Machine room inspection safety hazard prompting method and system

Abstract

The application provides a machine room inspection safety hidden danger prompting method and system, and relates to the technical field of artificial intelligence. First, the synchronous data streams of a non-visible light waveband thermal radiation sensing array and a visible light waveband high-definition camera array are collected, target inspection data pairs are extracted and registered, and a thermal radiation and visible light inspection data frame group is obtained. Then, the thermal radiation inspection data frame group is divided into temperature field abnormal zones, the abnormal heat zone boundary contour line set is extracted, the visible light inspection data frame group is subjected to equipment semantic partitioning, and a semantic mask map is generated. Then, the abnormal heat zone boundary contour line set and the semantic mask map are superimposed to generate an equipment and line abnormal heating distribution map. Finally, a hidden danger feature encoder and a hidden danger state transition prediction recurrent network are used to generate a safety hidden danger prompting instruction. The application can accurately locate and predict machine room safety hidden dangers.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method and system for alerting potential safety hazards during computer room inspections. Background Technology

[0002] In the daily operation and management of data centers, traditional inspection methods mainly rely on regular manual checks. This approach is not only inefficient but also easily affected by the experience and subjective factors of the inspectors, making it difficult to detect potential safety hazards in a timely manner. With technological advancements, some data center inspection systems based on single sensors have emerged, such as those using only visible light cameras to monitor the appearance of equipment or those using only temperature sensors to detect temperature changes within the data center.

[0003] However, these single-sensor monitoring methods have significant limitations. Relying solely on visible light cameras can only acquire external information about the equipment, failing to accurately assess potential problems such as internal overheating. While temperature sensors can detect temperature changes, they struggle to pinpoint the exact location of temperature anomalies and the corresponding equipment or cables, making it impossible to precisely locate safety hazards. Furthermore, most existing monitoring systems lack the capability for in-depth analysis and processing of monitoring data, hindering the effective prediction of potential safety hazards and making it difficult to implement preventative measures in advance, thus failing to meet the high security requirements of modern data centers. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, an embodiment of the present invention provides a method for identifying potential safety hazards during computer room inspections, the method comprising: The synchronous data stream jointly output by the non-visible light band thermal radiation sensor array and the visible light band high-definition camera array deployed at key inspection nodes in the computer room is collected under the control of a synchronous trigger clock. Thermal radiation sensor data and visible light image data at the same time stamp are extracted from the synchronous data stream to form a target inspection data pair set. Registration processing based on the spatial calibration parameters of the sensor array is performed on the target inspection data pair set to obtain thermal radiation inspection data frame group and visible light inspection data frame group with spatial position alignment relationship. Temperature field anomaly zoning processing is performed on the thermal radiation inspection data frame group. The thermal radiation change rate of each pixel position in the thermal radiation inspection data frame group is calculated using a preset spatial neighborhood gradient operator. Based on the thermal radiation change rate, the set of abnormal hot zone boundary contour lines with abrupt temperature distribution changes is extracted from the thermal radiation inspection data frame group. The visible light inspection data frame group is subjected to device semantic partitioning processing. A pre-built deep residual semantic segmentation network is called to extract semantic masks for the cabinet equipment area and cable routing area in the visible light inspection data frame group, and generate semantic mask maps of the cabinet equipment area and cable routing area. The set of abnormal hot zone boundary contour lines is spatially intersected and superimposed with the semantic mask of the cabinet equipment area and the semantic mask of the cable routing area to generate an abnormal heat distribution map of the equipment that coincides with the semantic mask of the cabinet equipment area and an abnormal heat distribution map of the lines that coincides with the semantic mask of the cable routing area. The preset hazard feature encoder is invoked to extract compressed features from the abnormal heat distribution map of the equipment and the abnormal heat distribution map of the line, generating a set of abnormal heat compression feature vectors. The set of abnormal heat compression feature vectors is then input into the hazard state transition prediction loop network to generate a hazard state evolution path description sequence with time continuity. Based on the hazard state evolution path description sequence, a safety hazard warning instruction is generated for the target cabinet equipment identifier or the target cable routing section identifier.

[0005] Furthermore, embodiments of the present invention also provide a data center inspection safety hazard alert system, comprising: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned data center inspection safety hazard alert method by executing the machine-executable instructions.

[0006] Based on the above, by collecting the synchronous data stream jointly output by the non-visible light band thermal radiation sensing array and the visible light band high-definition camera array, and extracting the thermal radiation sensing data and visible light image data at the same time stamp to form a target inspection data pair set, the target inspection data pair set is registered to obtain thermal radiation inspection data frame group and visible light inspection data frame group with spatial position alignment relationship. Temperature field anomaly zoning processing is performed on the thermal radiation inspection data frame group, which can accurately extract the boundary contour line set of abnormal hot areas with abrupt temperature distribution changes and promptly detect abnormal temperature areas in the computer room. By performing semantic partitioning on visible light inspection data frames, the system can accurately distinguish between cabinet equipment areas and cable routing areas. It also performs spatial intersection and overlay processing on the abnormal heat zone boundary contour set and semantic mask map to generate abnormal heat distribution maps of equipment and lines. Through the hidden danger feature encoder and the hidden danger state transition prediction recurrent network, it can extract compressed features and predict the state evolution path of abnormal heat conditions, which can detect potential safety hazards in advance and generate targeted safety hazard warning instructions. This significantly improves the efficiency and accuracy of data center inspection and reduces the risk of safety accidents. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the execution flow of the data center inspection safety hazard alert method provided in this embodiment of the invention.

[0008] Figure 2 This is a schematic diagram of exemplary hardware and software components of the computer room inspection safety hazard alert system provided in an embodiment of the present invention. Detailed Implementation

[0009] Figure 1 This is a flowchart illustrating a method for identifying potential safety hazards during computer room inspections, provided in one embodiment of the present invention. A detailed description follows.

[0010] The method for identifying potential safety hazards during computer room inspections provided in this application can be applied to automated inspection scenarios in data center computer rooms, communication base station computer rooms, or power control computer rooms. The computer room contains multiple standard server racks and cable trays connecting each rack. The ceiling of the computer room is equipped with a distributed array of non-visible light band thermal radiation sensors and a visible light band high-definition camera array.

[0011] Step S110: Collect the synchronous data stream jointly output by the non-visible light band thermal radiation sensor array and the visible light band high-definition camera array deployed at key inspection nodes in the computer room under the control of a synchronous trigger clock. Extract the thermal radiation sensor data and visible light image data at the same time stamp from the synchronous data stream to form a target inspection data pair set. Perform registration processing based on the spatial calibration parameters of the sensor array on the target inspection data pair set to obtain thermal radiation inspection data frame group and visible light inspection data frame group with spatial position alignment relationship.

[0012] Step S111: Obtain the physical arrangement spacing parameters of the sensing units of the non-visible light band thermal radiation sensing array and the pixel size parameters of the photosensitive element of the visible light band high-definition camera array, and construct a planar coordinate mapping function for the non-visible light thermal radiation data based on the physical arrangement spacing parameters of the sensing units. The planar coordinate mapping function is used to map the discrete temperature measurement point values ​​collected by the thermal radiation sensing array to a continuous two-dimensional spatial planar grid. Also, construct a perspective projection inverse calculation function for the visible light image data based on the pixel size parameters of the photosensitive element and the lens focal length parameters of the visible light band high-definition camera array. The perspective projection inverse calculation function is used to map the pixel coordinates of the visible light image to the physical coordinate system of the computer room inspection space.

[0013] The non-visible light band thermal radiation sensing array consists of m rows * n columns of sensing units arranged in an equally spaced rectangular grid. The lateral physical spacing between adjacent sensing units is dx, and the longitudinal physical spacing between adjacent sensing units is dy. A planar coordinate mapping function maps the discrete temperature measurement point coordinates (i, j) of the i-th row and j-th column sensing unit to physical planar coordinates (px, py), where px = j * dx and py = i * dy. The photosensitive element pixel size parameter of the visible light band high-definition camera array is the physical size s corresponding to each pixel, and the lens focal length parameter is f. The perspective projection inverse calculation function maps the visible light image pixel coordinates (u, v) to the ray direction in the physical coordinate system of the computer room inspection space. The mapping formula is that the ray direction vector is given by (u - u0) * s / f and (v - v0) * s / f, where (u0, v0) are the coordinates of the principal point of the image.

[0014] Step S112: In the physical coordinate system of the inspection space in the computer room, taking the geometric installation origin of the non-visible light band thermal radiation sensing array as the reference origin, calculate the spatial offset vector parameter of the optical center projection point of the visible light band high-definition camera array relative to the reference origin. Use the spatial offset vector parameter to perform affine transformation translation compensation processing on the physical coordinates of the visible light image output by the perspective projection inverse calculation function to obtain a set of visible light registration physical coordinates that share the same reference origin as the non-visible light thermal radiation data.

[0015] The geometric mounting origin of the non-visible light band thermal radiation sensing array is the mounting position of its upper left corner sensing unit. The spatial offset vector parameters of the optical center projection point of the visible light band high-definition camera array relative to this reference origin are (off_x, off_y, off_z). The visible light image physical coordinates (qx, qy) output by the perspective projection inverse calculation function are translated and compensated to obtain the compensated coordinates (rx, ry), where rx = qx - off_x, ry = qy - off_y. By traversing all pixel coordinates of the visible light image, a set of visible light registration physical coordinates that shares the same reference origin as the non-visible light thermal radiation data is obtained.

[0016] Step S113: Perform grid node density resampling processing on the thermal radiation two-dimensional planar grid output by the planar coordinate mapping function, and perform spatial bidirectional interpolation fitting processing on the resampled thermal radiation two-dimensional planar grid nodes and the corresponding pixel positions in the visible light registration physical coordinate set. During the spatial bidirectional interpolation fitting process, for visible light pixel areas not covered by the thermal radiation two-dimensional planar grid nodes, construct a spatial thermal radiation numerical diffusion surface function based on the thermal radiation values ​​of the thermal radiation nodes in the neighborhood of the visible light pixel area.

[0017] The original resolution of the two-dimensional planar mesh for thermal radiation is m*n, while the resolution of the visible light image is u*v, where u is typically much larger than m and v is much larger than n. A bicubic interpolation method is used to upsample and resample the two-dimensional planar mesh for thermal radiation, matching its node density to the pixel density of the visible light image. For the resampled thermal radiation mesh nodes, an inverse distance-weighted interpolation fitting method is used to calculate the thermal radiation value at each visible light pixel coordinate. For pixel regions not covered by the thermal radiation mesh nodes, a two-dimensional diffusion surface function is constructed using the thermal radiation values ​​of the four nearest thermal radiation mesh nodes in the neighborhood of that visible light pixel region. This two-dimensional diffusion surface function adopts a radial basis function form, using the positions and values ​​of the neighboring thermal radiation mesh nodes as constraints to solve the coefficients of the diffusion surface equation, thereby obtaining the thermal radiation value of that pixel region.

[0018] Step S114: Use the spatial thermal radiation numerical diffusion surface function to perform thermal radiation numerical assignment processing on each coordinate point in the visible light registration physical coordinate set, and generate a dense thermal radiation registration data frame group that is aligned point by point with the visible light image pixels.

[0019] For each coordinate point in the visible light registration physical coordinate set, the pixel region to which it belongs is determined based on the location of that coordinate point. The spatial thermal radiation numerical diffusion surface function corresponding to that pixel region is then called, and the coordinate point position is substituted into the diffusion surface function for calculation to obtain the thermal radiation value at that coordinate point. The thermal radiation values ​​calculated for all coordinate points are organized according to the pixel arrangement order of the original visible light image to generate a dense thermal radiation registration data frame group. The resolution of this dense thermal radiation registration data frame group is exactly the same as the resolution of the original visible light image, achieving pixel-by-pixel alignment between the thermal radiation data and the visible light image.

[0020] Step S115: The dense thermal radiation registration data frame group is used as a thermal radiation inspection data frame group with spatial alignment, and the original visible light image frame whose timestamp matches the dense thermal radiation registration data frame group is used as a visible light inspection data frame group with spatial alignment.

[0021] The dense thermal radiation registration data frame group generated in step S114 is used as the thermal radiation inspection data frame group. At the same time, the original visible light image frame with the same timestamp as the dense thermal radiation registration data frame group is extracted from the original synchronous data stream and used as the visible light inspection data frame group. The two constitute a registration data pair with spatial alignment for use in subsequent processing steps.

[0022] Step S120: Perform temperature field anomaly zoning processing on the thermal radiation inspection data frame group, calculate the thermal radiation change rate of each pixel position in the thermal radiation inspection data frame group using a preset spatial neighborhood gradient operator, and extract the set of abnormal hot zone boundary contour lines with abrupt temperature distribution changes in the thermal radiation inspection data frame group based on the thermal radiation change rate.

[0023] Step S121: Define a rectangular analysis window region in the thermal radiation inspection data frame group, centered on the pixel position to be processed and covering the adjacent peripheral pixel positions. Traverse all adjacent pixel positions within the rectangular analysis window region, calculate the positive and negative differences in the horizontal direction between the thermal radiation value of the center pixel position and the thermal radiation values ​​of each peripheral pixel position, and calculate the positive and negative differences in the vertical direction between the thermal radiation value of the center pixel position and the thermal radiation values ​​of each peripheral pixel position.

[0024] The rectangular analysis window size is set to (2k+1)*(2k+1), where k is the window radius parameter. For each pixel position (x, y) in the thermal radiation inspection data frame group, the rectangular analysis window region is extracted with that position as the center. Traversing all adjacent pixel positions within the window, the positive difference value Δtx_p=T(x+1, y)-T(x, y) and the negative difference value Δtx_n=T(x-1, y)-T(x, y) in the horizontal direction are calculated. The positive difference value Δty_p=T(x, y+1)-T(x, y) and the negative difference value Δty_n=T(x, y-1)-T(x, y) in the vertical direction are also calculated.

