the fault result; A fault warning will be issued based on the fault results.

[0007] Optionally, feature enhancement is performed on the infrared image to obtain an infrared feature image, including: Using the infrared image as the root node of a quadtree, decomposition parameters are set, and the brightness of the pixels in the root node is traversed to obtain the brightness range; the decomposition parameters include a range threshold and a minimum size. The root node is decomposed into multiple child nodes according to a preset direction until the brightness range is less than or equal to the range threshold and the node size is less than or equal to the minimum size. Then, the node is taken as the terminal node and the decomposition is stopped. The terminal nodes are combined to obtain a terminal image set. The coordinates of each terminal image in the terminal image set are normalized to obtain a normalized coordinate set. The normalized coordinates in the normalized coordinate set are input into the Bessel basis function to obtain a weight value set. Brightness sampling is performed on the target terminal image in the terminal image set to obtain multiple control points. The control points are combined to obtain a control point matrix. The weight value is multiplied by the corresponding control point matrix to obtain a sub-background image. The sub-background images are stitched together to obtain an infrared background image. Pixel subtraction is performed on the infrared image and the infrared background image to obtain an infrared feature image.

[0008] Optionally, the visible light image is filtered and enhanced to obtain an enhanced color image, including: A preset sliding window is obtained, and the visible light image is slid through the preset sliding window to obtain a window set. The pixel variance and pixel mean are calculated based on the target window in the window set. The target window is any one of the windows in the window set. A linear coefficient is calculated based on the pixel variance, a bias term is calculated based on the pixel mean, and a filtered output is obtained by calculating each pixel in the target window based on the linear coefficient and the bias term. The filtered outputs are then combined to obtain a smoothed image. The coordinate brightness is calculated based on the smoothed image, the global logarithmic average brightness is calculated based on the coordinate brightness, the maximum brightness of the entire image is determined based on the smoothed image, and a brightness adjustment image is calculated based on the coordinate brightness, the global logarithmic average brightness, and the global maximum brightness. The pixels in the brightness adjustment image are normalized to obtain an enhanced color image; the coordinate brightness is the brightness of any pixel in the smoothed image.

[0009] Optionally, feature labels are obtained by extracting features from the enhanced color image using a target model, wherein the target model is based on improvements to YOLOv10 and includes: The target model is obtained by replacing the neck structure of the YOLOv10 model with an improved neck structure. The workflow for improving the neck structure includes: The output of the C2f module at layer 2 in the backbone network is used as the first input feature of the improved neck structure, the output of the C2f module at layer 4 in the backbone network is used as the second input feature of the improved neck structure, the output of the C2f module at layer 6 in the backbone network is used as the third input feature of the improved neck structure, and the output of the PSA module at layer 10 in the backbone network is used as the fourth input feature of the improved neck structure. The first input feature is input into the Conv module to obtain the first tensor. The third and fourth input features are input into the multi-scale downsampling module to obtain the second tensor. The second tensor is input into the C2f module to obtain the third tensor. The third tensor and the second input feature are input into the multi-scale downsampling module to obtain the fourth tensor. The fourth tensor is input into the C2f module to obtain the fifth tensor. The first and fifth tensors are input into the BiFPN module to obtain the sixth tensor. The 6th tensor is input into the C2f module to obtain the 7th tensor. The 7th tensor is input into the Conv module to obtain the 8th tensor. The 8th tensor and the 3rd tensor are input into the BiFPN module to obtain the 9th tensor. The 9th tensor is input into the C2f module to obtain the 10th tensor. The 10th tensor is input into the SCDown module to obtain the 11th tensor. The 11th tensor is concatenated with the 4th input feature to obtain the 12th tensor. The 12th tensor is input into the C2fCIB module to obtain the 13th tensor. The 7th tensor, the 10th tensor, and the 13th tensor are used as the output of the improved neck structure.

[0010] Optionally, the working principle of the multi-scale downsampling module includes: Obtain high-level features and low-level features, downsample the high-level features to obtain the first feature, input the first feature into the EMA module to obtain the second feature, and downsample the second feature to obtain the third feature; The low-level features are input into the Conv module to obtain the fourth feature. The third feature and the fourth feature are fused to obtain the fifth feature. The fifth feature and the third feature are concatenated to obtain the sixth feature. The sixth feature is used as the output of the multi-scale downsampling module.

[0011] A second aspect of the present invention provides a fault monitoring device for a network cabinet, the device comprising: Infrared enhancement module, used to acquire visible light image and infrared image of target area, and perform feature enhancement on the infrared image to obtain infrared feature image; A color enhancement module is used to filter and enhance the visible light image to obtain an enhanced color image, and to extract feature labels from the enhanced color image using a target model. The feature fusion module is used to add the feature labels to the infrared feature image to obtain the final enhanced image, and input the final enhanced image into the fault classification model to obtain the fault result; The fault warning module is used to provide fault warnings based on the fault results.

