A method and device for monitoring and detecting longitudinal displacement and locked rail temperature of seamless railway rails

Abstract

This application relates to a method and device for image monitoring and detection of the longitudinal displacement and locked rail temperature of seamless railway rails. The method comprises: using a track inspection vehicle to collect multiple rail images of a seamless railway rail section; constructing an image stitching model to obtain a fused rail image; constructing a longitudinal displacement monitoring model, inputting the fused rail image into the longitudinal displacement monitoring model to obtain the longitudinal displacement of the rail; and calculating the actual locked rail temperature of the seamless railway rail section based on the longitudinal displacement of the rail. This method does not rely on observation point monitoring equipment or manual on-site measurement, reducing detection costs and improving detection convenience and applicability.

Description

Technical Field

[0001] The present application relates to the field of seamless rail displacement and locked rail temperature measurement, and specifically to a method and device for image monitoring and detection of seamless line rail longitudinal displacement and locked rail temperature. Background Art

[0002] A seamless rail is a long, jointless rail. Shorter rails are welded together to create continuous tracks up to several kilometers long. This design reduces the number of joints, improving track stability and comfort while also reducing maintenance. The actual locking rail temperature of a seamless rail refers to the ambient temperature at the time the seamless rail is laid and locked onto the sleeper. It determines the rail's initial thermal state and serves as a benchmark for evaluating the rail's thermal expansion and contraction behavior.

[0003] Measuring the actual locked rail temperature is crucial to railway transportation safety and efficiency. Excessively high temperatures can cause rail expansion, leading to deformation, cracking, bending, and even derailment. When temperatures are too low, rails are susceptible to frost heave and other effects, which can also negatively impact line safety. By locking rail temperature on seamless lines, rail temperature can be monitored and adjusted in a timely manner to ensure safe operation. Furthermore, changes in rail temperature can cause the line to expand and contract, affecting its geometry and position. Excessive rail temperature fluctuations can cause the gap between the track and other facilities to become excessively large or small, increasing maintenance and adjustment workload and potentially causing increased wear on the line.

[0004] The actual locked rail temperature can be measured using rail displacement. However, traditional measurement methods, such as manual wire drawing on displacement observation piles, optical instrument observation on displacement observation piles, and laser measurement of rail longitudinal displacement, suffer from poor accuracy and cannot meet the high-efficiency and high-precision measurement requirements of high-speed railways. Furthermore, each measuring point requires monitoring equipment, and the number of monitoring devices directly impacts costs, making them relatively expensive. Furthermore, on kilometer-span bridges, the complexity of seamless rail tracks and the significant additional forces make existing testing methods incapable of accurately measuring the actual locked rail temperature.

[0005] Therefore, designing an accurate and convenient method for monitoring and detecting the longitudinal displacement and locked rail temperature of seamless railway rails is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method and device for monitoring and detecting the longitudinal displacement and locked rail temperature of seamless line rails. A track inspection vehicle is used to collect images of seamless line rails, and the longitudinal displacement of the rails is obtained by splicing and analyzing the rail images. The actual locked rail temperature is determined by the longitudinal displacement of the rails, thereby improving the accuracy of the actual locked rail temperature detection of the rails, achieving convenience in detection, and reducing the cost of detection.

[0007] This application first discloses a method for monitoring and detecting the longitudinal displacement and locked rail temperature of a seamless railway rail. The specific steps are as follows:

[0008] Step 1: Use a track inspection vehicle to collect multiple rail images of a seamless track unit section;

[0009] Step 2: construct an image stitching model, input the multiple rail images into the image stitching model, the image stitching model includes a feature extraction module, a feature description module, a feature matching module and a feature fusion module, the feature fusion module divides the overlapping area of ​​the stitched image into multiple units of equal size, calculates the weight coefficient of the center point of each unit, and processes each pixel in the unit using the weight coefficient to obtain a fused rail image;

[0010] Step 3. Construct a longitudinal displacement monitoring model, input the fused rail image into the longitudinal displacement monitoring model, and obtain the longitudinal displacement of the rail. The longitudinal displacement monitoring model includes an encoder and a decoder. The fused rail image and the reference rail image are respectively input into two channels of the encoder. Each channel includes a feature extraction layer, a first encoding layer, and a second encoding layer connected in sequence. The outputs of the two channels are input into the mutual information layer to obtain a mutual information feature map. The mutual information feature map is input into the interactive enhancement layer after passing through the third encoding layer and the fourth encoding layer. The decoder includes a first decoding layer, a second decoding layer, a third decoding layer, and a fourth decoding layer connected in sequence. The output of the interactive enhancement layer and the output of the third encoding layer are input into the first decoding layer, and the mutual information feature map is input into the second decoding layer. The outputs of the first encoding layers of the two channels are merged and input into the second decoding layer. The outputs of the feature extraction layers of the two channels are merged and input into the first decoding layer. The output of the first decoding layer passes through a 1×1 convolutional layer to output the longitudinal displacement of the rail.