[0025] Step S122: Sum the squares of the positive and negative difference values ​​in the horizontal direction and the positive and negative difference values ​​in the vertical direction, and then calculate the square root to obtain the gradient magnitude response parameter that characterizes the degree of thermal radiation change at the pixel location; calculate the arctangent angle value based on the difference components of the positive difference values ​​in the horizontal and vertical directions to obtain the gradient direction angle parameter that characterizes the direction of thermal radiation change at the pixel location.

[0026] Calculate the gradient magnitude response parameter G_mag = sqrt(Δtx_p^2 + Δtx_n^2 + Δty_p^2 + Δty_n^2). Calculate the gradient direction angle parameter G_ang = arctan2(Δty_p, Δtx_p), where arctan2 is the arctangent function in the four quadrants, returning an angle value in the range (-π, π]. A larger gradient magnitude response parameter G_mag indicates a more drastic change in thermal radiation at that pixel location, and the gradient direction angle parameter G_ang represents the direction of the fastest change in thermal radiation.

[0027] Step S123: Traverse all pixel positions in the thermal radiation inspection data frame group to generate a global thermal radiation gradient magnitude distribution map containing the gradient magnitude response parameters corresponding to all pixel positions. In the global thermal radiation gradient magnitude distribution map, retrieve a set of pixel chains where the gradient magnitude response parameters undergo a step change and the gradient direction angle parameter differences between adjacent pixels satisfy the preset direction consistency constraint condition.

[0028] For each pixel location in the thermal radiation inspection data frame group, steps S121 and S122 are repeated to arrange the calculated gradient magnitude response parameter G_mag according to pixel coordinates, generating a global thermal radiation gradient magnitude distribution map. In this distribution map, an edge tracking algorithm is used to retrieve pixel chains: starting from a pixel whose gradient magnitude response parameter exceeds a preset threshold, neighboring pixels are searched along the vertical direction of the gradient direction angle parameter G_ang. If the gradient magnitude response parameter of an adjacent pixel also exceeds the threshold and the difference in gradient direction angle parameters between the two pixels is less than a preset angle difference threshold, then the adjacent pixel is included in the current pixel chain. This search is repeated until no further extension is possible, resulting in a set of pixel chains.

[0029] Step S124: Perform contour tracing processing based on connected component analysis on the pixel chain set, mark the pixel chains with closed ends as candidate closed boundary contours for thermal radiation anomalies, perform thermal radiation mean statistical processing on the internal region surrounded by the candidate closed boundary contours for thermal radiation anomalies and the thermal radiation mean statistical processing on the external background region, and confirm the closed boundary contours whose deviation between the internal thermal radiation mean and the external thermal radiation mean is greater than a preset deviation threshold as the set of boundary contours of abnormal hot areas with abrupt temperature distribution changes.

[0030] For each pixel chain in the pixel chain set, determine whether its first and last pixel positions satisfy the closure condition (the distance between the first and last pixels is less than a preset closure distance threshold). If the closure condition is satisfied, mark the pixel chain as a candidate closed boundary contour for thermal radiation anomalies. Calculate the mean thermal radiation value T_in of all pixels within the inner region enclosed by the closed boundary contour, and simultaneously calculate the mean thermal radiation value T_out of all pixels within the outer buffer region of the closed boundary contour. Calculate the deviation D_dev = (T_in - T_out) / T_out. If the absolute value of D_dev is greater than a preset deviation threshold η, then confirm the closed boundary contour as an anomalous thermal zone boundary contour line with abrupt temperature distribution changes. Collect all confirmed boundary contour lines to form a set of anomalous thermal zone boundary contour lines.

[0031] Step S130: Perform device semantic partitioning processing on the visible light inspection data frame group, call the pre-built deep residual semantic segmentation network to extract semantic masks for the cabinet equipment area and cable routing area in the visible light inspection data frame group, and generate semantic mask map of cabinet equipment area and semantic mask map of cable routing area.

[0032] The specific architecture, training process, and application of the deep residual semantic segmentation network are as follows.

[0033] This deep residual semantic segmentation network employs an encoder-decoder symmetric architecture. The encoder is based on the ResNet-50 backbone network, while the decoder uses a progressive upsampling and skip connection structure. The encoder consists of five stages: the first stage is an initial convolutional downsampling module, containing a 7x7 convolutional layer with a stride of 2 (64 output channels) and a 3x3 max-pooling layer with a stride of 2, reducing the spatial resolution of the input image to 1 / 4 of its original value. The second to fifth stages are residual convolutional unit groups, each consisting of multiple stacked residual blocks. Each residual block contains two 3x3 convolutional layers and an identity mapping connection, where the identity mapping directly adds the input to the convolutional output. The second stage contains 3 residual blocks with 256 output channels; the third stage contains 4 residual blocks with 512 output channels; the fourth stage contains 6 residual blocks with 1024 output channels; and the fifth stage contains 3 residual blocks with 2048 output channels.

[0034] The decoder consists of a dilated spatial pyramid pooling module and an upsampling path. The dilated spatial pyramid pooling module comprises six parallel branches with dilation rates of 1, 2, 4, 8, 12, and 16. Each branch first undergoes 1x1 convolution to reduce the dimensionality to 256 channels, then passes through a dilated convolution with the corresponding dilation rate. Finally, the outputs of all branches are concatenated along the channel dimension and fused using a 1x1 convolution before output. The upsampling path comprises four upsampling stages. Each stage first upsamples the feature map by a factor of 2 using bilinear interpolation, then performs a skip connection (channel concatenation) with the feature map from the corresponding stage in the encoder, followed by feature fusion through two 3x3 convolutional layers. The final upsampling stage outputs a channel count equal to the number of categories (3 categories: rack equipment, cable routing, and background), and outputs a semantic segmentation probability map after softmax activation.

[0035] The training process of this deep residual semantic segmentation network is as follows. The training dataset contains 5000 visible light images of a server room, annotating the equipment areas and cable routing areas of the server racks, with an image resolution of 1024x1024. Data augmentation uses random horizontal flipping, random rotation (-10 degrees to 10 degrees), random scaling (0.8x to 1.2x), and color dithering. The loss function is a weighted sum of cross-entropy loss and Dice loss, where the cross-entropy loss has a weight of 0.6 and the Dice loss has a weight of 0.4. The optimizer uses stochastic gradient descent with momentum, with an initial learning rate of 0.01, a momentum parameter of 0.9, and a weight decay parameter of 0.0005. The learning rate uses a multinomial decay strategy, multiplying by (1-epoch / total epochs)^0.9 after each training epoch. The batch size is set to 8, and the number of training epochs is set to 200. The evaluation metric is the average intersection-union ratio (OCR), and training stops when the OCR on the validation set fails to improve for 20 consecutive epochs.

[0036] When applying the model, the original visible light image frames in the visible light inspection data frame set are preprocessed: scaled to 1024x1024 resolution, and pixel values ​​are normalized to the range of 0 to 1. The preprocessed image is input into the depth residual semantic segmentation network, and after forward propagation, a probability distribution map is output. An argmax operation is performed on the probability distribution map to obtain the predicted category of each pixel, and then semantic mask maps of the cabinet equipment area and cable routing area are extracted through connected component analysis.

[0037] Step S131: Input the original visible light image frame in the visible light inspection data frame group into the initial convolutional downsampling module of the deep residual semantic segmentation network to extract a set of shallow visual feature maps containing edge texture gradient responses. Input the set of shallow visual feature maps into the first residual convolutional unit group of the deep residual semantic segmentation network. Perform nonlinear interactive transformation processing between feature channels on the shallow visual feature maps through a stacked identity mapping connection structure to generate the first deep residual feature map.

[0038] The deep residual semantic segmentation network employs an encoder-decoder architecture. The initial convolutional downsampling module contains a 7x7 convolutional layer with a stride of 2 and a 3x3 max-pooling layer with a stride of 2, reducing the spatial resolution of the input image to one-quarter of its original value while expanding the number of channels to 64. The shallow visual feature map set contains low-level visual features such as edges, textures, and corners. The first residual convolutional unit group consists of multiple stacked residual blocks, each containing two 3x3 convolutional layers and an identity mapping connection (directly adding the input to the output), used to extract deep semantic features without causing the gradient vanishing problem. The first residual feature map is generated after passing through the first residual convolutional unit group.

[0039] Step S132: Input the first deep residual feature map into the dilated spatial pyramid pooling module of the deep residual semantic segmentation network, and perform multi-scale receptive field feature recoding on the first deep residual feature map using parallel dilated convolution kernels with different dilation rate parameters to generate a multi-scale receptive field fusion feature map set. Input the multi-scale receptive field fusion feature map set into the second residual convolution unit group of the deep residual semantic segmentation network, and perform channel dimensionality reduction and spatial detail restoration on the multi-scale receptive field fusion feature map through depth-separable convolution operation to generate the second deep residual feature map.

[0040] The dilated spatial pyramid pooling module comprises six parallel branches, each employing dilated convolutional kernels with different dilation rates: r1, r2, r3, r4, r5, and r6. Each branch outputs a feature map of the same spatial size. The output feature maps from all branches are concatenated along the channel dimension and then fused through a 1x1 convolutional layer to obtain a multi-scale receptive field fused feature map set. The second residual convolutional unit group uses depthwise separable convolution instead of standard convolution, separating spatial and channel convolutions to reduce the number of parameters. The depthwise separable convolution first performs a 3x3 spatial convolution independently on each input channel, then uses a 1x1 convolution to fuse inter-channel information. After passing through the second residual convolutional unit group, a second depthwise residual feature map is generated.

[0041] Step S133: Input the second deep residual feature map into the first semantic prediction branch of the deep residual semantic segmentation network, use a pixel-by-pixel classifier to perform foreground confidence score prediction processing for each spatial location of the second deep residual feature map to belong to the rack equipment category, generate a rack equipment category probability distribution map, perform binarization threshold filtering processing on the rack equipment category probability distribution map, aggregate the pixel locations of rack equipment category foreground confidence scores exceeding the preset semantic extraction threshold into rack equipment connected regions, and generate a rack equipment region semantic mask map.

[0042] The first semantic prediction branch contains a 3x3 convolutional layer and a 1x1 convolutional layer, mapping the second deep residual feature map from the high-dimensional feature space to the category space. The output channel is a probability distribution map representing the total number of categories (including rack equipment, cable routing, background, etc.). For the rack equipment category, its corresponding probability channels are extracted to obtain the rack equipment category probability distribution map. A semantic extraction threshold γ is set, and pixels with probability values ​​greater than γ are marked as foreground, while others are marked as background. Connectivity analysis is performed on the foreground pixels, aggregating adjacent foreground pixels into connected regions to generate a semantic mask map of the rack equipment region.

[0043] Step S134: Input the second deep residual feature map into the second semantic prediction branch of the deep residual semantic segmentation network, use a pixel-by-pixel classifier to perform foreground confidence score prediction processing for each spatial location of the second deep residual feature map belonging to the cable routing category, generate a cable routing category probability distribution map, perform binarization threshold filtering processing on the cable routing category probability distribution map, and aggregate the pixel locations where the foreground confidence score of the cable routing category exceeds the preset semantic extraction threshold into a cable routing connected region, generating a cable routing region semantic mask map.

[0044] The second semantic prediction branch has the same structure as the first semantic prediction branch, but it is used to predict cable routing categories. The probability channels corresponding to the cable routing categories are extracted from the output probability distribution map to obtain the cable routing category probability distribution map. Binarization is performed using the same semantic extraction threshold γ, and connected component analysis is performed on the foreground pixels to generate a semantic mask map of the cable routing region.

[0045] Step S135: Store the semantic mask of the cabinet equipment area and the semantic mask of the cable routing area as the output of the equipment semantic partitioning process.

[0046] The semantic mask map of the cabinet equipment area generated in step S133 and the semantic mask map of the cable routing area generated in step S134 are associated and stored.

[0047] Step S140: Perform spatial intersection and overlay processing on the set of abnormal hot zone boundary contour lines and the semantic mask of the cabinet equipment area and the semantic mask of the cable routing area respectively to generate an abnormal heat distribution map of the equipment that overlaps with the semantic mask of the cabinet equipment area and an abnormal heat distribution map of the lines that overlaps with the semantic mask of the cable routing area.

[0048] Step S141: Obtain a list of coordinates of the internal pixel region enclosed by each closed boundary contour line in the set of abnormal hot zone boundary contour lines.

[0049] For each closed boundary contour line in the set of abnormal hot zone boundary contour lines, a scan line filling algorithm or a polygon filling algorithm is used to generate all pixel coordinates of the internal region enclosed by the closed boundary contour line, forming a list of internal pixel region coordinates.

[0050] Step S142: In the semantic mask map of the cabinet equipment area, traverse each coordinate point in the internal pixel area coordinate list and determine whether the semantic mask pixel value corresponding to the coordinate point in the semantic mask map of the cabinet equipment area is a valid activation value representing the cabinet equipment area.

[0051] For each coordinate point (x, y) in the list of internal pixel region coordinates, find the corresponding mask pixel value mask_dev(x, y) in the semantic mask map of the rack equipment region. If mask_dev(x, y) is equal to the preset rack equipment category code value C_dev (e.g., C_dev=1), then the semantic mask pixel value of that coordinate point represents the effective activation value of the rack equipment region.

[0052] Step S143: If the semantic mask pixel value is an effective activation value representing the cabinet equipment area, then the coordinate point is marked as a candidate pixel point for effective equipment hidden danger, and the thermal radiation value within the closed boundary contour line where the coordinate point is located is retained; and if the semantic mask pixel value is an invalid activation value representing the background area, then the coordinate point is marked as an invalid background interference pixel point, and the thermal radiation value within the closed boundary contour line where the coordinate point is located is set to zero.

[0053] For the coordinates of pixels marked as valid candidate points for potential equipment hazards, extract the corresponding thermal radiation value from the thermal radiation inspection data frame group and retain this value unchanged. For the coordinates of pixels marked as invalid background interference pixels, forcibly set the corresponding thermal radiation value to 0.