[0012] Optionally, the infrared enhancement module includes: The root node determination module is used to take the infrared image as the root node of a quadtree, set decomposition parameters, and traverse the brightness of the pixels in the root node to obtain the brightness range; the decomposition parameters include a range threshold and a minimum size. The node decomposition module is used to decompose the root node into multiple child nodes according to a preset direction until the brightness range is less than or equal to the range threshold and the node size is less than or equal to the minimum size. Then, the node is taken as the terminal node and the decomposition is stopped. The normalization module is used to combine each terminal node to obtain a terminal image set, normalize the coordinates of each terminal image in the terminal image set to obtain a normalized coordinate set, and input each normalized coordinate in the normalized coordinate set into the Bessel basis function to obtain a weight value set. The feature extraction module is used to sample the brightness of the target terminal image in the terminal image set to obtain multiple control points, combine the control points to obtain a control point matrix, multiply the weight value with the corresponding control point matrix to obtain a sub-background image, stitch the sub-background images to obtain an infrared background image, and perform pixel subtraction on the infrared image and the infrared background image to obtain an infrared feature image.

[0013] Optionally, the color enhancement module includes: A window sliding module is used to obtain a preset sliding window, slide the visible light image through the preset sliding window to obtain a window set, and calculate the pixel variance and pixel mean based on the target window in the window set; the target window is any one of the window sets. The image filtering module is used to calculate linear coefficients based on the pixel variance, calculate bias terms based on the pixel mean, calculate filtered outputs for each pixel in the target window based on the linear coefficients and the bias terms, and combine the filtered outputs to obtain a smoothed image. An image enhancement module is used to calculate coordinate brightness based on the smoothed image, calculate global logarithmic average brightness based on the coordinate brightness, determine the maximum brightness of the entire image based on the smoothed image, calculate a brightness adjustment image based on the coordinate brightness, the global logarithmic average brightness, and the global maximum brightness, and normalize each pixel in the brightness adjustment image to obtain an enhanced color image; the coordinate brightness is the brightness of any pixel in the smoothed image.

[0014] Optionally, feature labels are obtained by extracting features from the enhanced color image using a target model, wherein the target model is based on improvements to YOLOv10 and includes: The target model is obtained by replacing the neck structure of the YOLOv10 model with an improved neck structure. The workflow for improving the neck structure includes: The output of the C2f module at layer 2 in the backbone network is used as the first input feature of the improved neck structure, the output of the C2f module at layer 4 in the backbone network is used as the second input feature of the improved neck structure, the output of the C2f module at layer 6 in the backbone network is used as the third input feature of the improved neck structure, and the output of the PSA module at layer 10 in the backbone network is used as the fourth input feature of the improved neck structure. The first input feature is input into the Conv module to obtain the first tensor. The third and fourth input features are input into the multi-scale downsampling module to obtain the second tensor. The second tensor is input into the C2f module to obtain the third tensor. The third tensor and the second input feature are input into the multi-scale downsampling module to obtain the fourth tensor. The fourth tensor is input into the C2f module to obtain the fifth tensor. The first and fifth tensors are input into the BiFPN module to obtain the sixth tensor. The 6th tensor is input into the C2f module to obtain the 7th tensor. The 7th tensor is input into the Conv module to obtain the 8th tensor. The 8th tensor and the 3rd tensor are input into the BiFPN module to obtain the 9th tensor. The 9th tensor is input into the C2f module to obtain the 10th tensor. The 10th tensor is input into the SCDown module to obtain the 11th tensor. The 11th tensor is concatenated with the 4th input feature to obtain the 12th tensor. The 12th tensor is input into the C2fCIB module to obtain the 13th tensor. The 7th tensor, the 10th tensor, and the 13th tensor are used as the output of the improved neck structure.

[0015] Optionally, the working principle of the multi-scale downsampling module includes: Obtain high-level features and low-level features, downsample the high-level features to obtain the first feature, input the first feature into the EMA module to obtain the second feature, and downsample the second feature to obtain the third feature; The low-level features are input into the Conv module to obtain the fourth feature. The third feature and the fourth feature are fused to obtain the fifth feature. The fifth feature and the third feature are concatenated to obtain the sixth feature. The sixth feature is used as the output of the multi-scale downsampling module.