[0011] Step 4: Calculate the actual locking rail temperature of the seamless line unit rail section based on the longitudinal displacement of the rail.

[0012] Preferably, the feature extraction module uses the SURF algorithm to extract feature points of the multiple rail images, the feature description module uses the BRISK algorithm to generate feature point descriptors, the feature matching module uses the NNDR algorithm to perform bidirectional coarse matching on the feature point descriptors, and uses the MSAC algorithm to eliminate erroneous matching points to obtain a spliced ​​image.

[0013] Preferably, the feature fusion module divides the overlapping area of ​​the stitched image into a plurality of units of equal size, calculates a weight coefficient of the center point of each unit, and processes each pixel in the unit using the weight coefficient, including:

[0014] The feature fusion module divides the overlapping area into a plurality of units of equal size and calculates the weight coefficient of the center point of each unit. The weight coefficient is:

[0015]

[0016] in, is the weight coefficient, x is the coordinate of the center point, and are the starting and ending coordinates of the overlapping area respectively;

[0017] Use the weight coefficient of the center point as the weight coefficient of each pixel in the entire unit, and perform separate weight coefficient calculations or calculate weight coefficients for units of appropriate size for pixels close to the edge of the overlapping area and not entering the unit;

[0018] The feature fusion module processes each pixel in the unit of the overlapping area, which is expressed as:

[0019]

[0020] in, is the weight coefficient, is the overlapping area image, and are the two images to be stitched.

[0021] Preferably, the feature extraction layer includes a 7×7 convolutional layer and two 3×3 convolutional layers connected in sequence, and the 7×7 convolutional layer and the two 3×3 convolutional layers are spliced ​​to output a multi-level feature map;

[0022] The mutual information layer uses a local window of a preset size to slide on the feature maps output by the two channels with a preset step size and extracts local feature vectors, calculates the mutual information of the local feature vector pairs corresponding to the same position of the local window on the two feature maps, and constructs a matrix of the mutual information values ​​between all the local feature vector pairs to obtain the mutual information feature map;

[0023] The interactive enhancement layer includes two parallel branches, wherein the first branch includes a linear layer, a depth-separable convolutional layer, a selection scanning module and an activation layer connected in sequence, and the second branch includes a channel attention layer and a spatial attention layer connected in sequence. The feature maps output by the fourth encoding layer are respectively input to the two branches, and the outputs of the two branches are aggregated and subjected to a linear layer to obtain an enhanced feature map.

[0024] Preferably, the first encoding layer, the second encoding layer, the third encoding layer and the fourth encoding layer are composed of a depth-separable convolution layer, an activation layer, a two-dimensional convolution layer, a batch normalization layer and a maximum pooling layer, and the first decoding layer, the second decoding layer, the third decoding layer and the fourth decoding layer are composed of a deconvolution layer, a splicing layer, a two-dimensional convolution layer, an activation layer and a batch normalization layer.

[0025] Preferably, the calculating of the actual locking rail temperature of the seamless railway unit rail section according to the longitudinal displacement of the rail includes:

[0026] The actual locking rail temperature calculation formula is:

[0027]

[0028] Where Ts is the actual locked rail temperature, T is the on-site measured rail temperature, Δl is the longitudinal displacement of the rail, l is the length of the seamless line unit rail section, and α is the rail expansion coefficient.

[0029] On the other hand, the present application also provides a device for monitoring and detecting longitudinal displacement and locked rail temperature of a seamless railway rail, the device comprising:

[0030] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the above-mentioned method for monitoring and detecting the longitudinal displacement and locked rail temperature of a seamless line rail.