[0054] Step S144: Aggregate the position coordinates of all candidate pixels marked as valid equipment potential hazards and their corresponding retained thermal radiation values, construct a filtered thermal radiation data layer that only contains thermal radiation distribution information of the internal area of ​​the cabinet equipment, and perform layer overlay rendering processing on the filtered thermal radiation data layer and the equipment boundary outline of the semantic mask map of the cabinet equipment area to generate a visualized abnormal heat generation distribution map of the equipment with thermal radiation color mapping table filling the internal area of ​​the cabinet equipment outline.

[0055] The coordinates and thermal radiation values ​​of all valid candidate pixels for potential equipment hazards are organized into a sparse data layer. For rack equipment areas not covered by valid candidate pixels, interpolation is used to fill them. This data layer is then overlaid with the equipment boundary outline of the semantic mask map of the rack equipment area. A thermal radiation color mapping table is used to map the thermal radiation values ​​into pseudo-colors, which are then filled into the internal area of ​​the rack equipment outline to generate a visual distribution map of abnormal equipment heating.

[0056] Step S145: In the semantic mask map of the cable routing area, traverse each coordinate point in the coordinate list of the internal pixel area and determine whether the semantic mask pixel value corresponding to the coordinate point in the semantic mask map of the cable routing area is a valid activation value representing the cable routing area.

[0057] For each coordinate point (x, y) in the list of internal pixel region coordinates, find the corresponding mask pixel value mask_cab(x, y) in the semantic mask map of the cable routing region. If mask_cab(x, y) is equal to the preset cable routing category code value C_cab (e.g., C_cab=2), then the semantic mask pixel value of that coordinate point represents the effective activation value of the cable routing region.

[0058] Step S146: If the semantic mask pixel value is an effective activation value representing the cable routing area, then the coordinate point is marked as a candidate pixel for a valid line hazard, and the thermal radiation value within the closed boundary contour line where the coordinate point is located is retained; and if the semantic mask pixel value is an invalid activation value representing the background area, then the coordinate point is marked as an invalid background interference pixel, and the thermal radiation value within the closed boundary contour line where the coordinate point is located is set to zero.

[0059] For the coordinates of pixels marked as valid candidate points for potential line hazards, the corresponding thermal radiation value is extracted from the thermal radiation inspection data frame group and retained unchanged. For the coordinates of pixels marked as invalid background interference pixels, the corresponding thermal radiation value is forcibly set to 0.

[0060] Step S147: Aggregate the position coordinates of all candidate pixels marked as valid line hazards and their corresponding retained thermal radiation values, construct a filtered thermal radiation data layer that only contains thermal radiation distribution information of the cable routing area, and perform layer overlay rendering processing on the filtered thermal radiation data layer and the cable routing boundary contour of the semantic mask of the cable routing area to generate a visualized line abnormal heat distribution map with the thermal radiation color map table filling the inner area of ​​the cable routing contour.

[0061] The coordinates and thermal radiation values ​​of all valid candidate pixels for potential line hazards are organized into a sparse data layer. For cable routing areas not covered by valid candidate pixels, interpolation is used to fill them. This data layer is then overlaid with the cable routing boundary outline of the semantic mask image of the cable routing area. A thermal radiation color mapping table is used to map the thermal radiation values ​​into pseudo-colors, which are then filled into the internal area of ​​the cable routing outline to generate a visualized distribution map of abnormal line heating.

[0062] Step S150: Call the preset hidden danger feature encoder to extract compressed features from the abnormal heat distribution map of the equipment and the abnormal heat distribution map of the line, generate a set of abnormal heat compression feature vectors, and input the set of abnormal heat compression feature vectors into the hidden danger state transition prediction loop network to generate a hidden danger state evolution path description sequence with time continuity. Based on the hidden danger state evolution path description sequence, generate a safety hidden danger prompt instruction for the target cabinet equipment identifier or the target cable routing section identifier.

[0063] The specific architecture, training process, and application method of the hazard feature encoder are as follows.

[0064] The hazard feature encoder employs a dual-branch parallel compression architecture. The first convolutional compression branch consists of four 3x3 convolutional layers with a stride of 2, each followed by batch normalization and a ReLU activation function. The number of output channels for each convolutional layer is 32, 64, 128, and 256, respectively. After four layers of downsampling, the spatial size of the input image is compressed to 1 / 16 of the original image. The first global pooling compression branch performs global average pooling on the feature map output by the first convolutional compression branch, compressing the spatial features of each channel into a scalar, outputting a 256-dimensional global statistical compression vector of equipment hazards. The second and third convolutional compression branches have the same structure as the first branch, but they process the abnormal heat distribution map of the line, outputting a 256-dimensional global statistical compression vector of line hazards.

[0065] The training process for this hazard feature encoder employs a self-supervised contrastive learning approach. The training dataset contains 10,000 images of abnormal heat generation from equipment and 10,000 images of abnormal heat generation from power lines, all derived from historical data center inspection records. Positive samples are augmented versions of different data from the same equipment or power line at the same inspection time, while negative samples are data from different inspection times or different equipment. The contrastive loss function uses NT-Xent loss, with the temperature parameter set to 0.1. The optimizer uses Adam, with an initial learning rate of 0.0001 and a weight decay parameter of 0.0001. The batch size is set to 64, and the number of training epochs is 100.

[0066] When applying the model, the abnormal heating distribution map of the equipment and the abnormal heating distribution map of the line are scaled to 224x224 resolution, and the pixel values ​​are normalized to the range of 0 to 1. The input is forward propagated to the hidden danger feature encoder, and the global statistical compression vector of the equipment hidden danger and the global statistical compression vector of the line hidden danger are output respectively. The two are concatenated to obtain the set of abnormal heating compressed feature vectors.

[0067] The specific architecture, training process, and application of the hazard state transition prediction recurrent network are as follows.

[0068] This recurrent network for predicting hazard state transitions employs a Long Short-Term Memory (LSTM) network architecture, comprising an input gate, a forget gate, an output gate, and a memory unit. The hidden layer dimension is set to 128, and the memory unit dimension is also set to 128. The input gate weight matrix W_i has a dimension of 128 x (128 + input dimension), the forget gate weight matrix W_f has a dimension of 128 x (128 + input dimension), the output gate weight matrix W_o has a dimension of 128 x (128 + input dimension), and the candidate memory unit weight matrix W_c has a dimension of 128 x (128 + input dimension). The input feature dimension is 512 (256 + 256).

[0069] The training process of the recurrent network for predicting the state transition of potential hazards is as follows. The training dataset contains 5000 time-series samples from data center inspections. Each sample contains a sequence of compressed feature vectors of abnormal heating at 20 consecutive inspection times, along with a corresponding sequence of hazard state labels (annotated by operations and maintenance experts). The input sequence length is 20, and the output sequence length is 5 (predicting the hazard state at the next 5 times). The mean squared error loss function is used. The optimizer uses Adam, with an initial learning rate of 0.001, a batch size of 32, and 150 training epochs. An early stopping mechanism is employed, stopping training when the validation set loss does not decrease for 15 consecutive epochs.

[0070] When applying the model, the set of compressed feature vectors of abnormal heating at the current inspection time and the previous 19 time steps is organized into an input sequence (length 20) in chronological order and input into the hazard state transition prediction recurrent network. At each time step, the network outputs the hidden layer vector of the hazard state at the current time step, and outputs the hidden layer vectors of the hazard state at the most recent 5 time steps as a sequence describing the evolution path of the hazard state.

[0071] Step S151: Input the abnormal heat distribution map of the equipment into the first convolutional compression branch of the hidden danger feature encoder. Compress the spatial resolution of the abnormal heat distribution map of the equipment to a preset low-dimensional feature map size through a series of stride convolutional downsampling layers to obtain a low-dimensional spatial feature map of the equipment hidden danger. Input the low-dimensional spatial feature map of the equipment hidden danger into the first global pooling compression branch of the hidden danger feature encoder. Calculate the global average pooling response for each feature channel of the low-dimensional spatial feature map of the equipment hidden danger to obtain the global statistical compression vector of the equipment hidden danger.

[0072] The first convolutional compression branch of the hazard feature encoder consists of multiple 3*3 convolutional layers with a stride of 2 connected in series. With each convolutional layer, the spatial size of the feature map is halved, while the number of channels doubles. After L layers of convolutional downsampling, a low-dimensional spatial feature map of the equipment hazard is obtained, with a spatial size of w*h and the number of channels c. The first global pooling compression branch calculates the global average pooling response for each channel of this feature map: for the c-th channel, its pooling response value is the arithmetic mean of the feature values ​​at all spatial locations in that channel. The pooling response values ​​of all channels are arranged into a vector to obtain the global statistical compression vector v_dev for the equipment hazard.

[0073] Step S152: Input the abnormal heat distribution map of the line into the second convolutional compression branch of the hidden danger feature encoder. Compress the spatial resolution of the abnormal heat distribution map of the line to a preset low-dimensional feature map size through a series of stride convolutional downsampling layers to obtain a low-dimensional spatial feature map of the line hidden danger. Input the low-dimensional spatial feature map of the line hidden danger into the second global pooling compression branch of the hidden danger feature encoder. Calculate the global average pooling response for each feature channel of the low-dimensional spatial feature map of the line hidden danger to obtain the global statistical compression vector of the line hidden danger.

[0074] The second convolutional compression branch has the same structure as the first convolutional compression branch, but it processes the abnormal heating distribution map of the line. After the same downsampling operation, a low-dimensional spatial feature map of line hazards is obtained. The second global pooling compression branch calculates the global average pooling response for each channel of this feature map to obtain the global statistical compression vector v_cab of line hazards.

[0075] Step S153: The global statistical compression vector of the equipment hidden danger and the global statistical compression vector of the line hidden danger are concatenated and spliced ​​in the feature dimension direction to generate an abnormal heating compressed feature vector set representing the overall thermal hidden danger state of the computer room at the current inspection time. The abnormal heating compressed feature vector set is input into the input gate structure of the hidden danger state transition prediction recurrent network. The abnormal heating compressed feature vector set at the current time is linearly transformed using the input gate weight matrix. It is then weighted and fused with the hidden layer state vector of the hidden danger state transition prediction recurrent network at the previous time to generate the candidate memory unit state vector at the current time.

[0076] The global statistical compression vectors of equipment hazards (v_dev) and line hazards (v_cab) are concatenated along the feature dimension to obtain the set of abnormal heating compressed feature vectors x_t=[v_dev, v_cab]. The recurrent network for hazard state transition prediction adopts a long short-term memory network structure. The input gate structure calculates the candidate memory unit state vector c'_t=tanh(W_c*[h_{t-1}, x_t]+b_c), where W_c is the weight matrix of the candidate memory unit, h_{t-1} is the hidden layer state vector of the previous time step, and b_c is the bias term.

[0077] Step S154: Input the set of abnormal heating compressed feature vectors and the hidden layer state vector of the hidden danger state transition prediction recurrent network at the previous moment into the forget gate structure of the hidden danger state transition prediction recurrent network, calculate the forget gate coefficient vector representing the degree of retention of historical hidden danger state information, and use the forget gate coefficient vector to perform element-wise product update processing on the long-term memory unit state vector at the previous moment.

[0078] The forget gate structure calculates the forget gate coefficient vector f_t = sigmoid(W_f * [h_{t-1}, x_t] + b_f), where W_f is the weight matrix of the forget gate, and sigmoid is the sigmoid activation function with an output value between 0 and 1. The forget gate coefficient vector f_t is then element-wise multiplied with the previous long-term memory unit state vector c_{t-1} to obtain the updated long-term memory unit state vector c'_t = f_t * c_{t-1}.

[0079] Step S155: The candidate memory unit state vector at the current moment is added element by element to the long-term memory unit state vector updated by the forgetting gating coefficient vector to generate the long-term memory unit state vector at the current moment.

[0080] The current long-term memory unit state vector c_t = c'_t + i_t * c'_t, where i_t is the input gating coefficient vector, i_t = sigmoid(W_i * [h_{t-1}, x_t] + b_i), and W_i is the weight matrix of the input gate.

[0081] Step S156: Input the abnormal heating compressed feature vector set, the hidden layer state vector of the previous time-stage hidden danger state transition prediction recurrent network, and the current time-stage long-term memory unit state vector into the output gate structure of the hidden danger state transition prediction recurrent network, calculate the output gating coefficient vector, and perform element-wise multiplication of the output gating coefficient vector with the current time-stage long-term memory unit state vector processed by the activation function to generate the hidden layer vector of the hidden danger state at the current time.

[0082] The output gate structure calculates the output gating coefficient vector o_t = sigmoid(W_o * [h_{t-1}, x_t] + b_o), where W_o is the weight matrix of the output gate. The hidden layer vector of the current hidden danger state is h_t = o_t * tanh(c_t).

[0083] Step S157: Repeat the forward propagation process of the hidden danger state transition prediction loop network at multiple consecutive inspection times, and form a hidden danger state evolution path description sequence by arranging the hidden danger state hidden layer vectors output each time in chronological order. Analyze the monotonic change trend direction and slope of the vector values ​​in the hidden danger state evolution path description sequence, and generate a warning level identification code and suggested handling action code for the target cabinet equipment identification or target cable routing section identification based on the monotonic change trend direction and slope.

[0084] Steps S151 to S156 are repeated for T consecutive inspection times to obtain the hidden danger status vector sequence h_1, h_2, ..., h_T. For each dimension component in the sequence, its monotonic trend direction (increasing or decreasing) and trend slope (the sum of the differences between adjacent times) are calculated. Based on the trend direction and slope magnitude, they are mapped to the warning level identifier code (e.g., normal, attention, warning, and emergency levels) and the suggested action code (e.g., inspection, maintenance, replacement, etc.).

[0085] Step S158: Encapsulate the warning level identifier code and the suggested action code into a safety hazard warning instruction.

[0086] The warning level identifier code and suggested action code generated in step S157 are encapsulated according to the preset instruction format to generate a security hazard warning instruction, which is then sent to the data center operation and maintenance management terminal equipment.