[0016] The beneficial effects of this invention are: This invention proposes a fault monitoring method for network cabinets. By acquiring, enhancing, and fusing visible light and infrared images in a dual-modal manner, it achieves multi-dimensional and accurate perception of the status of equipment within the cabinet: the infrared feature image effectively captures abnormal heat points, while the enhanced visible light image clearly identifies the appearance of the equipment, the status of indicator lights, and connection details. The final enhanced image formed after feature alignment and fusion of the two images enables intelligent identification of typical faults such as overheating, fan failure, and port abnormalities through a fault classification model, and triggers an automated early warning system. This improves the automation level and early detection capability of cabinet fault monitoring, reduces the cost of manual inspection, and enhances the reliability and security of data center or communication equipment room operations. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a fault monitoring method for a network cabinet, provided as an embodiment of the present invention; Figure 2 A schematic diagram of a model structure for a fault monitoring method for network cabinets provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a fault monitoring device for a network cabinet provided in an embodiment of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] This invention provides a fault monitoring method for network cabinets. See also... Figure 1 , Figure 1 A flowchart illustrating a fault monitoring method for a network cabinet, provided as an embodiment of the present invention. The method includes the following steps: S101, acquire visible light and infrared images of the target area, and perform feature enhancement on the infrared image to obtain an infrared feature image; S102, the visible light image is filtered and enhanced to obtain an enhanced color image, and feature labels are obtained by feature extraction from the enhanced color image through the target model; S103, add feature labels to the infrared feature image to obtain the final enhanced image, and input the final enhanced image into the fault classification model to obtain the fault result; S104, provide fault warning based on fault results.

[0020] This invention provides a fault monitoring method for network cabinets. By fusing infrared and visible light images, the thermal characteristics and appearance details of faults are enhanced. Infrared images are used to extract features that highlight abnormal heating, while visible light images are filtered to enhance texture information. The dual-modal fusion is then input into a classification model to accurately identify faults such as overheating and damage, triggering automatic early warnings and improving the accuracy and efficiency of network cabinet monitoring.

[0021] In one implementation, multi-dimensional feature complementarity enhancement of the target area is achieved by fusing visible light and infrared dual-modal image information. Feature extraction from the infrared image effectively highlights heat distribution and abnormal heating characteristics, while filtering and enhancement of the visible light image optimizes texture and color details. The combination of the two provides more comprehensive and higher-quality input data for subsequent analysis, improving the perception capability and robustness of potential fault features in complex scenes.

[0022] In one implementation, visible light image feature labels extracted based on the target model are fused with infrared feature images, achieving structured alignment and enhancement of semantic information and thermal features. This cross-modal feature fusion mechanism not only enriches the semantic context of the infrared image but also makes fault-related regions more prominent in the feature space, providing more concentrated and discriminative input for subsequent fault classification, thereby improving the expressive power and analytical accuracy of fault features.

[0023] In one implementation, the fault classification model can be a deep convolutional neural network (CNN) based model, such as ResNet, EfficientNet, or a lightweight MobileNet. Alternatively, a customized network structure supporting multi-scale feature input can be used to adapt to the fused visible light texture and infrared thermal features in the final enhanced image. Training process: The model is trained under supervision using a visible light-infrared fused image dataset labeled with fault categories. The cross-entropy loss between the predicted results and the true labels is calculated through forward propagation, and the model parameters are optimized using backpropagation. The parameters updated during training mainly include the weights and biases of the convolutional layers, the scaling and translation parameters of the batch normalization layers, and the weight matrix and bias terms of the fully connected classification layers. The entire process iteratively optimizes the model to gradually learn the discriminative patterns of the fault region in multimodal features.

[0024] In one implementation, when the fault classification model outputs a specific fault type (such as overheating, line abnormality, equipment failure, etc.), the system will trigger a corresponding early warning mechanism based on the fault level. This includes displaying highlighted alarm information on the monitoring interface, providing on-site prompts through audio-visual devices, automatically generating work orders and pushing them to the operation and maintenance management system, and sending SMS or email notifications to relevant responsible persons according to preset rules.

[0025] In one embodiment, feature enhancement of an infrared image to obtain an infrared feature image includes: Using the infrared image as the root node of a quadtree, decomposition parameters are set, and the brightness of the pixels in the root node is traversed to obtain the brightness range; the decomposition parameters include the range threshold and the minimum size. The root node is decomposed into multiple child nodes according to a preset direction until the brightness range is less than or equal to the range threshold and the node size is less than or equal to the minimum size. Then, the node is taken as the terminal node and the decomposition is stopped. The terminal nodes are combined to obtain a terminal image set. The coordinates of each terminal image in the terminal image set are normalized to obtain a normalized coordinate set. The normalized coordinates in the normalized coordinate set are input into the Bessel basis function to obtain a weight value set. Brightness sampling is performed on the target terminal image in the terminal image set to obtain multiple control points. These control points are combined to form a control point matrix. The weight values ​​are multiplied by the corresponding control point matrix to obtain a sub-background image. These sub-background images are then stitched together to obtain an infrared background image. Pixel subtraction is performed between the infrared image and the infrared background image to obtain an infrared feature image. In one implementation, image initialization involves using the infrared image as the root node of a quadtree and defining the spatial coordinate range of the image (x, y, y). min ~x max y min ~y max ); Set decomposition parameters: Threshold T (brightness range threshold): Set the maximum allowable brightness difference within a node (e.g., T=20, which can be adjusted according to the noise level of the infrared image). When the brightness range (maximum brightness - minimum brightness) within a node > T, it is determined that decomposition can continue; Minimum size Z (node ​​termination condition): Set the minimum pixel size of the node (e.g., Z=4×4 or 8×8). When the node size reaches Z, decomposition stops regardless of whether the brightness range exceeds the limit; Calculate the current node brightness range: For the currently processed node (initially the root node), traverse all pixels within it and calculate the maximum brightness value I. max and minimum value I min The brightness range ΔI=I is obtained. max -I minDetermine whether to decompose: If ΔI > T and node size > Z: Divide the current node into 4 child nodes along the horizontal and vertical directions, with each child node having a size of 1 / 2 of the original node; If ΔI ≤ T or node size ≤ Z: Mark the current node as a terminal node and stop decomposition; For all newly generated child nodes, repeat the brightness range calculation and decomposition judgment in steps 1-2 until all nodes become terminal nodes; Traverse the complete quadtree structure, collect the image regions corresponding to all terminal nodes, and form a terminal image set (each segmented image is an independent image sub-region).