[0031] This application proposes a method and device for monitoring and detecting the longitudinal displacement and locked rail temperature of seamless line rails. A track inspection vehicle is used to collect rail images, which does not rely on observation point monitoring equipment or manual on-site measurement, thereby reducing the cost of detection and improving the convenience and applicability of detection; an image stitching model is constructed to improve the accuracy of image stitching and the image quality of the overlapping area of ​​image stitching, thereby providing accurate data for subsequent longitudinal displacement analysis; a longitudinal displacement monitoring model is constructed to improve the performance of displacement measurement by using mutual information layers and interactive enhancement layers, thereby improving the accuracy of measured displacement. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flow chart of a method for monitoring and detecting longitudinal displacement and locked rail temperature of a seamless railway rail provided by an exemplary embodiment of the present invention;

[0033] Figure 2 A schematic structural diagram of a longitudinal displacement monitoring model provided by an exemplary embodiment of the present invention;

[0034] Figure 3 A schematic structural diagram of an interactive enhancement layer provided by an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0035] The detailed description set forth below in conjunction with the accompanying drawings is intended to be a description of various exemplary embodiments of the present invention, and is not intended to represent the only embodiment that can put the present invention into practice. For the purpose of providing a thorough understanding of the present invention, the detailed description includes specific details. However, it will be apparent to those skilled in the art that the present invention can be put into practice without these specific details. In some instances, well-known structures and components are shown in block diagram form to avoid blurring the concept of the present invention.

[0036] Example 1

[0037] This embodiment provides a method for monitoring and detecting the longitudinal displacement and locked rail temperature of a seamless railway rail. Figure 1 As shown, the method provided in this embodiment includes the following steps:

[0038] Step 1: Use a track inspection vehicle to collect multiple rail images of a seamless line unit rail section.

[0039] Among them, the existing rail longitudinal displacement monitoring usually uses visual acquisition equipment fixed at the monitoring point to collect rail images. However, due to the large scale of the railway network, the number of visual acquisition equipment will be huge, which will bring huge monitoring costs. At the same time, manual on-site measurements can also be used, which also leads to a large amount of manpower consumption and reduced inspection efficiency and accuracy. The track inspection vehicle is a special type of railway vehicle, mainly used to detect track status and evaluate track quality. The track inspection vehicle can quickly and accurately obtain rail image data. Therefore, using the track inspection vehicle to collect rail images as the basis for rail longitudinal displacement and actual locked rail temperature will reduce the use of fixed monitoring point equipment and manpower consumption, thereby reducing monitoring costs and improving detection accuracy.

[0040] Step 2: Construct an image stitching model and input the multiple rail images into the image stitching model. The image stitching model includes a feature extraction module, a feature description module, a feature matching module, and a feature fusion module. The feature fusion module divides the overlapping area of ​​the stitched image into multiple units of equal size, calculates the weight coefficient of the center point of each unit, and processes each pixel in the unit using the weight coefficient to obtain a fused rail image.

[0041] Among them, the track inspection vehicle is limited by the installation position of its CCD camera. It can only collect partial images of the seamless line unit rail section at a time, and cannot collect the image of the entire rail section at one time. Therefore, the track inspection vehicle will collect multiple continuous rail images, and these rail images are spliced ​​together to form the image of the entire rail section.

[0042] Step 3: Construct a longitudinal displacement monitoring model, such as Figure 2 As shown, the fused rail image is input into the longitudinal displacement monitoring model to obtain the longitudinal displacement of the rail. The longitudinal displacement monitoring model includes an encoder and a decoder. The fused rail image and the reference rail image are respectively input into the two channels of the encoder. Each channel includes a feature extraction layer, a first encoding layer and a second encoding layer connected in sequence. The outputs of the two channels are input into the mutual information layer to obtain a mutual information feature map. The mutual information feature map is input into the interactive enhancement layer after passing through the third encoding layer and the fourth encoding layer. The decoder includes a first decoding layer, a second decoding layer, a third decoding layer and a fourth decoding layer connected in sequence. The output of the interactive enhancement layer and the output of the third encoding layer are input into the first decoding layer, and the mutual information feature map is input into the second decoding layer. The outputs of the first encoding layers of the two channels are merged and input into the second decoding layer. The outputs of the feature extraction layers of the two channels are merged and input into the first decoding layer. The output of the first decoding layer passes through a 1×1 convolutional layer to output the longitudinal displacement of the rail.

[0043] Step 4: Calculate the actual locking rail temperature of the seamless line unit rail section based on the longitudinal displacement of the rail.

[0044] In step 2, the feature extraction module uses the SURF algorithm to extract feature points of the multiple rail images.

[0045] Among them, the SURF algorithm is an algorithm used for feature detection and matching in the field of computer vision. It has the characteristics of scale invariance, rotation invariance, robustness and high efficiency, and is widely used in image stitching, object recognition, image registration and other fields.