[0087] Step S210: Obtain the historical thermal radiation inspection data record sequence output by the non-visible light band thermal radiation sensor array in the target inspection data set under the synchronous trigger clock control, perform periodic baseline drift modeling processing on the historical thermal radiation inspection data record sequence, and use the long-term sliding window statistical algorithm to calculate the mean parameter and variance parameter of thermal radiation value of each sensing unit in the historical thermal radiation inspection data record sequence within a fixed time period.

[0088] For each sensing unit (i, j) in the non-visible light band thermal radiation sensing array, its historical thermal radiation value record sequence R(i, j, t) over the past L inspection cycles is obtained, where t is the inspection time index. A long-term sliding window statistical algorithm is used, with a window width of W inspection cycles. Within each window, the arithmetic mean μ_ij = (1 / W)*ΣR(i, j, t) and variance σ_ij^2 = (1 / W)*Σ(R(i, j, t) - μ_ij)^2 of the thermal radiation values ​​are calculated. The window slides forward with a step size S, obtaining the mean parameter sequence and variance parameter sequence for each sensing unit at different time periods.

[0089] Step S220: Construct a non-parametric thermal radiation baseline distribution function for each sensing unit based on the mean parameter and variance parameter of the thermal radiation value. The non-parametric thermal radiation baseline distribution function is used to describe the allowable range of thermal radiation value fluctuation of the sensing unit under normal operating conditions.

[0090] For each sensing unit (i, j), a non-parametric thermal radiation baseline distribution function f_ij(r) = (1 / (W*h))*ΣK((rR(i, j, t)) / h) is constructed using the kernel density estimation method, where K is the Gaussian kernel function and h is the bandwidth parameter. Based on this distribution function, the confidence interval [Q_α / 2, Q_1-α / 2] is calculated at a given confidence level α, where Q_α / 2 is the α / 2 quantile and Q_1-α / 2 is the 1-α / 2 quantile. This confidence interval represents the allowable range of thermal radiation numerical fluctuations of the sensing unit under normal operating conditions.

[0091] Step S230: When acquiring the real-time thermal radiation inspection data frame group, substitute the real-time thermal radiation value corresponding to each sensing unit in the real-time thermal radiation inspection data frame group into the non-parametric thermal radiation reference baseline distribution function for deviation evaluation processing, calculate the standard deviation multiple of the real-time thermal radiation value relative to the mean parameter of the thermal radiation value, and if the standard deviation multiple exceeds the preset baseline deviation multiple threshold, mark the position coordinate of the sensing unit as the reference deviation abnormal sensing point.

[0092] For each sensor unit (i, j) in the real-time thermal radiation inspection data frame group, obtain its real-time thermal radiation value R_real(i, j). Calculate the standard deviation multiple z_ij = (R_real(i, j) - μ_ij) / σ_ij. Set a baseline deviation multiple threshold Z_th. If |z_ij| > Z_th, then mark the position coordinates (i, j) of that sensor unit as the reference deviation abnormal sensing point.

[0093] Step S240: Form a local thermal field anomaly verification region by all sensing units in the spatial neighborhood of the reference deviation anomaly sensing point, and search within the local thermal field anomaly verification region for whether there are continuously distributed sensing units that simultaneously meet the trigger condition that the standard deviation multiple exceeds the preset baseline deviation multiple threshold.

[0094] Centered on each reference deviation anomaly sensing point, a neighborhood radius r_neighbor is defined. All sensing units with a Manhattan or Euclidean distance less than r_neighbor are included in the local thermal field anomaly verification region. Within this region, a four-connected or eight-connected region growing algorithm is used to search for whether there are continuously distributed sensing units that simultaneously satisfy the condition |z_ij|>Z_th. If such a continuously distributed region exists, its boundary is recorded.

[0095] Step S250: If there are continuously distributed sensing units within the local thermal field anomaly verification area that simultaneously meet the triggering condition, then the geometric outer contour of the local thermal field anomaly verification area is added to the set of abnormal thermal field boundary contours as a supplementary abnormal thermal zone boundary.

[0096] For the continuous distribution region identified in step S240, its minimum convex hull or minimum bounding rectangle is calculated as the geometric bounding contour. This geometric bounding contour is then added to the set of abnormal hot zone boundary contours generated in step S120 as a supplementary abnormal hot zone boundary.

[0097] Step S260: Based on the supplemented set of abnormal hot zone boundary contours, re-execute the spatial intersection and overlay processing with the semantic mask map of the cabinet equipment area and the semantic mask map of the cable routing area to update the abnormal heat distribution map of the equipment and the abnormal heat distribution map of the line.

[0098] Using the supplemented set of abnormal hot zone boundary contours as new input, repeat the spatial intersection overlay processing in step S140 to regenerate the updated equipment abnormal heat distribution map and the updated line abnormal heat distribution map.

[0099] Step S270: Regenerate the abnormal heat distribution vector set using the updated equipment abnormal heat distribution map and the line abnormal heat distribution map, and update the hidden danger state evolution path description sequence. Based on the updated hidden danger state evolution path description sequence, dynamically correct the warning level identifier code in the safety hazard warning instruction.

[0100] The updated equipment abnormal heat distribution map and the updated line abnormal heat distribution map are input into the hazard feature encoder and the hazard state transition prediction recurrent network. Following steps S151 to S157, the abnormal heat compression feature vector set is regenerated, and the hazard state evolution path description sequence is updated. Based on the updated hazard state evolution path description sequence, the warning level identifier code is recalculated, and the newly calculated warning level identifier code replaces the original warning level identifier code in the safety hazard warning instruction, thus achieving dynamic correction of the warning level.

[0101] Step S310: Obtain the sequence of historical visible light inspection data frames collected by the visible light band high-definition camera array in multiple consecutive synchronous trigger clock cycles. The sequence of historical visible light inspection data frames is the cumulative set of visible light inspection data frames with spatial alignment in multiple consecutive synchronous trigger clock cycles.

[0102] Read visible light inspection data frame groups collected in the past P synchronous trigger clock cycles from the data storage system, arrange them in the order of timestamps to form the historical visible light inspection data frame group sequence F_vis(1), F_vis(2), ..., F_vis(P).

[0103] Step S320: Perform a three-channel separation operation of hue, saturation and brightness based on color space transformation on each visible light inspection data frame group in the historical visible light inspection data frame group sequence to extract the hue channel image frame sequence that reflects the aging state of the coating on the device surface.

[0104] The visible light image in each frame of visible light inspection data is converted from the RGB color space to the HSV color space. The conversion formulas are: V=max(R,G,B), S=(V-min(R,G,B)) / V (if V≠0), H=60*(GB) / (V-min(R,G,B)) (if V=R), H=60*(2+(BR) / (V-min(R,G,B))) (if V=G), H=60*(4+(RG) / (V-min(R,G,B))) (if V=B). The tone channel H is extracted to obtain the tone channel image frame sequence H_seq(1), H_seq(2), ..., H_seq(P).

[0105] Step S330: Perform hue offset cumulative analysis processing based on pixel value difference operation on the hue channel image frame sequence. Take the first hue channel image in the historical visible light inspection data frame group sequence as the reference hue frame, and calculate the absolute offset parameter of each pixel position in each subsequent hue channel image relative to the reference hue frame in terms of hue value.

[0106] The first frame's tone channel image H_seq(1) is used as the reference tone frame H_ref. For the k-th frame (k>1) of the tone channel image, the absolute offset parameter ΔH_k(x,y) = |H_seq(k)(x,y) - H_ref(x,y)| is calculated for each pixel position (x,y). The unit of the absolute offset parameter is tone angle (a value between 0 and 360 degrees).

[0107] Step S340: Perform linear regression fitting on the absolute offset parameters calculated from multiple consecutive frames along the time axis to obtain the tone offset rate slope parameter and tone offset fitting confidence parameter at each pixel position.

[0108] For each pixel location (x, y), its absolute offset parameter ΔH_k(x, y) at multiple time points is used as the dependent variable, and the corresponding timestamp t_k is used as the independent variable. A linear regression is performed using the least squares method to obtain the regression equation ΔH = a_xy*t + b_xy. Here, a_xy is the hue shift rate slope parameter, representing the rate at which the hue at that pixel location changes over time. Simultaneously, the regression determination coefficient R^2_xy is calculated as the confidence parameter for the hue shift fit.

[0109] Step S350: In the current visible light image frame corresponding to the visible light inspection data frame group, extract the tone offset rate slope parameter corresponding to all pixel positions that coincide with the semantic mask map of the cabinet equipment area, and mark the pixel positions where the tone offset rate slope parameter exceeds the preset coating aging rate threshold as candidate pixels for abnormal coating aging.

[0110] Obtain the semantic mask M_dev corresponding to the rack equipment area of ​​the current visible light image frame. For all pixel positions in the mask image with a value of C_dev (representing the rack equipment area), extract the tone shift rate slope parameter a_xy corresponding to that position. Set the coating aging rate threshold A_th. If a_xy > A_th, then mark the pixel position as a candidate pixel for abnormal coating aging.

[0111] Step S360: Perform region aggregation processing based on spatial connectivity on the candidate pixels of abnormal coating aging, and aggregate the candidate pixels of abnormal coating aging that are spatially adjacent and meet the preset connected neighborhood distance condition into a connected region block of abnormal coating aging.

[0112] An eight-connected region growing algorithm is used to aggregate candidate pixels for abnormal coating aging: starting from any unvisited candidate pixel, all candidate pixels within its eight neighborhoods are recursively added to the same region block until no new candidate pixels are added. Each region block forms a connected region block for abnormal coating aging.

[0113] Step S370: Calculate the mean of the area parameter and the slope parameter of the color shift rate within each of the abnormal aging connected regions of the coating, and determine the abnormal aging connected regions of the coating that have the area parameter exceeding the preset minimum alarm area threshold and the mean value exceeding the preset severe aging rate threshold as severe aging alarm regions of the coating.

[0114] For each connected region R exhibiting abnormal coating aging, calculate its area Area_R (total number of pixels within the region). Calculate the arithmetic mean μ_a_R of the tone shift rate slope parameter a_xy for all pixels within the region. Set a minimum alarm area threshold A_min and a severe aging rate threshold A_severe. If Area_R > A_min and μ_a_R > A_severe, then the region R is identified as a severe coating aging alarm region.

[0115] Step S380: Obtain the geometric outer polygon outline coordinate sequence of the severely aged coating alarm area, and map the geometric outer polygon outline coordinate sequence to the corresponding spatial position of the abnormal heat distribution map of the device. In the abnormal heat distribution map of the device, determine whether there is a situation where the distribution density of effective device hidden danger candidate pixels exceeds the preset hidden danger distribution density threshold within the area covered by the geometric outer polygon outline coordinate sequence.

[0116] For each severely aged coating alarm area, calculate its geometric outer polygon contour coordinate sequence (convex hull vertex coordinates). Map this contour to the same spatial coordinate system as the abnormal heat distribution map of the equipment. In the abnormal heat distribution map of the equipment, count the number of valid candidate pixels for potential equipment hazards, N_hot, within the area covered by this contour, and calculate the distribution density ρ_hot = N_hot / Area_R. Set a hazard distribution density threshold ρ_th; if ρ_hot > ρ_th, the condition is met.

[0117] Step S390: If the distribution density of valid candidate pixels for equipment hazards exceeds a preset hazard distribution density threshold within the area covered by the geometric outer polygon contour coordinate sequence, a multi-hazard associated alarm identifier representing the coupling of coating aging and thermal hazards is generated. The multi-hazard associated alarm identifier, the geometric outer polygon contour coordinate sequence of the severe coating aging alarm area, and the information of the abnormal heat generation distribution area of ​​the equipment that coincides with the location of the severe coating aging alarm area are encapsulated together into a coating thermal coupling hazard supplementary data segment. The coating thermal coupling hazard supplementary data segment is appended to the end field of the data payload of the safety hazard warning instruction.

[0118] If the condition of step S380 is satisfied, generate a device multiple hazard association warning identifier flag_couple (e.g., set to 1). Package this identifier, the geometric外包 polygon contour coordinate sequence Poly_coating of the severely aged coating warning area, and the information on the abnormal heat generation distribution area (statistical characteristics of the thermal radiation numerical distribution) that coincides with the position of Poly_coating into a coating thermal coupling hazard supplementary data segment. Append this supplementary data segment to the end field of the data payload of the safety hazard prompt instruction generated in step S158.

[0119] Step S410: Obtain the sequence of original thermal radiation numerical matrices collected by the non-visible light band thermal radiation sensing array within a continuous plurality of synchronous trigger clock cycles, and perform thermal diffusion direction field calculation processing based on the neighborhood difference operator on each frame of the original thermal radiation numerical matrix in the sequence of original thermal radiation numerical matrices to generate a set of thermal diffusion direction vector field maps corresponding to each frame of the original thermal radiation numerical matrix.

[0120] For each frame of the original thermal radiation numerical matrix T(x, y), calculate the horizontal direction gradient G_x and the vertical direction gradient G_y at each pixel position using the Sobel operator. The horizontal convolution kernel of the Sobel operator is [[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]], and the vertical convolution kernel is [[-1, -2, -1], [0, 0, 0], [1, 2, 1]]. The thermal diffusion direction vector is defined as the unit vector in the negative gradient direction, i.e., (-G_x / pow(G_x^2 + G_y^2, 0.5), -G_y / pow(G_x^2 + G_y^2, 0.5)). Organize the direction vectors at all pixel positions into a two-channel vector field map of the same size as the thermal radiation numerical matrix to form a set of thermal diffusion direction vector field maps.

[0121] Step S420: Perform heat source convergence analysis processing based on the vector field divergence operator on each frame of the thermal diffusion direction vector field map in the set of thermal diffusion direction vector field maps, calculate the two-dimensional spatial divergence value of the thermal diffusion direction vector at each pixel position, and mark the pixel positions where the two-dimensional spatial divergence value is less than the preset negative divergence threshold as pixel points in the heat flow convergence candidate area.