[0026] In one implementation, terminal nodes are combined to obtain a terminal image set. The coordinates of each terminal image in the terminal image set are normalized to obtain a normalized coordinate set. The normalized coordinates in the normalized coordinate set are input into the Bessel basis function to obtain a weight value set. Since the size of each terminal image may be different (e.g., some are 8×8 pixels, some are 16×16 pixels), directly using intra-block coordinates for calculation would be confusing. Therefore, the coordinates of each block are scaled to between 0 and 1 (normalization). For example, if the x-direction range of the image is x minimum = 1 and x maximum = 16, and the x-coordinate of a certain point is 8, then the normalized u... λ =(8-1) / (16-1)=7 / 15≈0.47; Similarly, vᵧ can be obtained in the y direction.

[0027] In one implementation, brightness sampling is performed on the target terminal image block in the terminal image block set to obtain multiple control points. Within the k-th terminal image, 16 control points are sampled according to a uniform grid rule (e.g., a 4×4 grid distribution), and the coordinates (x, y, y) of each control point are recorded. p ,y p ) and the corresponding original infrared brightness value I p By multiplying the weight value by the control point matrix, we can obtain the pure background brightness of this pixel.

[0028] In one implementation, an infrared image is adaptively segmented using a quadtree structure, combined with background generation via Bezier surface fitting, effectively separating background and target features in the infrared image. This method dynamically adjusts the decomposition granularity based on local brightness differences, using larger nodes in smooth background regions to improve computational efficiency, and performing finer segmentation in regions rich in detail or with drastic brightness changes to preserve key information. Subsequently, normalized coordinates and Bezier basis functions are used to perform smooth surface fitting on each terminal node, generating a continuous and natural infrared background estimate. Finally, infrared feature images are accurately extracted through differential operations.

[0029] In one embodiment, filtering and enhancing a visible light image to obtain an enhanced color image includes: Obtain a preset sliding window, slide the visible light image through the preset sliding window to obtain a window set, and calculate the pixel variance and pixel mean based on the target window in the window set; the target window is any one of the windows in the window set. The linear coefficients are calculated based on the pixel variance, the bias term is calculated based on the pixel mean, and the filtered output is obtained by calculating each pixel in the target window based on the linear coefficients and the bias term. The filtered outputs are then combined to obtain a smoothed image. The coordinate brightness is calculated from the smoothed image, the global logarithmic average brightness is calculated from the coordinate brightness, the maximum brightness of the entire image is determined from the smoothed image, and the brightness adjustment image is calculated from the coordinate brightness, the global logarithmic average brightness, and the global maximum brightness. The enhanced color image is obtained by normalizing each pixel in the brightness adjustment image; the coordinate brightness is the brightness of any pixel in the smoothed image.

[0030] In one implementation, the preset sliding window can be set to a size of 3×3 or 5×5. The filtered output is calculated individually for each pixel within the window. The pixel variance and pixel mean are calculated based on the target window in the window set. Each window contains multiple pixels; the variance (brightness variance) and mean (brightness average) of the brightness of each pixel within each window are calculated. The linear coefficient is calculated as follows: , where a x σ is a linear coefficient. x The value is luminance variance, and ε is a smoothing regularization parameter with a value of 0.01~0.1 to avoid a denominator of 0; the bias term is calculated as follows: , where b x As a bias term, a x is the linear coefficient, and is the average brightness value.

[0031] In one implementation, the coordinate brightness is calculated based on the smoothed image, which is the pixel (x,y) in the smoothed image. A weighted brightness is then calculated for that pixel, for example: L(x,y) = 0.299 × T. R +0.587×T G +0.114×T B T R For the red channel, T G For the green channel, T B The blue channel is used; the global logarithmic average brightness is calculated based on the coordinate brightness. The formula for calculating the global logarithmic average brightness is as follows: ,in, The global logarithmic average brightness is given by δ, which is a very small constant (with a value of 1e-6 to avoid numerical overflow when taking the logarithm of pure black pixels), avg[・] is the arithmetic mean operator, and exp[・] is the natural exponential function. The maximum brightness of the entire image is determined based on the smoothed image, that is, the pixel with the largest value in the entire image is taken as the maximum brightness of the entire image.

