A contact network wear detection method based on deep learning

Abstract

The present invention discloses a contact network wear detection method based on deep learning, wherein the method utilizes a contact network image acquisition module, a vehicle bottom vibration compensation module and an on-board detection host installed on rail vehicles such as inspection vehicles and maintenance vehicles to cooperate with each other to perform real-time detection of the contact network surface wear degree during vehicle driving. The contact network image acquisition module includes two high-speed industrial cameras, a high-definition imaging trigger module, and an auxiliary light source module. The on-board detection host includes an image processing module, a stereo matching module and a point cloud computing module. The on-board detection host is connected to the high-definition imaging trigger module via a bus, controls the auxiliary light source module and the two industrial cameras to synchronously capture contact network image data, receives data through the image processing module, and realizes image filtering and noise reduction. The present invention utilizes a deep learning algorithm to realize three-dimensional reconstruction of the contact network, which can detect the contact network wear degree more intuitively and clearly.

Description

Technical Field

[0001] The present invention relates to a contact network wear detection method based on deep learning, belonging to the technical field of contact network status detection. Background Art

[0002] As of December 31, 2020, my country's high-speed railway operating mileage reached 38,000 kilometers, carrying 235.833 million passengers, accounting for 64.4% of all railway passenger traffic. The focus of my country's high-speed rail development has gradually shifted from construction to operation and maintenance. The continuous improvement in both performance and the quality of operation and maintenance has placed higher demands on the safe and reliable operation of the catenary power supply in traction power supply systems. Rapidity, automation, and intelligence will be the future trends in railway inspection. Since the pantograph and catenary maintain high-speed contact and friction during train operation, long-term operation inevitably leads to wear on the contact wire surface. If the wear is not promptly monitored and the contact wire is not replaced in severely worn sections, unstable power supply between the pantograph and the catenary may occur. In severe cases, arcing may occur, causing damage to the catenary and even paralyzing the entire railway line. Therefore, real-time monitoring of catenary wear is essential to ensure safe and reliable catenary power supply.

[0003] Catenary wear detection methods are primarily categorized into three types: manual inspection, non-contact inspection, and non-contact image-based inspection. Non-contact inspection offers a significant advantage in wear detection due to the significant workload and the uncertainty surrounding the error range of the data collected by the inspector. Compared to the other two methods, machine vision-based catenary inspection utilizes industrial cameras to simulate binocular vision for edge detection and 3D reconstruction of the catenary surface, achieving 3D visualization. This method offers high flexibility, intelligent equipment, and minimal disruption to normal traffic flow, leading to its increasing adoption.

[0004] In view of the above reasons, in order to meet the needs of contact network wear detection and maintenance, it is very necessary to design a contact network wear detection method based on deep learning. Summary of the Invention

[0005] To achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a contact network wear detection method based on deep learning, comprising the steps of:

[0006] S100, initializing parameters of each module of the contact network wear detection system;

[0007] S200, the on-board detection host is connected to the high-definition imaging trigger module through the bus, controlling the auxiliary light source module and two high-speed industrial cameras to synchronously capture the contact network image data;

[0008] S300, the on-board detection host receives data through the image processing module, performs image smoothing and filtering noise reduction, and realizes 3D reconstruction of the contact network through the stereo matching module based on the STransMNet algorithm and the point cloud computing module;

[0009] S400, the on-board detection host uses the vibration compensation module to modify the model, compare the contact network parameters in the database, and realize the wear degree assessment;

[0010] The catenary wear detection system includes a catenary image acquisition module, a vehicle bottom vibration compensation module and a vehicle-mounted detection host;

[0011] The contact network image acquisition module includes two high-speed linear array CCD cameras, auxiliary light sources and high-definition imaging trigger equipment;

[0012] The vibration compensation module is composed of a laser radar, which measures the deviation angle through multi-prism scattering scanning. The vehicle-mounted detection host includes an image processing module, a stereo matching module, and a point cloud computing module, and is connected to each module through a bus.

[0013] Use the image processing module to implement Gaussian filtering and denoising of images:

[0014] The Gaussian filter can be expressed as follows to implement image smoothing and filtering noise reduction:

[0015]

[0016] The first-order directional derivative of G(x,y) is:

[0017]

[0018]

[0019] Where: ▽G represents the gradient vector, and n represents the direction vector.