[0122] For the thermal diffusion direction vector field map F(x, y) = (u(x, y), v(x, y)), calculate the two-dimensional spatial divergence value div(F) = ∂u / ∂x + ∂v / ∂y. Use the central difference method to calculate the partial derivatives: ∂u / ∂x ≈ (u(x + 1, y) - u(x - 1, y)) / 2, ∂v / ∂y ≈ (v(x, y + 1) - v(x, y - 1)) / 2. Set the negative divergence threshold D_neg (negative value). If div(F) < D_neg, then mark this pixel position as a pixel point in the heat flow convergence candidate area.

[0123] Step S430: Statistically process the cumulative number of times the same pixel position in the set of heat diffusion direction vector field maps in multiple consecutive frames is marked as a candidate region for heat flow convergence, generate a cumulative distribution map of heat flow convergence frequency that characterizes the persistence of heat flow convergence, extract connected pixel regions in the cumulative distribution map of heat flow convergence frequency that exceed a preset persistence frequency threshold, and use the geometric outline of the connected pixel regions as the set of boundary outlines of the core candidate region of the hidden heat source.

[0124] For each pixel location, the cumulative number of times it is marked as a candidate region for heat flow convergence in the consecutive Q-frame heat diffusion direction vector field map is counted, freq(x, y). A cumulative distribution map of heat flow convergence frequency is generated. A persistence frequency threshold F_th is set, and pixel locations where freq(x, y) > F_th are extracted and aggregated into connected regions using an eight-connected region growing algorithm. The convex hull of each connected region is calculated as the geometric outer contour line, resulting in a set of boundary contour lines for the core candidate regions of potential heat sources.

[0125] Step S440: Perform thermal radiation temporal stability assessment processing on the internal region enclosed by each closed boundary contour line in the set of candidate core regions of the potential heat source, and calculate the temporal variation coefficient parameter of thermal radiation values ​​of all sensing units in the internal region within multiple consecutive synchronous trigger clock cycles.

[0126] For the internal region enclosed by each closed boundary contour line, collect the thermal radiation numerical sequences of all sensing units within that region for Q consecutive time periods. Calculate the temporal mean μ_seq and temporal standard deviation σ_seq for each sensing unit, and then calculate the arithmetic mean of the coefficient of variation CV = (σ_seq / μ_seq) for all sensing units within that region.

[0127] Step S450: The internal region corresponding to the closed boundary contour line where the temporal variation coefficient parameter is lower than the preset stability variation threshold is determined as the core region of the stable hidden heat source, and the position coordinates of the geometric center sensing unit of the core region of the stable hidden heat source are used as the coordinates of the source point of the hidden heat source. The coordinates of the source point of the hidden heat source are mapped to the visible light image pixel coordinate system corresponding to the visible light inspection data frame group, and a local visible light image block with a preset spatial size range is cropped with the mapped pixel point of the source point of the hidden heat source in the visible light image as the center.

[0128] Set the stability variation threshold CV_th, and determine the internal area corresponding to the closed boundary contour line with CV < CV_th as the core area of the stable hidden danger heat source. Calculate the geometric center coordinates (x_c, y_c) of this area as the coordinates of the occurrence source point of the hidden danger heat source. Using the perspective projection inverse calculation function and its inverse transformation in step S111, map (x_c, y_c) to the coordinates (u_c, v_c) in the visible light image pixel coordinate system. Centered on (u_c, v_c), crop a local visible light image block with a size of W_crop * H_crop.

[0129] Step S460: Input the local visible light image block into a pre-constructed fine-grained recognition convolutional network for computer room equipment components, and output the category name of the computer room equipment components and the coordinates of the component space bounding box contained in the local visible light image block through the multi-level convolutional feature extraction layer and region proposal generation layer of the fine-grained recognition convolutional network for computer room equipment components.

[0130] The fine-grained recognition convolutional network for computer room equipment components adopts the Faster R-CNN architecture. The multi-level convolutional feature extraction layer is composed of the first several layers of VGG16 or ResNet50, and outputs a feature map. The region proposal generation layer slides anchor boxes on this feature map and outputs candidate region bounding boxes and foreground confidence levels. Subsequently, through the region of interest pooling layer and the classification and regression network, the category name of each component (such as "power module", "fan module", "main board", etc.) and the precise coordinates of the component space bounding box are output.

[0131] Step S470: Generate targeted hidden danger description text information for specific equipment components based on the category name of the computer room equipment components and the coordinates of the component space bounding box, and append the targeted hidden danger description text information to the hidden danger description field of the safety hidden danger prompt instruction.

[0132] Combine the recognized category name of the computer room equipment components and the coordinates of the component space bounding box into a natural language description text in the format of "Stable heat source convergence detected in the [component category name] area (coordinate range [bounding box coordinates])". Append this targeted hidden danger description text information to the hidden danger description field of the safety hidden danger prompt instruction to provide more accurate hidden danger location information.

[0133] For example, the method may further include: Step S510: Obtain the continuous multi-frame visible light image frame data corresponding to the visible light inspection data frame group collected by the visible light band high-definition camera array within a continuous plurality of synchronous trigger clock cycles.

[0134] Read the visible light image frame data collected within the past R synchronous trigger clock cycles from the data storage system to form a continuous multi-frame visible light image frame sequence G(1), G(2),..., G(R).

[0135] Step S520: Perform morphological skeleton extraction processing on the semantic mask map sequence of the cable routing region in the continuous multi-frame visible light image data to obtain the set of center line pixels of the cable skeleton representing the center path of the cable routing.

[0136] For each frame of visible light image data, obtain its corresponding semantic mask map M_cab for the cable routing region. A morphological skeleton extraction algorithm is employed: repeatedly perform erosion operations, then subtract the result from the opening operation, until the image no longer changes. Merge the skeleton points obtained from different iteration levels to obtain the centerline pixel set of the cable skeleton. The pixels in this centerline pixel set constitute the single-pixel width center path of the cable routing.

[0137] Step S530: Perform segmented straight line fitting processing on the pixel set of the center line of the cable skeleton, and divide the pixel set of the center line of the cable skeleton into a set of cable segmented straight line segments with different orientation angles.

[0138] The Douglas-Puk algorithm or iterative endpoint fitting algorithm is used to segment the pixel set of the cable skeleton centerline. For example, the pixel sequence is recursively divided into sub-segments, each of which can be approximated by a straight line segment, such that the maximum distance from the original pixel to the fitted straight line segment is less than a preset tolerance. The final result is a set of cable segmented straight line segments, each with start-point coordinates, end-point coordinates, direction angle, and length parameters.

[0139] Step S540: Calculate the spatial offset pixel distance parameter of the set of straight line segments of the cable segment at the corresponding position in the two-dimensional image plane in the continuous multi-frame visible light image frame data of two adjacent frames.

[0140] For two adjacent frames t and t+1, optical flow matching or nearest neighbor matching algorithms are used to establish the correspondence between the cable segment lines between the two frames. For each pair of matched line segments, the Euclidean distance between their center points is calculated as the spatial offset pixel distance parameter d_pixel.

[0141] Step S550: Based on the six-degree-of-freedom pose transformation parameter sequence of the visible light band high-definition camera array at the corresponding moment, the spatial offset pixel distance parameter in the two-dimensional image plane is converted into the actual displacement offset of the cable in three-dimensional space through the triangulation principle. The six-degree-of-freedom pose transformation parameter sequence is obtained by performing visual real-time localization and map construction processing on the continuous multi-frame visible light image frame data.

[0142] Visual real-time localization and mapping algorithms (such as ORB-SLAM) are used to process multiple consecutive frames of visible light image data to estimate the six-degree-of-freedom pose transformation parameters (rotation matrix R and translation vector T) of the camera for each frame. For the matching line segment center point pair (p_t, p_{t+1}), the three-dimensional spatial coordinates are calculated using the triangulation principle. Then, the actual displacement offset in three-dimensional space is calculated as d_3d = ||P_{t+1} - P_t||, where P_t and P_{t+1} are the coordinates of the three-dimensional points.

[0143] Step S560: Mark the straight segments of the cable segment whose actual displacement exceeds the preset cable deformation displacement threshold as abnormal cable segments.

[0144] Set a cable deformation displacement threshold D_th. If d_3d>D_th, then mark the straight segment of the cable as a cable segment with abnormal deformation.

[0145] Step S570: Extract the start and end pixel coordinate index values ​​of the deformed cable segment in the pixel set of the center line of the cable skeleton, and crop out the local magnified image block of the deformed cable segment from the continuous multi-frame visible light image data according to the start and end pixel coordinate index values.

[0146] For each abnormally deformed cable segment, obtain the coordinates of its starting and ending pixels in the original cable skeleton centerline pixel set. Using the line segment determined by these two coordinate points as the centerline, extend it to both sides by a preset width W_pad, and crop out a magnified image block in the visible light image frame.

[0147] Step S580: Input the magnified image block of the abnormally deformed cable segment into the pre-constructed fine-grained classification convolutional network for cable sheath damage, and extract the microscopic damage feature response map of the cable sheath texture.

[0148] The fine-grained classification convolutional network for cable sheath damage employs an attention-based fine-grained classification network structure, comprising multiple convolutional layers, channel attention modules, and spatial attention modules. After a magnified local image patch is propagated forward through the network, the final convolutional layer outputs a micro-damage feature response map. High-response regions in this micro-damage feature response map correspond to locations where sheath damage may exist.

[0149] Step S590: Determine whether there is a potential for insulation layer rupture in the abnormally deformed cable section based on the microscopic damage characteristic response map. If there is a potential for insulation layer rupture, generate an additional potential physical damage identifier for the cable.

[0150] Global average pooling is performed on the microscopic damage feature response map to obtain the damage probability score p_damage. A damage judgment threshold p_th is set. If p_damage > p_th, it is determined that there is a potential for insulation layer breakage, and an additional potential physical damage identifier flag_cable_damage is generated (for example, set to 1).

[0151] Step S5100: Associate and bind the cable physical damage hazard identifier with the target cable routing section identifier, and append the bound hazard information to the extended data field of the safety hazard warning instruction.

[0152] Associate and bind the physical damage hazard identifier flag_cable_damage with its corresponding target cable routing section identifier cab_id. Then, append the bound hazard information (including cable routing section identifier, spatial location of the abnormally deformed cable section, damage probability score, etc.) to the extended data field of the safety hazard warning instruction.

[0153] Step S610: Obtain the original thermal radiation numerical matrix sequence collected by the non-visible light band thermal radiation sensing array in multiple consecutive synchronous trigger clock cycles, perform thermal diffusion direction field calculation processing based on the neighborhood difference operator on each frame of the original thermal radiation numerical matrix sequence, and generate a set of thermal diffusion direction vector field maps corresponding to each frame of thermal radiation numerical matrix.

[0154] For each frame of the original thermal radiation numerical matrix T(x, y), the Sobel operator is used to calculate the horizontal gradient G_x and the vertical gradient G_y at each pixel location. The horizontal convolution kernel of the Sobel operator is [[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]], and the vertical convolution kernel is [[-1, -2, -1], [0, 0, 0], [1, 2, 1]]. The thermal diffusion direction vector is defined as the unit vector of the negative gradient direction, i.e., (-G_x / pow(G_x^2+G_y^2, 0.5), -G_y / pow(G_x^2+G_y^2, 0.5)). The direction vectors of all pixel locations are organized into a two-channel vector field map of the same size as the thermal radiation numerical matrix, forming the thermal diffusion direction vector field map set F_dir(1), F_dir(2), ..., F_dir(Q).

[0155] Step S620: Perform heat source convergence analysis processing based on vector field divergence operator on each frame of the heat diffusion direction vector field map set, calculate the two-dimensional spatial divergence value of the heat diffusion direction vector at each pixel position, and mark the pixel positions with the two-dimensional spatial divergence value less than the preset negative divergence threshold as candidate pixels for heat flow convergence region.

[0156] For the heat diffusion direction vector field map F_dir(x, y) = (u(x, y), v(x, y)), calculate the two-dimensional spatial divergence div(F) = ∂u / ∂x + ∂v / ∂y. The central difference method is used to calculate the partial derivatives: ∂u / ∂x ≈ (u(x + 1, y) - u(x - 1, y)) / 2, ∂v / ∂y ≈ (v(x, y + 1) - v(x, y - 1)) / 2. Set a negative divergence threshold D_neg (take a negative value). If div(F) < D_neg, it indicates that this position is a heat flow convergence point (heat flow flows to this point from all around), and mark the pixel position as a pixel point in the candidate area of heat flow convergence.

[0157] Step S630: Statistically process the cumulative number of times that the pixel positions in the same pixel position in the set of consecutive multi-frame heat diffusion direction vector field maps are marked as pixel points in the candidate area of heat flow convergence, generate a cumulative distribution map of heat flow convergence frequency characterizing the persistence of heat flow convergence. Extract the connected pixel regions in the cumulative distribution map of heat flow convergence frequency where the heat flow convergence frequency exceeds the preset persistence frequency threshold, and use the geometric outer envelope contour line of the connected pixel region as the set of boundary contour lines of the candidate area of the hidden danger heat source core.

[0158] For each pixel position (x, y), statistically process the cumulative number of times freq(x, y) that it is marked as a pixel point in the candidate area of heat flow convergence in the consecutive Q-frame heat diffusion direction vector field maps. Arrange freq(x, y) according to the pixel coordinates to generate a cumulative distribution map of heat flow convergence frequency. Set a persistence frequency threshold F_th, and extract the pixel positions where freq(x, y) > F_th. Use the eight-connected region growing algorithm to aggregate adjacent pixels that meet the conditions into a connected region. Calculate the convex hull or the minimum bounding rectangle of each connected region as the geometric outer envelope contour line to obtain the set of boundary contour lines H_contour of the candidate area of the hidden danger heat source core.