[0032] In one implementation, pixel variance and mean are calculated region by region using a preset sliding window. Based on this, linear coefficients and bias terms are dynamically generated, achieving adaptive local filtering. This method can distinguish between flat regions and detail edge regions in an image: in flat regions with low variance, the filtered output relies more on the mean of surrounding pixels, smoothing noise; in edge regions with high variance, more original pixel information is preserved, avoiding detail blurring. Finally, a smoothed image is obtained by combining the filtered outputs of each window, maintaining the structural integrity of the image while reducing noise, providing a higher-quality base image for subsequent brightness adjustments.

[0033] In one implementation, coordinate brightness, global logarithmic average brightness, and maximum brightness across the entire image are calculated based on a smoothed image. Brightness adjustments and normalization are then performed accordingly, enhancing the image's visibility and dynamic range under different lighting conditions. By combining local brightness characteristics with global statistics, it can adaptively improve dark details, suppress overexposed areas, and improve overall contrast and color performance. The final output enhanced color image maintains a natural look while increasing visual information, making it suitable for image enhancement applications in unevenly lit or low-light scenarios.

[0034] In one embodiment, feature labels are obtained by extracting features from an enhanced color image using a target model. Improvements to the target model based on YOLOv10 include: The target model is obtained by replacing the neck structure of the YOLOv10 model with an improved neck structure. The workflow for improving neck structure includes: The output of the C2f module at layer 2 in the backbone network is used as the first input feature for improving the neck structure; the output of the C2f module at layer 4 in the backbone network is used as the second input feature for improving the neck structure; the output of the C2f module at layer 6 in the backbone network is used as the third input feature for improving the neck structure; and the output of the PSA module at layer 10 in the backbone network is used as the fourth input feature for improving the neck structure. The first input feature is input into the Conv module to obtain the first tensor. The third and fourth input features are input into the multi-scale downsampling module to obtain the second tensor. The second tensor is input into the C2f module to obtain the third tensor. The third tensor and the second input feature are input into the multi-scale downsampling module to obtain the fourth tensor. The fourth tensor is input into the C2f module to obtain the fifth tensor. The first and fifth tensors are input into the BiFPN module to obtain the sixth tensor. The 6th tensor is input into the C2f module to obtain the 7th tensor. The 7th tensor is input into the Conv module to obtain the 8th tensor. The 8th and 3rd tensors are input into the BiFPN module to obtain the 9th tensor. The 9th tensor is input into the C2f module to obtain the 10th tensor. The 10th tensor is input into the SCDown module to obtain the 11th tensor. The 11th tensor is concatenated with the 4th input feature to obtain the 12th tensor. The 12th tensor is input into the C2fCIB module to obtain the 13th tensor. Tensors 7, 10, and 13 are used as the output for the improved neck structure.

[0035] In one implementation, see [link to implementation details]. Figure 2 , Figure 2 This is a schematic diagram of the model structure of a fault monitoring method for network cabinets provided in an embodiment of the present invention. Input 1 is the first input feature, input 2 is the second input feature, input 3 is the third input feature, and input 4 is the fourth input feature. The detection head is the core component of the target detection model, responsible for receiving the feature map output by the feature extraction network. Through convolution, anchor box mechanism, or key point prediction, it directly outputs the target's category confidence, bounding box coordinates, and key point positions, etc. By using features at different levels in the backbone network (outputs of C2f and PSA modules) as multi-scale inputs to improve the neck structure, it achieves comprehensive capture and fusion of target detail information and semantic information. Utilizing the high spatial resolution of shallow features and the rich semantics of deep features, it provides an information foundation for subsequent multi-scale feature interaction and enhancement, improves the model's ability to perceive targets of different sizes in the image, and improves the feature extraction effect for small and blurry targets.

[0036] In one implementation, a multi-scale downsampling module and a BiFPN module are introduced to perform cross-level feature fusion and adaptive weight adjustment, enhancing the efficiency of feature integration and information flow in the neck structure. The multi-scale downsampling module enables flexible fusion of shallow and deep features, while the BiFPN module, through bidirectional cross-scale connections and learnable feature weights, achieves efficient aggregation and rebalancing of multi-scale features, strengthening the model's ability to express target features at different scales, reducing information loss during transmission, and improving the quality and discriminative power of feature maps.

[0037] In one implementation, the SCDown module performs efficient downsampling and concatenates it with deep features, followed by deep feature extraction and interaction via the C2fCIB module, ultimately outputting three enhanced feature maps at different scales. This structure further optimizes and stabilizes the feature pyramid, expanding the receptive field, preserving key spatial information, and enhancing the robustness and discriminative power of the features. The three output feature maps correspond to the detection of large, medium, and small targets, respectively, providing subsequent detection heads with hierarchical and information-rich feature representations, thereby improving the overall accuracy and generalization ability of the target detection model in different scenarios.