[0046] The feature description module uses the BRISK algorithm to generate feature point descriptors and binary descriptors to reduce memory usage and improve matching efficiency.

[0047] Among them, the BRISK algorithm is a feature detection and description algorithm for computer vision and image processing. It was proposed by Stefan Leutenegger et al. in 2011 and is widely used in applications that require fast and robust feature matching.

[0048] The feature matching module uses the NNDR algorithm to perform a rough bidirectional matching of feature point descriptors. Since some incorrect matching pairs may still exist after the initial screening, a bidirectional matching method is used to eliminate incorrect matching pairs to improve matching accuracy. Specifically, after the unidirectional matching, the matching pairs are first saved. Then, the relationship between the matching image and the original image is swapped, and a reverse matching is performed. The common matching pairs are selected as the final matching results, and the unidirectional matching pairs are deleted.

[0049] Among them, the NNDR (Nearest Neighbor Distance Ratio) algorithm is a strategy for feature point matching. It screens matching pairs based on the distance ratio between the nearest neighbor and the next nearest neighbor. The core idea of ​​this method is that if the ratio of the nearest neighbor distance to the next nearest neighbor distance of a feature point is less than a certain threshold, the match is considered reliable.

[0050] The MSAC algorithm is used to eliminate incorrect matching points and obtain the stitched image.

[0051] The Minimum Sample Algorithm for Consensus (MSAC) algorithm is a robust estimation algorithm for estimating mathematical model parameters from a set of data. Proposed by Fischler and Bolles in 1981, it is primarily used for geometric estimation problems in computer vision, such as feature point matching, camera calibration, and 3D reconstruction. The core idea of ​​the MSAC algorithm is to find the optimal model parameters from a set of data that may contain outliers, thereby maximizing the consistency between the model and the data.

[0052] In step 2, traditional image fusion algorithms directly fuse the registered images, which will produce very obvious seams and color differences and seriously lead to excessive color unevenness or artifacts. The existing weighted fusion algorithm improves the traditional fusion algorithm. The weight is determined by the distance between the pixels in the overlapping area and the boundary of the overlapping area, thereby achieving a smooth transition and eliminating seams, as shown below:

[0053]

[0054] in, and is the weight coefficient, is the overlapping area image, and are the two images to be stitched, and are the starting and ending coordinates of the overlapping area respectively.

[0055] However, since the intersection point of the two weight functions and the inflection point of the fusion function are the same, the overlapping area will have shadows after registration. In addition, the proposed method does not work well when the brightness difference between the two stitched images is large.

[0056] Therefore, in this embodiment, the weight function is adjusted to accommodate the changing trend from 1 to 0, and the parabolic property is used to optimize the weight function. Furthermore, to improve fusion efficiency, a unit-based integration acceleration strategy is proposed. The overlapping region is divided into multiple units, and a weight coefficient is calculated for each unit's center point, representing the weight of all pixels within the unit. This results in better image fusion performance at different brightness levels, angles, resolutions, and scales, effectively resolving ghosting and seam issues and providing accurate data for subsequent longitudinal displacement analysis.

[0057] The feature fusion module divides the overlapping area of ​​the stitched image into multiple units of equal size, calculates the weight coefficient of the center point of each unit, and processes each pixel in the unit using the weight coefficient, including:

[0058] The feature fusion module divides the overlapping area into a plurality of units of equal size and calculates the weight coefficient of the center point of each unit. The weight coefficient is:

[0059]

[0060] in, is the weight coefficient, x is the coordinate of the center point, and are the starting and ending coordinates of the overlapping area respectively;

[0061] Use the weight coefficient of the center point as the weight coefficient of each pixel in the entire unit, and perform separate weight coefficient calculations or calculate weight coefficients for units of appropriate size for pixels close to the edge of the overlapping area and not entering the unit;

[0062] The feature fusion module processes each pixel in the unit of the overlapping area, which is expressed as:

[0063]

[0064] in, is the weight coefficient, is the overlapping area image, and are the two images to be stitched.

[0065] In step 3, the longitudinal displacement monitoring model includes a dual-channel input and mutual information layer to extract deep-level feature information from the two rail images and perform fusion and matching. Furthermore, to enhance the network's feature extraction capabilities during the encoding phase, a feature extraction module is introduced that extracts displacement features at different scales and depths through depthwise separable convolutions of varying sizes. Furthermore, an interactive enhancement layer, including a selective scanning module, a channel attention module, and a spatial attention module, is incorporated into the longitudinal displacement monitoring model to enhance the model's representation of feature maps, particularly focusing on channels containing rich displacement field information.