[0159] Step S640: Perform a thermal radiation temporal stability evaluation process on the internal region surrounded by each closed boundary contour line in the set of boundary contour lines of the candidate area of the hidden danger heat source core, and calculate the temporal variation coefficient parameter of the thermal radiation values of all sensing units in the internal region within consecutive multiple synchronous trigger clock cycles.

[0160] For each closed boundary contour line C in the set of boundary contour lines of the core candidate regions of potential hazard heat sources, all the sensing units within the internal region enclosed by it are extracted. For each sensing unit, a sequence of thermal radiation values R(t1), R(t2), …, R(tQ) at Q consecutive moments is obtained. The mean μ_seq and standard deviation σ_seq of this sequence are calculated. The coefficient of variation CV_i of this sensing unit is CV_i = σ_seq / μ_seq. The arithmetic mean CV_region of the coefficients of variation of all the sensing units within region C is calculated as CV_region = (1 / N) * ΣCV_i, where N is the total number of sensing units within the region.

[0161] Step S650: Determine the internal region corresponding to the closed boundary contour line with the temporal coefficient of variation parameter lower than the preset stability variation threshold as the stable core region of the potential hazard heat source, and use the position coordinates of the sensing unit at the geometric center of the stable core region of the potential hazard heat source as the coordinates of the occurrence source point of the potential hazard heat source. Map the coordinates of the occurrence source point of the potential hazard heat source to the pixel coordinate system of the visible light image corresponding to the visible light inspection data frame group, and crop a local visible light image block with a preset spatial size range centered on the mapped pixel point of the occurrence source point of the potential hazard heat source in the visible light image.

[0162] Set the stability variation threshold CV_th (for example, 0.1), and determine the internal region corresponding to the closed boundary contour line with CV_region < CV_th as the stable core region of the potential hazard heat source, indicating that the thermal radiation anomaly in this region has temporal stability and is not accidental fluctuation. Calculate the geometric center coordinates (x_c, y_c) = ((x_min + x_max) / 2, (y_min + y_max) / 2) of this region as the coordinates of the occurrence source point of the potential hazard heat source. Using the perspective projection transformation relationship established in step S111, map (x_c, y_c) to the coordinates (u_c, v_c) in the pixel coordinate system of the visible light image. Crop a local visible light image block with a size of W_crop * H_crop centered on (u_c, v_c).

[0163] Step S660: Input the local visible light image block into the pre - constructed fine - granularity recognition convolutional network for computer room equipment components. Output the category name of the computer room equipment components and the coordinates of the component spatial bounding box within the local visible light image block through the multi - level convolutional feature extraction layer and region proposal generation layer of the fine - granularity recognition convolutional network for computer room equipment components.

[0164] The fine - granularity recognition convolutional network for computer room equipment components adopts a region - proposal - based convolutional neural network architecture (Faster R - CNN). The multi - level convolutional feature extraction layer is composed of 13 convolutional layers and 4 pooling layers alternately stacked (the first 13 layers of VGG16), and the output spatial size is the original Figure 1A 16 / 16 feature map is generated. The region proposal generation layer generates nine anchor boxes of different scales and aspect ratios centered at each location on this feature map. Binary classification is used to determine whether the anchor boxes contain the target object, and bounding box regression is used to correct the anchor box positions. The region of interest pooling layer pools the candidate region feature maps of different sizes into a fixed size. The classification and regression network consists of two fully connected layers, outputting the component category probability distribution (e.g., "power module," "fan module," "motherboard," "hard drive," etc.) and precise component spatial bounding box coordinates for each candidate region.

[0165] Step S670: Generate targeted hazard description text information for a specific equipment component based on the category name of the equipment component in the computer room and the spatial boundary coordinates of the component, and append the targeted hazard description text information to the hazard description field of the safety hazard warning instruction.

[0166] The identified equipment component category names and component spatial boundary coordinates are combined into a natural language description text, formatted as "A stable heat source convergence was detected in the [component category name] (coordinate range [x1, y1, x2, y2]) area. It is recommended to conduct a special thermal imaging re-inspection of this component." This targeted hazard description text is then appended to the hazard description field of the safety hazard warning instruction.

[0167] Step S710: Obtain the historical visible light inspection data frame sequence acquired by the visible light band high-definition camera array within multiple consecutive synchronous trigger clock cycles, perform a three-channel separation operation of hue, saturation and brightness based on color space transformation processing on each visible light image in the historical visible light inspection data frame sequence, and extract the hue channel image frame sequence that reflects the aging state of the coating on the surface of the device.

[0168] The visible light inspection data frame groups collected within the past P synchronous trigger clock cycles are read from the data storage system and arranged in timestamp order to form a historical visible light inspection data frame group sequence F_vis(1), F_vis(2), ..., F_vis(P). For each visible light image frame, the color space is converted from RGB to HSV. The conversion formula is: V=max(R,G,B), S=(V-min(R,G,B)) / V (if V≠0), H=60*(GB) / (V-min(R,G,B)) (if V=R), H=60*(2+(BR) / (V-min(R,G,B))) (if V=G), H=60*(4+(RG) / (V-min(R,G,B))) (if V=B). Extract the tone channel H (value range 0-360 degrees) to obtain the tone channel image frame sequence H_seq(1), H_seq(2), ..., H_seq(P).

[0169] Step S720: Perform hue offset cumulative analysis processing based on pixel value difference operation on the hue channel image frame sequence. Take the first hue channel image in the historical visible light inspection data frame group sequence as the reference hue frame, and calculate the absolute offset parameter of each pixel position in each subsequent hue channel image relative to the reference hue frame in terms of hue value.

[0170] The first frame's tone channel image H_seq(1) is used as the reference tone frame H_ref. For the k-th frame (k>1) of the tone channel image, the absolute offset parameter ΔH_k(x,y) = |H_seq(k)(x,y) - H_ref(x,y)| is calculated for each pixel position (x,y). The unit of the absolute offset parameter is tone angle (a value between 0 and 360 degrees). The larger the value of ΔH_k(x,y), the more significant the color change at that pixel position.

[0171] Step S730: Perform linear regression fitting on the absolute offset parameters calculated from multiple consecutive frames along the time axis to obtain the tone change slope parameter and tone offset fitting confidence parameter at each pixel position.

[0172] For each pixel location (x, y), its absolute offset parameter ΔH_k(x, y) at multiple time points is used as the dependent variable, and the corresponding timestamp t_k is used as the independent variable. A linear regression is performed using the least squares method. The regression equation is ΔH = a_xy*t + b_xy, where a_xy is the hue change slope parameter, representing the rate (degrees / day) of hue change at that pixel location over time. Simultaneously, the regression determination coefficient R^2_xy = 1 - (SS_res / SS_tot) is calculated, where SS_res is the sum of squared residuals, SS_tot is the total sum of squares, and R^2_xy serves as the confidence parameter for the hue offset fit.

[0173] Step S740: In the current visible light image frame corresponding to the visible light inspection data frame group, extract the tone change slope parameter corresponding to all pixel positions that coincide with the semantic mask map of the cabinet equipment area, and mark the pixel positions where the tone change slope parameter exceeds the preset coating aging rate threshold as candidate pixels for abnormal coating aging.

[0174] Obtain the semantic mask M_dev corresponding to the rack equipment area of ​​the current visible light image frame. For all pixel positions (x, y) with a value of C_dev (representing the rack equipment area) in the mask image, extract the tone change slope parameter a_xy corresponding to that position. Set a coating aging rate threshold A_th (degrees / day). If a_xy > A_th, it indicates that the coating aging rate at that position is abnormal, and the pixel position is marked as a candidate pixel for abnormal coating aging.

[0175] Step S750: Perform region aggregation processing based on spatial connectivity on the candidate pixels of abnormal coating aging, and aggregate the candidate pixels of abnormal coating aging that are spatially adjacent and meet the preset connected neighborhood distance condition into a connected region block of abnormal coating aging.

[0176] An eight-connected region growing algorithm is used to aggregate candidate pixels for abnormal coating aging. Starting from any unvisited candidate pixel, all candidate pixels within its eight neighborhoods are recursively added to the same region block until no new candidate pixels are added. Each region block forms a connected region block for abnormal coating aging. The pixel coordinate set and boundary of each region block are recorded.

[0177] Step S760: Calculate the mean of the area parameter and the slope parameter of the color shift rate within each of the abnormal aging connected regions of the coating, and determine the abnormal aging connected regions of the coating that have the area parameter exceeding the preset minimum alarm area threshold and the mean value exceeding the preset severe aging rate threshold as severe aging alarm regions of the coating.

[0178] For each connected region R exhibiting abnormal coating aging, calculate its area Area_R (total number of pixels within the region). Calculate the arithmetic mean μ_a_R of the tone change slope parameter a_xy for all pixels within the region. Set a minimum alarm area threshold A_min (number of pixels) and a severe aging rate threshold A_severe (degrees / day). If Area_R > A_min and μ_a_R > A_severe, then the region R is identified as a severe coating aging alarm region.

[0179] Step S770: Obtain the geometric outer polygon outline coordinate sequence of the severely aged coating alarm area, and map the geometric outer polygon outline coordinate sequence to the corresponding spatial position of the abnormal heat distribution map of the device. In the abnormal heat distribution map of the device, determine whether there is a situation where the distribution density of effective device hidden danger candidate pixels exceeds the preset hidden danger distribution density threshold within the area covered by the geometric outer polygon outline coordinate sequence.

[0180] For each severely aged coating alarm area, calculate its convex hull as the geometric outer polygon contour coordinate sequence Poly_coating. Map this contour coordinate sequence to the same spatial coordinate system as the abnormal heat distribution map of the equipment. In the abnormal heat distribution map of the equipment, count the number N_hot of valid equipment hazard candidate pixels (i.e., the valid equipment hazard candidate pixels marked in step S143) within the area covered by Poly_coating. Calculate the distribution density ρ_hot = N_hot / Area_R. Set a hazard distribution density threshold ρ_th. If ρ_hot > ρ_th, it indicates that the coating aging area and the heat hazard area highly overlap.

[0181] Step S780: If the distribution density of valid candidate pixels for equipment hazards exceeds the preset hazard distribution density threshold within the area covered by the geometric outer polygon contour coordinate sequence, a multi-hazard association alarm identifier for equipment that characterizes the coupling between coating aging and thermal hazards is generated.

[0182] If the conditions in step S770 are met, a device multiple hazard association alarm identifier flag_couple is generated and set to 1 (indicating the existence of coupled hazards). This device multiple hazard association alarm identifier is used to prompt maintenance personnel that coating aging and thermal anomalies may have a common physical root cause (such as long-term equipment overload causing the shell temperature to rise and accelerate coating aging).

[0183] Step S790: The device's multiple hidden danger associated alarm identifier, the geometric outer polygon contour coordinate sequence of the coating severe aging alarm area, and the device's abnormal heat distribution area information that coincides with the location of the coating severe aging alarm area are encapsulated together into a coating thermal coupling hidden danger supplementary data segment, and the coating thermal coupling hidden danger supplementary data segment is appended to the end field of the data payload of the safety hidden danger prompt instruction.

[0184] The equipment's multiple hidden danger associated alarm identifier flag_couple, the geometric outer polygon contour coordinate sequence Poly_coating of the severe coating aging alarm area, and the information on the abnormal heat distribution area of ​​the equipment coinciding with the Poly_coating location (including statistical characteristics such as the maximum value, average value, and spatial distribution variance of thermal radiation within this area) are encapsulated into a supplementary data segment for coating thermal coupling hazards. This supplementary data segment is appended to the end field of the data payload of the safety hazard warning instruction to form a complete enhanced safety hazard warning instruction.

[0185] In the above embodiments, the convolutional network for fine-grained identification of equipment components in the computer room adopts the Faster R-CNN architecture, which includes a feature extraction backbone network, a region proposal network, a region of interest pooling layer, and a classification and regression network. The feature extraction backbone network uses the first 13 convolutional layers and 4 pooling layers of VGG16, and the spatial size of the output feature map is 1 / 16 of the input image, with 512 channels. The region proposal network generates 9 anchor boxes (3 scales x 3 aspect ratios) centered on each location on the feature map, and outputs the foreground confidence and bounding box regression offset of each anchor box through a 3x3 convolutional layer and a 1x1 convolutional layer. The region of interest pooling layer pools the candidate region feature maps of different sizes into a fixed-size 7x7 feature map. The classification and regression network consists of two fully connected layers (4096 dimensions each) and an output layer, which outputs the component category probability distribution for each candidate region (10 categories in total: power module, fan module, motherboard, hard drive, memory module, expansion card, cable interface, heat sink, capacitor array, and others) and the bounding box regression offset.

[0186] The training process of the convolutional network for fine-grained identification of equipment components in the computer room is as follows. The training dataset contains 20,000 cropped local visible light image patches, each labeled with the equipment component category and bounding box. Data augmentation employs random horizontal flipping, random rotation, and random brightness and contrast adjustments. The loss function includes the classification loss (binary cross-entropy) and regression loss (smooth L1 loss) for the region proposal network, and the classification loss (multi-class cross-entropy) and regression loss (smooth L1 loss) for the classification and regression network. The total loss is the sum of the four losses. The optimizer uses stochastic gradient descent with momentum, with an initial learning rate set to 0.001, momentum parameter set to 0.9, and weight decay parameter set to 0.0005. The batch size is set to 16, and the training epochs are set to 100. A learning rate decay strategy is adopted, multiplying the learning rate by 0.1 at epochs 60 and 80.

[0187] When applying the model, the cropped local visible light image patch is scaled to 800x800 resolution, and the pixel values ​​are normalized to the range of 0 to 1. This patch is then input into a fine-grained convolutional network for identifying equipment components in the computer room. After forward propagation, the network outputs the category name of the detected equipment component and the coordinates of the component's spatial bounding box. Detection results with a confidence score greater than 0.7 are taken as the final output.

[0188] The specific architecture, training process, and application of the fine-grained classification convolutional network for cable sheath damage are as follows.