[0038] In one embodiment, the working principle of the multi-scale downsampling module includes: Obtain high-level and low-level features, downsample the high-level features to obtain the first feature, input the first feature into the EMA module to obtain the second feature, and downsample the second feature to obtain the third feature; The low-level features are input into the Conv module to obtain the fourth feature. The third and fourth features are fused to obtain the fifth feature. The fifth feature is concatenated with the third feature to obtain the sixth feature. The sixth feature is used as the output of the multi-scale downsampling module.

[0039] In one implementation, a hierarchical extraction and enhancement of high-level semantic features is effectively achieved through two downsampling operations combined with an EMA attention mechanism. The first downsampling extracts the coarse-grained spatial structure of the high-level features, and the EMA module further focuses on their key channels and spatial regions, suppressing redundant information. The second downsampling completes resolution matching while preserving the semantic core. This process not only enhances the representativeness of high-level features but also maintains the information density of features during the downsampling process through attention guidance, providing high-quality high-level input for subsequent cross-level fusion.

[0040] In one implementation, low-level features are adjusted through convolution and then fused and concatenated with the processed high-level features, achieving effective complementarity and dual enhancement of semantic and detailed information. The fusion operation allows low-level detailed features to receive high-level semantic guidance, improving their discriminative power; while the concatenation operation preserves the independence of the original high-level features, forming a multi-scale composite feature output that combines rich details and deep semantics.

[0041] Based on the same inventive concept, this invention also provides a fault monitoring device for network cabinets. See also Figure 2 , Figure 2 A schematic diagram of a fault monitoring device for a network cabinet provided in an embodiment of the present invention includes: The infrared enhancement module is used to acquire visible light and infrared images of the target area, and to perform feature enhancement on the infrared image to obtain an infrared feature image; The color enhancement module is used to filter and enhance visible light images to obtain enhanced color images, and to extract features from the enhanced color images using a target model to obtain feature labels. The feature fusion module is used to add feature labels to the infrared feature image to obtain the final enhanced image, and input the final enhanced image into the fault classification model to obtain the fault result; The fault warning module is used to provide fault warnings based on the fault results.

[0042] The fault monitoring device for network cabinets provided in this invention enhances the thermal characteristics and appearance details of faults by fusing infrared and visible light images. Infrared images are used to extract features that highlight abnormal heating, while visible light images are filtered to enhance texture information. The dual-modal fusion is then input into a classification model to accurately identify faults such as overheating and damage, triggering automatic warnings and improving the accuracy and efficiency of network cabinet monitoring.

[0043] In one embodiment, the infrared enhancement module includes: The root node determination module is used to take the infrared image as the root node of the quadtree, set decomposition parameters, and traverse the brightness of the pixels in the root node to obtain the brightness range; the decomposition parameters include the range threshold and the minimum size. The node decomposition module is used to decompose the root node into multiple child nodes according to a preset direction until the brightness range is less than or equal to the range threshold and the node size is less than or equal to the minimum size. Then, the node is taken as the terminal node and the decomposition stops. The normalization module is used to combine the terminal nodes to obtain a terminal image set, normalize the coordinates of each terminal image in the terminal image set to obtain a normalized coordinate set, and input the normalized coordinates in the normalized coordinate set into the Bessel basis function to obtain a weight value set. The feature extraction module is used to sample the brightness of the target terminal image in the terminal image set to obtain multiple control points, combine the control points to obtain a control point matrix, multiply the weight value with the corresponding control point matrix to obtain a sub-background image, stitch the sub-background images to obtain an infrared background image, and perform pixel subtraction on the infrared image and the infrared background image to obtain an infrared feature image.

[0044] In one embodiment, the color enhancement module includes: The window sliding module is used to obtain a preset sliding window, slide the visible light image through the preset sliding window to obtain a window set, and calculate the pixel variance and pixel mean based on the target window in the window set; the target window is any one of the windows in the window set. The image filtering module is used to calculate linear coefficients based on pixel variance, calculate bias terms based on pixel mean, calculate filtered outputs for each pixel in the target window based on linear coefficients and bias terms, and combine the filtered outputs to obtain a smoothed image. The image enhancement module is used to calculate the coordinate brightness based on the smoothed image, calculate the global logarithmic average brightness based on the coordinate brightness, determine the maximum brightness of the entire image based on the smoothed image, calculate the brightness adjustment image based on the coordinate brightness, the global logarithmic average brightness, and the global maximum brightness, and normalize each pixel in the brightness adjustment image to obtain the enhanced color image; the coordinate brightness is the brightness of any pixel in the smoothed image.

[0045] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should still fall within the scope of the claims of the present invention.