[0066] The feature extraction layer includes a 7×7 convolutional layer and two 3×3 convolutional layers connected in sequence. The 7×7 convolutional layer and the two 3×3 convolutional layers are spliced ​​to output a multi-level feature map.

[0067] The mutual information layer uses a local window of a preset size to slide on the feature maps output by the two channels with a preset step size and extracts local feature vectors, calculates the mutual information of the local feature vector pairs corresponding to the same position of the local window on the two feature maps, and constructs a matrix of the mutual information values ​​between all the local feature vector pairs to obtain the mutual information feature map.

[0068] Mutual information is a concept in information theory that measures the interdependence between two random variables. Specifically, it describes the amount of information one random variable contains about another. A larger mutual information value indicates a stronger dependence between the two variables; a smaller mutual information value indicates a weaker dependence.

[0069] The interaction enhancement layer includes two parallel branches, such as Figure 3 As shown, the first branch includes a linear layer, a depth-separable convolutional layer, a selection scanning module and an activation layer connected in sequence, the second branch includes a channel attention layer and a spatial attention layer connected in sequence, the feature maps output by the fourth encoding layer are respectively input to the two branches, and the outputs of the two branches are aggregated and subjected to linear layers to obtain enhanced feature maps.

[0070] Among them, the interaction enhancement layer aims to integrate global and local information processing capabilities, screen features through channel and spatial attention mechanisms, highlight important features and suppress unimportant information, while achieving efficient feature extraction at a lower computational cost.

[0071] The Selective Scan Module (SSM) is a two-dimensional state-space model in the Mamba architecture. It can capture the dependency between any two points in the image by performing selective scanning in four directions. It has linear time complexity when capturing dependencies, which means it can process large-scale image data while maintaining computational efficiency.

[0072] Preferably, the first encoding layer, the second encoding layer, the third encoding layer and the fourth encoding layer are composed of a depth-separable convolution layer, an activation layer, a two-dimensional convolution layer, a batch normalization layer and a maximum pooling layer, and the first decoding layer, the second decoding layer, the third decoding layer and the fourth decoding layer are composed of a deconvolution layer, a splicing layer, a two-dimensional convolution layer, an activation layer and a batch normalization layer.

[0073] In step 4, the thermal expansion and contraction of rails refers to the phenomenon in which rails expand when the temperature increases and contract when the temperature decreases. This physical property of rails can be quantified using the coefficient of linear expansion (α), which represents the change in length per unit length of rail for a 1°C temperature change. For rails, the coefficient of linear expansion is typically around 12 microns / meter·°C, meaning that for every 1°C temperature change, the length of the rail changes by approximately 12 microns.

[0074] The locked rail temperature refers to the temperature of the rail when it is laid, at which point the rail length is considered standard. When the actual temperature deviates from the locked rail temperature, the rail will shift longitudinally.

[0075] Calculating the actual locked rail temperature of the seamless railway unit rail section according to the longitudinal displacement of the rail includes:

[0076] The actual locking rail temperature calculation formula is:

[0077]

[0078] Where Ts is the actual locked rail temperature, T is the on-site measured rail temperature, Δl is the longitudinal displacement of the rail, l is the length of the seamless line unit rail section, and α is the rail expansion coefficient.

[0079] Example 2

[0080] This embodiment provides a device for monitoring and detecting longitudinal displacement and locked rail temperature of a seamless railway rail, the device comprising:

[0081] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the above-mentioned method for monitoring and detecting the longitudinal displacement and locked rail temperature of a seamless line rail.