[0020] G n Convolve with the image f(x,y) and adjust the direction of n. When the edge directions are orthogonal, n is obtained.

[0021] The stereo matching module based on the STransMNet algorithm includes: a swing transformation module for extracting left and right image features, an optimal transmission module for matching cost constraints, a Modify-WTA module for disparity map regression, and a layer adjustment module for optimizing the regressed disparity map;

[0022] The stereo matching based on the STransMNet algorithm obtains a disparity map, including the following steps:

[0023] S301, extracting features from the left and right images through a swing transformation module, so that shallow features can contain rich local detail information and deep features can contain global correlation information;

[0024] S302, calculating feature similarity through correlation calculation to obtain a matching cost volume;

[0025] S303, through feature differentiation loss, improve the model training's ability to pay attention to details, specific loss L diff It can be calculated as:

[0026]

[0027]

[0028] Where: H represents the characteristic height; L ploe,j represents the jth pole line loss; z p,i represents the i-th feature on the epipolar line; y i For z p,i Corresponding labels; σ is softmax calculation;

[0029] Then construct the loss function:

[0030] L=w1L d1,r +w2L d1,f +w3L be,f +w4L rr +w5L diff

[0031] Among them: w1-w5 represents the loss function weight; L d1,r With L d1,f Average SmoothL1 loss of sub-resolution and original resolution disparity maps; L be,f is the cross entropy loss of the occlusion map, which represents the error between the predicted occlusion map and the true occlusion map:

[0032]

[0033] where z occ Indicates the occluded part, z noc Represents the non-occluded part, M and N represent the number of pixels in the corresponding area, L rr Express the relative response loss:

[0034]

[0035] where t i represents the elements in the matching pixel set in the matching matrix T, Represents the elements in the set of pixels that are not matched due to occlusion in the matching matrix T, N T With MT Respectively represent the total number of the above two sets;

[0036] S304, uniquely constraining the matching cost through the optimal transmission module;

[0037] S305, using Modify-WTA regression to generate a sub-resolution disparity map, and restoring it to the disparity map of the original image resolution through upsampling;

[0038] S306, fusing the left image with the restored disparity map and occlusion map information through a layer adjustment module to generate optimized disparity map and occlusion map;

[0039] The point cloud computing module combines the disparity map generated by the stereo matching module to achieve 3D reconstruction of the contact network:

[0040] The point cloud module converts the disparity map generated by the stereo matching module into a depth map:

[0041]

[0042] Where D and d represent depth and time difference respectively, B represents baseline length, f(x,y) represents pixel unit, and x 0l with x 0r Represents the column coordinates of the principal points of the left and right views respectively.

[0043] 3D reconstruction is achieved by calculating the point cloud in the camera coordinate system through the depth map:

[0044] Z=D

[0045]

[0046]

[0047] Among them, x and y are the column and row coordinates of the pixel.

[0048] The high-speed linear array CCD camera has a resolution of 4K and a scanning speed of 26K. The auxiliary light source uses LCD to provide high-quality lighting.

[0049] Compared with the existing technology, the present invention has the following beneficial effects: 1. Compared with traditional manual contact network wear detection, the present invention has higher detection efficiency and can avoid the instability of manual detection errors. 2. Compared with the lidar solution, the present invention is simpler and more economical. 3. Compared with the monocular detection method, the present invention has no recognition rate limit, because in principle, there is no need to identify and then measure, but all image objects are directly measured, reducing dependence on database samples. 4. The present invention realizes real-time three-dimensional reconstruction of the contact network based on deep learning, with higher accuracy and wider application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0051] Figure 1 The present invention is a schematic diagram of a contact network wear detection process.

[0052] Figure 2 This is the on-vehicle distribution diagram of the contact network wear detection equipment of the present invention.

[0053] Figure 3 This is a structural diagram of the contact network image acquisition module of the present invention.

[0054] Figure 4 This is the stereo matching flowchart based on STransMNet of the present invention. DETAILED DESCRIPTION

[0055] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments. Although the present application is further described in detail below in conjunction with the accompanying drawings and specific embodiments, the embodiments of the present invention are not limited to these embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in this application without making any creative efforts are within the scope of protection of this application.

[0056] In one embodiment, Figure 1 As shown, the present invention proposes a contact network wear detection method based on deep learning.