Claims

1. A fault monitoring method for network cabinets, characterized in that, The method includes: Acquire visible light and infrared images of the target area, and perform feature enhancement on the infrared image to obtain an infrared feature image; The visible light image is filtered and enhanced to obtain an enhanced color image, and feature labels are obtained by extracting features from the enhanced color image using a target model; The feature labels are added to the infrared feature image to obtain the final enhanced image, and the final enhanced image is input into the fault classification model to obtain the fault result; A fault warning will be issued based on the fault results.

2. The fault monitoring method for network cabinets according to claim 1, characterized in that, The infrared image is enhanced to obtain an infrared feature image, including: Using the infrared image as the root node of a quadtree, decomposition parameters are set, and the brightness of the pixels in the root node is traversed to obtain the brightness range; the decomposition parameters include a range threshold and a minimum size. The root node is decomposed into multiple child nodes according to a preset direction until the brightness range is less than or equal to the range threshold and the node size is less than or equal to the minimum size. Then, the node is taken as the terminal node and the decomposition is stopped. The terminal nodes are combined to obtain a terminal image set. The coordinates of each terminal image in the terminal image set are normalized to obtain a normalized coordinate set. The normalized coordinates in the normalized coordinate set are input into the Bessel basis function to obtain a weight value set. Brightness sampling is performed on the target terminal image in the terminal image set to obtain multiple control points. The control points are combined to obtain a control point matrix. The weight value is multiplied by the corresponding control point matrix to obtain a sub-background image. The sub-background images are stitched together to obtain an infrared background image. Pixel subtraction is performed on the infrared image and the infrared background image to obtain an infrared feature image.

3. The fault monitoring method for network cabinets according to claim 1, characterized in that, The process of filtering and enhancing the visible light image to obtain an enhanced color image includes: A preset sliding window is obtained, and the visible light image is slid through the preset sliding window to obtain a window set. The pixel variance and pixel mean are calculated based on the target window in the window set. The target window is any one of the windows in the window set. A linear coefficient is calculated based on the pixel variance, a bias term is calculated based on the pixel mean, and a filtered output is obtained by calculating each pixel in the target window based on the linear coefficient and the bias term. The filtered outputs are then combined to obtain a smoothed image. The coordinate brightness is calculated based on the smoothed image, the global logarithmic average brightness is calculated based on the coordinate brightness, the maximum brightness of the entire image is determined based on the smoothed image, and a brightness adjustment image is calculated based on the coordinate brightness, the global logarithmic average brightness, and the global maximum brightness. The pixels in the brightness adjustment image are normalized to obtain an enhanced color image; the coordinate brightness is the brightness of any pixel in the smoothed image.

4. The fault monitoring method for network cabinets according to claim 1, characterized in that, Feature labels are obtained by extracting features from the enhanced color image using a target model. The target model, based on improvements to YOLOv10, includes: The target model is obtained by replacing the neck structure of the YOLOv10 model with an improved neck structure. The workflow for improving the neck structure includes: The output of the C2f module at layer 2 in the backbone network is used as the first input feature of the improved neck structure, the output of the C2f module at layer 4 in the backbone network is used as the second input feature of the improved neck structure, the output of the C2f module at layer 6 in the backbone network is used as the third input feature of the improved neck structure, and the output of the PSA module at layer 10 in the backbone network is used as the fourth input feature of the improved neck structure. The first input feature is input into the Conv module to obtain the first tensor. The third and fourth input features are input into the multi-scale downsampling module to obtain the second tensor. The second tensor is input into the C2f module to obtain the third tensor. The third tensor and the second input feature are input into the multi-scale downsampling module to obtain the fourth tensor. The fourth tensor is input into the C2f module to obtain the fifth tensor. The first and fifth tensors are input into the BiFPN module to obtain the sixth tensor. The 6th tensor is input into the C2f module to obtain the 7th tensor. The 7th tensor is input into the Conv module to obtain the 8th tensor. The 8th tensor and the 3rd tensor are input into the BiFPN module to obtain the 9th tensor. The 9th tensor is input into the C2f module to obtain the 10th tensor. The 10th tensor is input into the SCDown module to obtain the 11th tensor. The 11th tensor is concatenated with the 4th input feature to obtain the 12th tensor. The 12th tensor is input into the C2fCIB module to obtain the 13th tensor. The 7th tensor, the 10th tensor, and the 13th tensor are used as the output of the improved neck structure.

5. A fault monitoring method for network cabinets according to claim 4, characterized in that, The working principle of the multi-scale downsampling module includes: Obtain high-level features and low-level features, downsample the high-level features to obtain the first feature, input the first feature into the EMA module to obtain the second feature, and downsample the second feature to obtain the third feature; The low-level features are input into the Conv module to obtain the fourth feature. The third feature and the fourth feature are fused to obtain the fifth feature. The fifth feature and the third feature are concatenated to obtain the sixth feature. The sixth feature is used as the output of the multi-scale downsampling module.