[0189] This fine-grained classification convolutional network for cable sheath damage employs an attention-based architecture, comprising a feature extraction backbone, a channel attention module, a spatial attention module, and a classification layer. The feature extraction backbone uses ResNet-50, taking the feature map before its global average pooling layer as the output; the feature map dimension is 2048x7x7. The channel attention module performs global average pooling and global max pooling on the feature map, outputting a channel attention weight vector (2048 dimension) after passing through two fully connected layers. This weight is then multiplied channel-wise by the original feature map. The spatial attention module performs channel-wise average pooling and max pooling on the enhanced feature map, concatenating the two pooling results and outputting a spatial attention weight map (7x7 dimension) after passing through a 7x7 convolutional layer and a sigmoid activation function. This weight is then multiplied spatially by the feature map. The classification layer consists of a global average pooling layer and a fully connected layer (output dimension 2, representing lossless and damaged classes).

[0190] The training process of this fine-grained classification convolutional network for cable sheath damage is as follows. The training dataset contains 15,000 magnified images of cable sections, of which 7,500 are labeled as having damaged sheaths and 7,500 are labeled as having intact sheaths. Data augmentation uses random horizontal flipping, random rotation, random scaling, and random Gaussian noise addition. The loss function is binary cross-entropy loss. The optimizer uses Adam, with an initial learning rate of 0.0001 and a weight decay parameter of 0.0001. The batch size is set to 32, and the number of training epochs is set to 80. The evaluation metrics are accuracy, precision, and recall. Training stops when the validation set accuracy fails to improve for 10 consecutive epochs.

[0191] When applying the model, the magnified image patch of the abnormally deformed cable section is scaled to 224x224 resolution, and the pixel values ​​are normalized to the range of 0 to 1. This patch is then input into a fine-grained classification convolutional network for cable sheath damage. After forward propagation, the network outputs a damage probability score, which is used as the global pooling result of the microscopic damage feature response map. If the damage probability score is greater than 0.7, a potential insulation layer breakage is identified.

[0192] Based on the same inventive concept, please refer to Figure 2 This diagram illustrates a schematic block diagram of the computer room inspection safety hazard alert system provided in an embodiment of this application. It includes a central processing unit (CPU), a system memory comprising random access memory (RAM) and read-only memory (ROM), and a system bus connecting the system memory and the CPU. The computer room inspection safety hazard alert system also includes a basic input / output system to facilitate information transmission between various devices within the computer, and a large-capacity storage device for storing the operating system, applications, and other program modules.

[0193] A basic input / output system includes a display for showing information and input devices such as a mouse and keyboard for user input. Both the display and the input devices are connected to the central processing unit via an input / output controller connected to the system bus. The basic input / output system may also include an input / output controller for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input / output controller also provides output to a display screen, printer, or other types of output devices.

[0194] A mass storage device is connected to the central processing unit via a mass storage controller connected to the system bus. The mass storage device and its associated computer-readable medium provide non-volatile storage for the data center inspection security hazard alert. Without loss of generality, computer-readable media can include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. According to various embodiments of this application, the data center inspection security hazard alert can also be connected to a remote computer on a network, such as the Internet. That is, the data center inspection security hazard alert can be connected to a network via a network interface unit connected to the system bus, or it can be used to connect to other types of networks or remote computer systems.

[0195] In addition, in the specific embodiments of this application, data such as user information are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0196] The above are merely exemplary embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application shall be included within the protection scope of this application.

Claims

1. A method for identifying potential safety hazards during computer room inspections, characterized in that, The method includes: The synchronous data stream jointly output by the non-visible light band thermal radiation sensor array and the visible light band high-definition camera array deployed at key inspection nodes in the computer room is collected under the control of a synchronous trigger clock. Thermal radiation sensor data and visible light image data at the same time stamp are extracted from the synchronous data stream to form a target inspection data pair set. Registration processing based on the spatial calibration parameters of the sensor array is performed on the target inspection data pair set to obtain thermal radiation inspection data frame group and visible light inspection data frame group with spatial position alignment relationship. Temperature field anomaly zoning processing is performed on the thermal radiation inspection data frame group. The thermal radiation change rate of each pixel position in the thermal radiation inspection data frame group is calculated using a preset spatial neighborhood gradient operator. Based on the thermal radiation change rate, the set of abnormal hot zone boundary contour lines with abrupt temperature distribution changes is extracted from the thermal radiation inspection data frame group. The visible light inspection data frame group is subjected to device semantic partitioning processing. A pre-built deep residual semantic segmentation network is called to extract semantic masks for the cabinet equipment area and cable routing area in the visible light inspection data frame group, and generate semantic mask maps of the cabinet equipment area and cable routing area. The set of abnormal hot zone boundary contour lines is spatially intersected and superimposed with the semantic mask of the cabinet equipment area and the semantic mask of the cable routing area to generate an abnormal heat distribution map of the equipment that coincides with the semantic mask of the cabinet equipment area and an abnormal heat distribution map of the lines that coincides with the semantic mask of the cable routing area. The preset hazard feature encoder is invoked to extract compressed features from the abnormal heat distribution map of the equipment and the abnormal heat distribution map of the line, generating a set of abnormal heat compression feature vectors. The set of abnormal heat compression feature vectors is then input into the hazard state transition prediction loop network to generate a hazard state evolution path description sequence with time continuity. Based on the hazard state evolution path description sequence, a safety hazard warning instruction is generated for the target cabinet equipment identifier or the target cable routing section identifier.

2. The method for identifying potential safety hazards during computer room inspections according to claim 1, characterized in that, The synchronous data stream jointly output by the non-visible light band thermal radiation sensor array and the visible light band high-definition camera array, deployed at key inspection nodes in the computer room, under synchronous trigger clock control, is collected. Thermal radiation sensor data and visible light image data at the same timestamp are extracted from this synchronous data stream to form a target inspection data pair set. Registration processing based on sensor array spatial calibration parameters is performed on this target inspection data pair set to obtain thermal radiation inspection data frame groups and visible light inspection data frame groups with spatial alignment, including: The physical arrangement spacing parameters of the sensing units of the non-visible light band thermal radiation sensing array and the pixel size parameters of the photosensitive elements of the visible light band high-definition camera array are obtained. A planar coordinate mapping function for non-visible light thermal radiation data is constructed based on the physical arrangement spacing parameters of the sensing units. The planar coordinate mapping function is used to map the discrete temperature measurement point values ​​collected by the thermal radiation sensing array to a continuous two-dimensional spatial planar grid. Furthermore, a perspective projection inverse calculation function for visible light image data is constructed based on the pixel size parameters of the photosensitive elements and the lens focal length parameters of the visible light band high-definition camera array. The perspective projection inverse calculation function is used to map the pixel coordinates of the visible light image to the physical coordinate system of the computer room inspection space. In the physical coordinate system of the inspection space in the computer room, the geometric installation origin of the non-visible light band thermal radiation sensing array is used as the reference origin. The spatial offset vector parameter of the optical center projection point of the visible light band high-definition camera array relative to the reference origin is calculated. The spatial offset vector parameter is used to perform affine transformation translation compensation processing on the physical coordinates of the visible light image output by the perspective projection inverse calculation function to obtain a set of visible light registration physical coordinates that share the same reference origin as the non-visible light thermal radiation data. The thermal radiation two-dimensional planar mesh output by the planar coordinate mapping function is subjected to mesh node density resampling processing. The resampled thermal radiation two-dimensional planar mesh nodes are then subjected to spatial bidirectional interpolation fitting processing with the corresponding pixel positions in the visible light registration physical coordinate set. During the spatial bidirectional interpolation fitting process, for visible light pixel areas not covered by the thermal radiation two-dimensional planar mesh nodes, a spatial thermal radiation numerical diffusion surface function is constructed based on the thermal radiation values ​​of the thermal radiation nodes in the neighborhood of the visible light pixel area. The spatial thermal radiation numerical diffusion surface function is used to assign thermal radiation numerical values ​​to each coordinate point in the visible light registration physical coordinate set, generating a dense thermal radiation registration data frame group that is aligned point by point with the visible light image pixels. The dense thermal radiation registration data frame group is used as a thermal radiation inspection data frame group with spatial alignment, and the original visible light image frame whose timestamp matches the dense thermal radiation registration data frame group is used as a visible light inspection data frame group with spatial alignment.

3. The method for identifying potential safety hazards during computer room inspections according to claim 1, characterized in that, The process involves performing temperature field anomaly zoning processing on the thermal radiation inspection data frame group, calculating the rate of change of thermal radiation at each pixel location in the thermal radiation inspection data frame group using a preset spatial neighborhood gradient operator, and extracting a set of boundary contour lines of anomalous hot zones with abrupt temperature distribution changes in the thermal radiation inspection data frame group based on the rate of change of thermal radiation, including: In the thermal radiation inspection data frame group, a rectangular analysis window area is defined with the pixel position to be processed as the center and covering the adjacent peripheral pixel positions. All adjacent pixel positions within the rectangular analysis window area are traversed. The positive and negative differences between the thermal radiation value of the center pixel position and the thermal radiation values ​​of each peripheral pixel position in the horizontal direction are calculated. The positive and negative differences between the thermal radiation value of the center pixel position and the thermal radiation values ​​of each peripheral pixel position in the vertical direction are also calculated. The positive and negative difference values ​​in the horizontal direction and the positive and negative difference values ​​in the vertical direction are squared and summed, and then the square root is calculated to obtain the gradient magnitude response parameter that characterizes the degree of drastic change in thermal radiation at the pixel location. The arctangent angle value is calculated based on the positive difference component in the horizontal direction and the positive difference component in the vertical direction, thus obtaining the gradient direction angle parameter that characterizes the direction of thermal radiation change at the pixel position. Traverse all pixel positions in the thermal radiation inspection data frame group to generate a global thermal radiation gradient magnitude distribution map containing the gradient magnitude response parameters corresponding to all pixel positions. In the global thermal radiation gradient magnitude distribution map, retrieve a set of pixel chains where the gradient magnitude response parameters undergo a step change and the gradient direction angle parameter differences between adjacent pixels satisfy a preset direction consistency constraint. The pixel chain set is subjected to contour tracking processing based on connected component analysis. The pixel chains with closed ends are marked as candidate closed boundary contours for thermal radiation anomalies. The internal region enclosed by the candidate closed boundary contours for thermal radiation anomalies is subjected to statistical processing of the mean thermal radiation value and the mean thermal radiation value of the surrounding background region. The closed boundary contours whose deviation between the mean thermal radiation value of the internal region and the mean thermal radiation value of the surrounding region is greater than a preset deviation threshold are identified as the set of boundary contours of abnormal hot zones with abrupt temperature distribution changes.

4. The method for identifying potential safety hazards during computer room inspections according to claim 1, characterized in that, The step of performing device semantic partitioning processing on the visible light inspection data frame group involves calling a pre-built deep residual semantic segmentation network to extract semantic masks for the cabinet equipment area and cable routing area in the visible light inspection data frame group, generating semantic mask maps for the cabinet equipment area and cable routing area, including: The original visible light image frames in the visible light inspection data frame group are input into the initial convolutional downsampling module of the deep residual semantic segmentation network to extract a set of shallow visual feature maps containing edge texture gradient responses. The set of shallow visual feature maps is then input into the first residual convolutional unit group of the deep residual semantic segmentation network. The shallow visual feature maps are processed by nonlinear interactive transformation between feature channels through a stacked identity mapping connection structure to generate the first deep residual feature map. The first deep residual feature map is input into the dilated spatial pyramid pooling module of the deep residual semantic segmentation network. Parallel dilated convolution kernels with different dilation parameters are used to perform multi-scale receptive field feature recoding on the first deep residual feature map to generate a set of multi-scale receptive field fusion feature maps. The set of multi-scale receptive field fusion feature maps is input into the second residual convolution unit group of the deep residual semantic segmentation network. The multi-scale receptive field fusion feature map is processed by channel dimensionality reduction and spatial detail restoration through depth-separable convolution operation to generate the second deep residual feature map. The second deep residual feature map is input into the first semantic prediction branch of the deep residual semantic segmentation network. A pixel-wise classifier is used to predict the foreground confidence score of each spatial location of the second deep residual feature map to belong to the cabinet equipment category, generating a cabinet equipment category probability distribution map. The cabinet equipment category probability distribution map is subjected to binarization threshold filtering. Pixel locations where the foreground confidence score of the cabinet equipment category exceeds the preset semantic extraction threshold are aggregated into cabinet equipment connected regions, generating a cabinet equipment region semantic mask map. The second deep residual feature map is input into the second semantic prediction branch of the deep residual semantic segmentation network. A pixel-wise classifier is used to predict the foreground confidence score of each spatial location of the second deep residual feature map to belong to the cable routing category, generating a cable routing category probability distribution map. The cable routing category probability distribution map is subjected to binarization threshold filtering. Pixel locations where the foreground confidence score of the cable routing category exceeds the preset semantic extraction threshold are aggregated into cable routing connected regions, generating a cable routing region semantic mask map. The semantic mask of the cabinet equipment area and the semantic mask of the cable routing area are stored as the output results of the equipment semantic partitioning process.