[0057] The specific implementation includes the following steps:

[0058] S100, initializing parameters of each module of the contact network wear detection system;

[0059] S200, the on-board detection host is connected to the high-definition imaging trigger module through the bus, controlling the auxiliary light source module and two high-speed industrial cameras to synchronously capture the contact network image data;

[0060] S300, the on-board detection host receives data through the image processing module, performs image smoothing and filtering noise reduction, and realizes 3D reconstruction of the contact network through the stereo matching module based on the STransMNet algorithm and the point cloud computing module;

[0061] S400, the on-board detection host uses the vibration compensation module to modify the model, compare the contact network parameters in the database, and realize the wear degree assessment;

[0062] The catenary wear detection system includes a catenary image acquisition module, a vehicle bottom vibration compensation module and a vehicle-mounted detection host;

[0063] The contact network image acquisition module includes two high-speed linear array CCD cameras, auxiliary light sources and high-definition imaging trigger equipment;

[0064] The vibration compensation module is composed of a laser radar, which measures the deviation angle through multi-prism scattering scanning. The vehicle-mounted detection host includes an image processing module, a stereo matching module, and a point cloud computing module, and is connected to each module through a bus.

[0065] Use the image processing module to implement Gaussian filtering and denoising of images:

[0066] The Gaussian filter can be expressed as follows to implement image smoothing and filtering noise reduction:

[0067]

[0068] The first-order directional derivative of G(x,y) is:

[0069]

[0070]

[0071] Where: ▽G represents the gradient vector, and n represents the direction vector.

[0072] G n Convolve with the image f(x,y) and adjust the direction of n. When the edge directions are orthogonal, n is obtained.

[0073] The stereo matching module based on the STransMNet algorithm includes: a swing transformation module for extracting left and right image features, an optimal transmission module for matching cost constraints, a Modify-WTA module for disparity map regression, and a layer adjustment module for optimizing the regressed disparity map;

[0074] The stereo matching based on the STransMNet algorithm obtains a disparity map, including the following steps:

[0075] S301, extracting features from the left and right images through a swing transformation module, so that shallow features can contain rich local detail information and deep features can contain global correlation information;

[0076] S302, calculating feature similarity through correlation calculation to obtain a matching cost volume;

[0077] S303, through feature differentiation loss, improve the model training's ability to pay attention to details, specific loss L diff It can be calculated as:

[0078]

[0079]

[0080] Where: H represents the characteristic height; L ploe,j represents the jth pole line loss; z p,i represents the i-th feature on the epipolar line; y i For z p,i Corresponding labels; σ is softmax calculation;

[0081] Then construct the loss function:

[0082] L=w1L d1,r +w2L d1,f +w3L be,f +w4L rr +w5L diff

[0083] Among them: w1-w5 represents the loss function weight; L d1,r With L d1,f Average SmoothL1 loss of sub-resolution and original resolution disparity maps; L be,f is the cross entropy loss of the occlusion map, which represents the error between the predicted occlusion map and the true occlusion map:

[0084]

[0085] where z occ Indicates the occluded part, z noc Represents the non-occluded part, M and N represent the number of pixels in the corresponding area, L rr Express the relative response loss:

[0086]

[0087] where t i represents the elements in the matching pixel set in the matching matrix T, Represents the elements in the set of pixels that are not matched due to occlusion in the matching matrix T, N T With M T Respectively represent the total number of the above two sets;

[0088] S304, uniquely constraining the matching cost through the optimal transmission module;

[0089] S305, using Modify-WTA regression to generate a sub-resolution disparity map, and restoring it to the disparity map of the original image resolution through upsampling;

[0090] S306, fusing the left image with the restored disparity map and occlusion map information through a layer adjustment module to generate optimized disparity map and occlusion map;

[0091] The point cloud computing module combines the disparity map generated by the stereo matching module to achieve 3D reconstruction of the contact network:

[0092] The point cloud module converts the disparity map generated by the stereo matching module into a depth map:

[0093]

[0094] Where D and d represent depth and time difference respectively, B represents baseline length, f(x,y) represents pixel unit, and x 0l with x 0r Represents the column coordinates of the principal points of the left and right views respectively.

[0095] 3D reconstruction is achieved by calculating the point cloud in the camera coordinate system through the depth map:

[0096] Z=D

[0097]

[0098]

[0099] Among them, x and y are the column and row coordinates of the pixel.