6. A fault monitoring device for network cabinets, characterized in that, The device includes: Infrared enhancement module, used to acquire visible light image and infrared image of target area, and perform feature enhancement on the infrared image to obtain infrared feature image; A color enhancement module is used to filter and enhance the visible light image to obtain an enhanced color image, and to extract feature labels from the enhanced color image using a target model. The feature fusion module is used to add the feature labels to the infrared feature image to obtain the final enhanced image, and input the final enhanced image into the fault classification model to obtain the fault result; The fault warning module is used to provide fault warnings based on the fault results.

7. A fault monitoring device for network cabinets according to claim 6, characterized in that, The infrared enhancement module includes: The root node determination module is used to take the infrared image as the root node of a quadtree, set decomposition parameters, and traverse the brightness of the pixels in the root node to obtain the brightness range; the decomposition parameters include a range threshold and a minimum size. The node decomposition module is used to decompose the root node into multiple child nodes according to a preset direction until the brightness range is less than or equal to the range threshold and the node size is less than or equal to the minimum size. Then, the node is taken as the terminal node and the decomposition is stopped. The normalization module is used to combine each terminal node to obtain a terminal image set, normalize the coordinates of each terminal image in the terminal image set to obtain a normalized coordinate set, and input each normalized coordinate in the normalized coordinate set into the Bessel basis function to obtain a weight value set. The feature extraction module is used to sample the brightness of the target terminal image in the terminal image set to obtain multiple control points, combine the control points to obtain a control point matrix, multiply the weight value with the corresponding control point matrix to obtain a sub-background image, stitch the sub-background images to obtain an infrared background image, and perform pixel subtraction on the infrared image and the infrared background image to obtain an infrared feature image.

8. A fault monitoring device for network cabinets according to claim 6, characterized in that, The color enhancement module includes: A window sliding module is used to obtain a preset sliding window, slide the visible light image through the preset sliding window to obtain a window set, and calculate the pixel variance and pixel mean based on the target window in the window set; the target window is any one of the window sets. The image filtering module is used to calculate linear coefficients based on the pixel variance, calculate bias terms based on the pixel mean, calculate filtered outputs for each pixel in the target window based on the linear coefficients and the bias terms, and combine the filtered outputs to obtain a smoothed image. An image enhancement module is used to calculate coordinate brightness based on the smoothed image, calculate global logarithmic average brightness based on the coordinate brightness, determine the maximum brightness of the entire image based on the smoothed image, calculate a brightness adjustment image based on the coordinate brightness, the global logarithmic average brightness, and the global maximum brightness, and normalize each pixel in the brightness adjustment image to obtain an enhanced color image; the coordinate brightness is the brightness of any pixel in the smoothed image.

9. A fault monitoring device for a network cabinet according to claim 6, characterized in that, Feature labels are obtained by extracting features from the enhanced color image using a target model. The target model, based on improvements to YOLOv10, includes: The target model is obtained by replacing the neck structure of the YOLOv10 model with an improved neck structure. The workflow for improving the neck structure includes: The output of the C2f module at layer 2 in the backbone network is used as the first input feature of the improved neck structure, the output of the C2f module at layer 4 in the backbone network is used as the second input feature of the improved neck structure, the output of the C2f module at layer 6 in the backbone network is used as the third input feature of the improved neck structure, and the output of the PSA module at layer 10 in the backbone network is used as the fourth input feature of the improved neck structure. The first input feature is input into the Conv module to obtain the first tensor. The third and fourth input features are input into the multi-scale downsampling module to obtain the second tensor. The second tensor is input into the C2f module to obtain the third tensor. The third tensor and the second input feature are input into the multi-scale downsampling module to obtain the fourth tensor. The fourth tensor is input into the C2f module to obtain the fifth tensor. The first and fifth tensors are input into the BiFPN module to obtain the sixth tensor. The 6th tensor is input into the C2f module to obtain the 7th tensor. The 7th tensor is input into the Conv module to obtain the 8th tensor. The 8th tensor and the 3rd tensor are input into the BiFPN module to obtain the 9th tensor. The 9th tensor is input into the C2f module to obtain the 10th tensor. The 10th tensor is input into the SCDown module to obtain the 11th tensor. The 11th tensor is concatenated with the 4th input feature to obtain the 12th tensor. The 12th tensor is input into the C2fCIB module to obtain the 13th tensor. The 7th tensor, the 10th tensor, and the 13th tensor are used as the output of the improved neck structure.

10. A fault monitoring device for a network cabinet according to claim 9, characterized in that, The working principle of the multi-scale downsampling module includes: Obtain high-level features and low-level features, downsample the high-level features to obtain the first feature, input the first feature into the EMA module to obtain the second feature, and downsample the second feature to obtain the third feature; The low-level features are input into the Conv module to obtain the fourth feature. The third feature and the fourth feature are fused to obtain the fifth feature. The fifth feature and the third feature are concatenated to obtain the sixth feature. The sixth feature is used as the output of the multi-scale downsampling module.

Images