5. The method for identifying safety hazards during computer room inspections according to claim 1, characterized in that, The step of performing spatial intersection and overlay processing on the set of abnormal hot zone boundary contours with the semantic mask map of the cabinet equipment area and the semantic mask map of the cable routing area, respectively, to generate an abnormal heat distribution map of the equipment that coincides with the semantic mask map of the cabinet equipment area and an abnormal heat distribution map of the lines that coincides with the semantic mask map of the cable routing area, includes: Obtain a list of coordinates of the internal pixel region enclosed by each closed boundary contour line in the set of abnormal hot zone boundary contour lines; In the semantic mask map of the cabinet equipment area, each coordinate point in the internal pixel area coordinate list is traversed, and it is determined whether the semantic mask pixel value corresponding to the coordinate point in the semantic mask map of the cabinet equipment area is a valid activation value representing the cabinet equipment area. If the semantic mask pixel value is an effective activation value representing the cabinet equipment area, then the coordinate point is marked as a candidate pixel point for effective equipment hidden danger, and the thermal radiation value within the closed boundary contour line where the coordinate point is located is retained; and if the semantic mask pixel value is an invalid activation value representing the background area, then the coordinate point is marked as an invalid background interference pixel point, and the thermal radiation value within the closed boundary contour line where the coordinate point is located is set to zero. Aggregate the location coordinates of all candidate pixels marked as valid equipment potential hazards and their corresponding retained thermal radiation values ​​to construct a filtered thermal radiation data layer that only contains thermal radiation distribution information of the internal area of ​​the cabinet equipment. Then, perform layer overlay rendering on the filtered thermal radiation data layer and the equipment boundary outline of the semantic mask of the cabinet equipment area to generate a visual abnormal heat distribution map of the equipment that fills the internal area of ​​the cabinet equipment outline with a thermal radiation color map. In the semantic mask map of the cable routing area, each coordinate point in the coordinate list of the internal pixel area is traversed, and it is determined whether the semantic mask pixel value corresponding to the coordinate point in the semantic mask map of the cable routing area is a valid activation value representing the cable routing area. If the semantic mask pixel value is an effective activation value representing a cable routing area, then the coordinate point is marked as a candidate pixel for a valid line hazard, and the thermal radiation value within the closed boundary contour line where the coordinate point is located is retained; and if the semantic mask pixel value is an invalid activation value representing a background area, then the coordinate point is marked as an invalid background interference pixel, and the thermal radiation value within the closed boundary contour line where the coordinate point is located is set to zero. The location coordinates of all candidate pixels marked as valid line hazards and their corresponding retained thermal radiation values ​​are aggregated to construct a filtered thermal radiation data layer that contains only thermal radiation distribution information of the cable routing area. This filtered thermal radiation data layer is then overlaid and rendered with the cable routing boundary contour of the semantic mask of the cable routing area to generate a visualized line abnormal heat distribution map that fills the inner area of ​​the cable routing contour with a thermal radiation color map.

6. The method for identifying potential safety hazards during computer room inspections according to claim 1, characterized in that, The process involves calling a preset hazard feature encoder to extract compressed features from the abnormal heat distribution maps of the equipment and the lines, generating a set of compressed feature vectors for abnormal heat generation. This set of compressed feature vectors is then input into a hazard state transition prediction recurrent network to generate a hazard state evolution path description sequence with time continuity. Based on this hazard state evolution path description sequence, a safety hazard warning instruction is generated targeting the equipment identifier in the target cabinet or the cable routing section identifier, including: The abnormal heat distribution map of the equipment is input into the first convolutional compression branch of the hidden danger feature encoder. The spatial resolution of the abnormal heat distribution map of the equipment is compressed step by step to the preset low-dimensional feature map size through a series of stride convolutional downsampling layers to obtain the low-dimensional spatial feature map of the equipment hidden danger. The low-dimensional spatial feature map of the equipment hidden danger is input into the first global pooling compression branch of the hidden danger feature encoder. The global average pooling response is calculated for each feature channel of the low-dimensional spatial feature map of the equipment hidden danger to obtain the global statistical compression vector of the equipment hidden danger. The abnormal heat distribution map of the line is input into the second convolutional compression branch of the hidden danger feature encoder. The spatial resolution of the abnormal heat distribution map of the line is compressed to the preset low-dimensional feature map size through a series of stride convolutional downsampling layers to obtain a low-dimensional spatial feature map of the line hidden danger. The low-dimensional spatial feature map of the line hidden danger is input into the second global pooling compression branch of the hidden danger feature encoder. The global average pooling response is calculated for each feature channel of the low-dimensional spatial feature map of the line hidden danger to obtain the global statistical compression vector of the line hidden danger. The global statistical compression vector of the equipment hidden danger and the global statistical compression vector of the line hidden danger are concatenated and spliced ​​in the feature dimension direction to generate an abnormal heat compression feature vector set representing the overall thermal hidden danger status of the computer room at the current inspection time. The abnormal heat compression feature vector set is input into the input gate structure of the hidden danger state transition prediction recurrent network. The abnormal heat compression feature vector set at the current time is linearly transformed using the input gate weight matrix. It is then combined with the hidden layer state vector of the hidden danger state transition prediction recurrent network at the previous time for weighted fusion to generate the candidate memory unit state vector at the current time. The set of abnormal heating compressed feature vectors and the hidden layer state vector of the hidden danger state transition prediction recurrent network at the previous moment are input into the forget gate structure of the hidden danger state transition prediction recurrent network. The forget gate coefficient vector representing the degree of retention of historical hidden danger state information is calculated, and the forget gate coefficient vector is used to perform element-wise product update processing on the long-term memory unit state vector at the previous moment. The current candidate memory unit state vector is added element-wise to the long-term memory unit state vector updated by the forgetting gating coefficient vector to generate the current long-term memory unit state vector. The abnormal heating compressed feature vector set, the hidden layer state vector of the previous hidden danger state transition prediction recurrent network, and the long-term memory unit state vector of the current time are input into the output gate structure of the hidden danger state transition prediction recurrent network to calculate the output gating coefficient vector. The output gating coefficient vector is then multiplied element-wise with the current long-term memory unit state vector after being processed by the activation function to generate the hidden layer vector of the hidden danger state at the current time. The forward propagation process of the hazard state transition prediction loop network is repeatedly executed at multiple consecutive inspection times. The hidden layer vectors of the hazard state output each time are arranged into a hazard state evolution path description sequence according to the time order. The monotonic change trend direction and slope of the vector values ​​in the hazard state evolution path description sequence are analyzed. Based on the monotonic change trend direction and slope, a warning level identification code and a suggested action code are generated for the target cabinet equipment identification or the target cable routing section identification. The warning level identification code and the suggested action code are encapsulated into a safety hazard prompt instruction.

7. The method for identifying potential safety hazards during computer room inspections according to claim 1, characterized in that, The method further includes: The target inspection data is obtained by acquiring the historical thermal radiation inspection data record sequence output by the non-visible light band thermal radiation sensor array in the set under the synchronous trigger clock control. Periodic baseline drift modeling processing is performed on the historical thermal radiation inspection data record sequence. The mean parameter and variance parameter of thermal radiation value of each sensor unit in the historical thermal radiation inspection data record sequence within a fixed time period are calculated using a long-term sliding window statistical algorithm. Based on the mean parameter and variance parameter of thermal radiation values, a non-parametric thermal radiation baseline distribution function is constructed for each sensing unit. The non-parametric thermal radiation baseline distribution function is used to describe the allowable range of thermal radiation value fluctuations of the sensing unit under normal operating conditions. When acquiring real-time thermal radiation inspection data frame group, the real-time thermal radiation value corresponding to each sensing unit in the real-time thermal radiation inspection data frame group is substituted into the non-parametric thermal radiation reference baseline distribution function for deviation evaluation processing. The standard deviation multiple of the real-time thermal radiation value relative to the mean parameter of the thermal radiation value is calculated. If the standard deviation multiple exceeds the preset baseline deviation multiple threshold, the position coordinates of the sensing unit are marked as the reference deviation abnormal sensing point. All sensing units within the spatial neighborhood of the reference deviation anomaly sensing point are combined to form a local thermal field anomaly verification area. Within the local thermal field anomaly verification area, it is searched to see if there are continuously distributed sensing units that simultaneously meet the trigger condition that the standard deviation multiple exceeds the preset baseline deviation multiple threshold. If there are continuously distributed sensing units within the local thermal anomaly verification area that simultaneously meet the triggering condition, then the geometric outer contour of the local thermal anomaly verification area is added to the set of abnormal thermal zone boundary contours as a supplementary abnormal thermal zone boundary. Based on the supplemented set of abnormal hot zone boundary contours, the spatial intersection and overlay processing with the semantic mask map of the cabinet equipment area and the semantic mask map of the cable routing area is re-executed to update the abnormal heat distribution map of the equipment and the abnormal heat distribution map of the line. The abnormal heating distribution map of the equipment and the abnormal heating distribution map of the line are used to regenerate the abnormal heating compressed feature vector set and update the hazard status evolution path description sequence. The warning level identification code in the safety hazard prompt instruction is dynamically corrected according to the updated hazard status evolution path description sequence.

8. The method for identifying potential safety hazards during computer room inspections according to claim 1, characterized in that, The method further includes: The sequence of historical visible light inspection data frames acquired by the visible light band high-definition camera array within multiple consecutive synchronous trigger clock cycles is obtained. The sequence of historical visible light inspection data frames is the cumulative set of the visible light inspection data frames with spatial alignment within multiple consecutive synchronous trigger clock cycles. Perform a three-channel separation operation of hue, saturation, and brightness based on color space transformation on each frame of the historical visible light inspection data frame group sequence to extract the hue channel image frame sequence that reflects the aging state of the coating on the device surface. Perform hue offset cumulative analysis processing based on pixel value difference operation on the hue channel image frame sequence. Take the first hue channel image in the historical visible light inspection data frame group sequence as the reference hue frame and calculate the absolute offset parameter of each pixel position in each subsequent hue channel image relative to the reference hue frame in terms of hue value. The absolute offset parameters calculated from multiple consecutive frames are subjected to linear regression fitting along the time axis for each pixel position to obtain the tone offset rate slope parameter and tone offset fitting confidence parameter at each pixel position. In the current visible light image frame corresponding to the visible light inspection data frame group, the hue offset rate slope parameter corresponding to all pixel positions that coincide with the semantic mask map of the cabinet equipment area is extracted, and the pixel positions where the hue offset rate slope parameter exceeds the preset coating aging rate threshold are marked as candidate pixels for abnormal coating aging. The candidate pixels of abnormal coating aging are aggregated based on spatial connectivity to form a connected region block of abnormal coating aging that are spatially adjacent and meet the preset connected neighborhood distance condition. Calculate the mean of the area parameter and the slope parameter of the color shift rate within each of the abnormal aging connected regions of the coating, and determine the abnormal aging connected regions of the coating that have the area parameter exceeding the preset minimum alarm area threshold and the mean value exceeding the preset severe aging rate threshold as the severe aging alarm regions of the coating. Obtain the geometric outer polygon outline coordinate sequence of the severely aged coating alarm area, and map the geometric outer polygon outline coordinate sequence to the corresponding spatial position of the abnormal heat distribution map of the device. In the abnormal heat distribution map of the device, determine whether there is a situation where the distribution density of effective device hidden danger candidate pixels exceeds the preset hidden danger distribution density threshold within the area covered by the geometric outer polygon outline coordinate sequence. If the distribution density of valid candidate pixels for equipment hazards exceeds a preset hazard distribution density threshold within the area covered by the geometrically enclosing polygon contour coordinate sequence, a multi-hazard associated alarm identifier representing the coupling of coating aging and thermal hazards is generated. The multi-hazard associated alarm identifier, the geometrically enclosing polygon contour coordinate sequence of the severe coating aging alarm area, and the information of the abnormal heat generation distribution area of ​​the equipment that overlaps with the severe coating aging alarm area are encapsulated together as a coating thermal coupling hazard supplementary data segment. The coating thermal coupling hazard supplementary data segment is appended to the end field of the data payload of the safety hazard warning instruction.

9. The method for identifying safety hazards during computer room inspections according to claim 1, characterized in that, The method further includes: The original thermal radiation numerical matrix sequence collected by the non-visible light band thermal radiation sensing array in multiple consecutive synchronous trigger clock cycles is obtained. The thermal diffusion direction field calculation processing based on the neighborhood difference operator is performed on each frame of the original thermal radiation numerical matrix sequence to generate a set of thermal diffusion direction vector field maps corresponding to each frame of thermal radiation numerical matrix. For each frame of the thermal diffusion direction vector field map in the set of thermal diffusion direction vector field maps, perform heat source convergence analysis processing based on vector field divergence operator, calculate the two-dimensional spatial divergence value of the thermal diffusion direction vector at each pixel position, and mark the pixel positions with the two-dimensional spatial divergence value less than the preset negative divergence threshold as candidate pixels for heat flow convergence region. The cumulative number of times the same pixel position in the set of thermal diffusion direction vector field maps of multiple consecutive frames is marked as a candidate region for heat flow convergence is statistically processed to generate a cumulative distribution map of heat flow convergence frequency that characterizes the persistence of heat flow convergence. In the cumulative distribution map of heat flow convergence frequency, connected pixel regions whose heat flow convergence frequency exceeds a preset sustained frequency threshold are extracted, and the geometric outline of the connected pixel regions is used as the set of boundary outlines of the core candidate regions of potential heat sources. For the internal region enclosed by each closed boundary contour line in the set of candidate core regions of the potential heat source, perform thermal radiation temporal stability assessment processing, and calculate the temporal variation coefficient parameter of thermal radiation values ​​of all sensing units in the internal region within multiple consecutive synchronous trigger clock cycles. The internal region corresponding to the closed boundary contour line where the time series variation coefficient parameter is lower than the preset stability variation threshold is determined as the core region of the stable hidden heat source, and the position coordinates of the geometric center sensing unit of the core region of the stable hidden heat source are used as the coordinates of the source point of the hidden heat source. Map the coordinates of the source point of the hidden heat source to the visible light image pixel coordinate system corresponding to the visible light inspection data frame group, and crop a local visible light image block with a preset spatial size range centered on the mapped pixel point of the source point coordinates of the hidden heat source in the visible light image. The local visible light image patch is input into a pre-constructed fine-grained recognition convolutional network for computer room equipment components. The multi-level convolutional feature extraction layer and region proposal generation layer of the fine-grained recognition convolutional network for computer room equipment components output the category name of the computer room equipment components contained in the local visible light image patch and the coordinates of the component spatial bounding box. Based on the category name of the equipment component in the computer room and the coordinates of the component's spatial boundary box, a targeted hazard description text is generated for a specific equipment component, and the targeted hazard description text is appended to the hazard description field of the safety hazard warning instruction.

10. A computer room inspection safety hazard alert system, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the data center inspection safety hazard alert method according to any one of claims 1 to 9 by executing the machine-executable instructions.

Images