[0100] The high-speed linear array CCD camera has a resolution of 4K and a scanning speed of 26K. The auxiliary light source uses LCD to provide high-quality lighting.

[0101] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A contact network wear detection method based on deep learning, characterized in that: Including steps: S100, initializing parameters of each module of the contact network wear detection system; S200, the on-board detection host is connected to the high-definition imaging trigger module through the bus, controlling the auxiliary light source module and two high-speed industrial cameras to synchronously capture the contact network image data; S300, the on-board detection host receives data through the image processing module, performs image smoothing and filtering noise reduction, and realizes 3D reconstruction of the contact network through the stereo matching module based on the STransMNet algorithm and the point cloud computing module; S400, the on-board detection host uses the vibration compensation module to modify the model, compare the contact network parameters in the database, and realize the wear degree assessment; The catenary wear detection system includes a catenary image acquisition module, a vehicle bottom vibration compensation module and a vehicle-mounted detection host; The contact network image acquisition module includes two high-speed linear array CCD cameras, auxiliary light sources and high-definition imaging trigger equipment; The vibration compensation module is composed of a laser radar, which measures the deviation angle through multi-prism scattering scanning; The vehicle-mounted detection host includes an image processing module, a stereo matching module and a point cloud computing module, and is connected to each module via a bus; Use the image processing module to implement Gaussian filtering and denoising of images: Gaussian filter can be used to achieve image smoothing and noise reduction: The first-order directional derivative of G(x,y) is: in: represents the gradient vector, n represents the direction vector; G n Convolve with the image f(x,y) and adjust the direction of n. When the edge directions are orthogonal, n is obtained. The stereo matching module based on the STransMNet algorithm includes: a swing transformation module for extracting left and right image features, an optimal transmission module for matching cost constraints, a Modify-WTA module for disparity map regression, and a layer adjustment module for optimizing the regressed disparity map; The stereo matching based on the STransMNet algorithm obtains a disparity map, including the following steps: S301, extracting features from the left and right images through a swing transformation module, so that shallow features can contain rich local detail information and deep features can contain global correlation information; S302, calculating feature similarity through correlation calculation to obtain a matching cost volume; S303, through feature differentiation loss, improve the model training's ability to pay attention to details, specific loss L diff It can be calculated as: Where: H represents the characteristic height; L ploe,j represents the jth pole line loss; z p,i represents the i-th feature on the epipolar line; y i For z p,i The corresponding label; σ is the softmax calculation; Then construct the loss function: L=w1L d1,r +w2L d1,f +w3L be,f +w4L rr +w5L diff Among them: w1-w5 represents the loss function weight; L d1,r With L d1,f Average SmoothL1 loss of sub-resolution and original resolution disparity maps; L be,f is the cross entropy loss of the occlusion map, which represents the error between the predicted occlusion map and the true occlusion map: where z occ Indicates the occluded part, z noc Represents the non-occluded part, M and N represent the number of pixels in the corresponding area, L rr Express the relative response loss: where t i represents the elements in the matching pixel set in the matching matrix T, Represents the elements in the set of pixels that are not matched due to occlusion in the matching matrix T, N T With M T Respectively represent the total number of the above two sets; S304, uniquely constraining the matching cost through the optimal transmission module; S305, using Modify-WTA regression to generate a sub-resolution disparity map, and restoring it to the disparity map of the original image resolution through upsampling; S306, fusing the left image with the restored disparity map and occlusion map information through a layer adjustment module to generate optimized disparity map and occlusion map; The point cloud computing module combines the disparity map generated by the stereo matching module to achieve 3D reconstruction of the contact network: The point cloud module converts the disparity map generated by the stereo matching module into a depth map: Where D and d represent depth and time difference respectively, B represents baseline length, f(x,y) represents pixel unit, and x 0l with x 0r Represents the column coordinates of the principal points of the left and right views respectively; 3D reconstruction is achieved by calculating the point cloud in the camera coordinate system through the depth map: Z=D Among them, x and y are the column and row coordinates of the pixel.

2. The method for detecting contact network wear based on deep learning according to claim 1, characterized in that: The high-speed linear array CCD camera has a resolution of 4K and a scanning speed of 26K. The auxiliary light source uses LCD to provide high-quality lighting.

Images