[0082] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for monitoring and detecting longitudinal displacement and locked rail temperature of a seamless railway track, characterized in that: The method comprises: Step 1: Use a track inspection vehicle to collect multiple rail images of a seamless track unit section; Step 2: construct an image stitching model, input the multiple rail images into the image stitching model, the image stitching model includes a feature extraction module, a feature description module, a feature matching module and a feature fusion module, the feature fusion module divides the overlapping area of ​​the stitched image into multiple units of equal size, calculates the weight coefficient of the center point of each unit, and processes each pixel in the unit using the weight coefficient to obtain a fused rail image; Step 3. Construct a longitudinal displacement monitoring model, input the fused rail image into the longitudinal displacement monitoring model, and obtain the longitudinal displacement of the rail. The longitudinal displacement monitoring model includes an encoder and a decoder. The fused rail image and the reference rail image are respectively input into two channels of the encoder. Each channel includes a feature extraction layer, a first encoding layer, and a second encoding layer connected in sequence. The outputs of the two channels are input into the mutual information layer to obtain a mutual information feature map. The mutual information feature map is input into the interactive enhancement layer after passing through the third encoding layer and the fourth encoding layer. The decoder includes a first decoding layer, a second decoding layer, a third decoding layer, and a fourth decoding layer connected in sequence. The output of the interactive enhancement layer and the output of the third encoding layer are input into the first decoding layer, and the mutual information feature map is input into the second decoding layer. The outputs of the first encoding layers of the two channels are merged and input into the second decoding layer. The outputs of the feature extraction layers of the two channels are merged and input into the first decoding layer. The output of the first decoding layer passes through a 1×1 convolutional layer to output the longitudinal displacement of the rail. Step 4: Calculate the actual locking rail temperature of the seamless line unit rail section based on the longitudinal displacement of the rail.

2. The method according to claim 1, characterized in that The feature extraction module uses the SURF algorithm to extract feature points of the multiple rail images, the feature description module uses the BRISK algorithm to generate feature point descriptors, the feature matching module uses the NNDR algorithm to perform bidirectional coarse matching on the feature point descriptors, and uses the MSAC algorithm to eliminate erroneous matching points to obtain a spliced ​​image.

3. The method according to claim 1, characterized in that The feature fusion module divides the overlapping area of ​​the stitched image into multiple units of equal size, calculates the weight coefficient of the center point of each unit, and processes each pixel in the unit using the weight coefficient, including: The feature fusion module divides the overlapping area into a plurality of units of equal size and calculates the weight coefficient of the center point of each unit. The weight coefficient is: in, is the weight coefficient, x is the coordinate of the center point, and are the starting and ending coordinates of the overlapping area respectively; Use the weight coefficient of the center point as the weight coefficient of each pixel in the entire unit, and perform separate weight coefficient calculations or calculate weight coefficients for units of appropriate size for pixels close to the edge of the overlapping area and not entering the unit; The feature fusion module processes each pixel in the unit of the overlapping area, which is expressed as: in, is the weight coefficient, is the overlapping area image, and are the two images to be stitched.

4. The method according to claim 1, wherein The feature extraction layer includes a 7×7 convolutional layer and two 3×3 convolutional layers connected in sequence, and the 7×7 convolutional layer and the two 3×3 convolutional layers are spliced ​​to output a multi-level feature map; The mutual information layer uses a local window of a preset size to slide on the feature maps output by the two channels with a preset step size and extracts local feature vectors, calculates the mutual information of the local feature vector pairs corresponding to the same position of the local window on the two feature maps, and constructs a matrix of the mutual information values ​​between all the local feature vector pairs to obtain the mutual information feature map; The interactive enhancement layer includes two parallel branches, wherein the first branch includes a linear layer, a depth-separable convolutional layer, a selection scanning module and an activation layer connected in sequence, and the second branch includes a channel attention layer and a spatial attention layer connected in sequence. The feature maps output by the fourth encoding layer are respectively input to the two branches, and the outputs of the two branches are aggregated and subjected to a linear layer to obtain an enhanced feature map.

5. The method according to claim 1, wherein The first encoding layer, the second encoding layer, the third encoding layer and the fourth encoding layer are composed of a depth-separable convolution layer, an activation layer, a two-dimensional convolution layer, a batch normalization layer and a maximum pooling layer, and the first decoding layer, the second decoding layer, the third decoding layer and the fourth decoding layer are composed of a deconvolution layer, a splicing layer, a two-dimensional convolution layer, an activation layer and a batch normalization layer.

6. The method according to claim 1, characterized in that Calculating the actual locked rail temperature of the seamless railway unit rail section according to the longitudinal displacement of the rail includes: The actual locking rail temperature calculation formula is: Where Ts is the actual locked rail temperature, T is the on-site measured rail temperature, Δl is the longitudinal displacement of the rail, l is the length of the seamless line unit rail section, and α is the rail expansion coefficient.

7. A seamless railway rail longitudinal displacement and locked rail temperature image monitoring and detection device, characterized in that: The device comprises: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute a method for monitoring and detecting the longitudinal displacement and locked rail temperature of a seamless line rail as described in any one of claims 1 to 6.

Images