Deep learning-based leaky coaxial cable clamp fault detection method

Abstract

The invention discloses a deep learning-based leaky coaxial cable clamp fault detection method. The method comprises the following steps of S1, carrying out image acquisition; s2, taking the images in the tunnel and out of the tunnel as a training sample set; s3, building a ResNet50 network model, and inputting the training sample set into the model for training; s4, inputting the video to be identified into the ResNet50 network model in sequence according to frames, and judging whether each image is in the tunnel or outside the tunnel; s5, a Yolov8-Retinexformer target detection model is built, and a target detection model is built; s6, using the target detection model to realize fault clamp identification and positioning; and S7, inferring the kilometer post position of the fault clamp in the tunnel by combining the vehicle speed information, the camera frame rate and the position information of the image. According to the method, classification of images inside and outside the tunnel is realized by adopting an image classification algorithm based on deep learning, fixture fault detection in the tunnel is realized by adopting the improved Yolov8, and the position of the fault fixture is obtained by combining vehicle speed information and a camera frame rate, so that the workload of detection personnel is greatly reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition, and in particular relates to a leaky coaxial cable clamp fault detection method based on deep learning. Background Art

[0002] Leaky coaxial cable (hereinafter referred to as: leaky cable) is an important component of railway wireless communication. Its function is to radiate and receive electromagnetic waves evenly along the line. It is mainly deployed in railway tunnels, curves and other areas where wireless signals are difficult to cover, to achieve coverage of electromagnetic field blind spots. Due to the wind tunnel effect generated when the train is running at high speed, the leaky cable and fixed components are prone to loosening and falling off, posing a serious hidden danger to driving safety. At present, the inspection and inspection of leaky cables along the railway are carried out manually once a month. The manual inspection method has low inspection efficiency and high labor intensity for operators; because it is operated during the night "window" (high-speed rail and intercity lines), the lighting conditions are poor and the detection rate of hidden dangers and faults is low; and some locations are located in long tunnels and wilderness areas, which poses a hidden danger to the personal safety of operators.

[0003] In order to replace manual walking inspections of leaky coaxial cables along the railway, vehicle-mounted image acquisition equipment is currently used to collect images of leaky coaxial cables along the railway. However, the existing image acquisition and analysis system can only collect video data of one cable of the public network or private network at the same time when taking pictures, and cannot collect video data of two cables on one side of the tunnel at the same time. The existing image acquisition and fault identification system has the following main disadvantages:

[0004] (1) Only one cable of the public network or private network can be collected at the same time, and it is impossible to collect video data of both cables on one side of the tunnel at the same time;

[0005] (2) The changes in the location and environment of the leaky cable (both inside and outside the tunnel) cause the collected images to be either dim or overexposed.

[0006] (3) There are deviations in the standard heights of the cable installation in each tunnel on different lines, the field of view of the acquisition camera is insufficient, and the height position of the cable installation is offset, resulting in missed shots;

[0007] (4) The acquisition camera configuration is insufficient and the shooting frame rate is low, resulting in the omission of some areas and failing to meet the shooting requirements at the highest speed of the high-speed rail;

[0008] (5) Tunnel positioning can only rely on staff to view the collected video frame by frame, and cannot intelligently locate the tunnel area, which leads to large positioning deviation of the faulty fixture;

[0009] (6) The fault identification algorithm is not advanced enough, resulting in problems such as low fixture identification accuracy and slow processing speed. Summary of the invention

[0010] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for detecting leakage coaxial cable clamp faults based on deep learning, which adopts an image classification algorithm based on deep learning to realize the classification of images inside and outside the tunnel, and adopts an improved Yolov8 to realize the detection of clamp faults in the tunnel, thereby greatly reducing the workload of detection personnel and increasing the accuracy of tunnel detection.

[0011] The object of the present invention is to achieve the following technical solution: a leaky coaxial cable fixture fault detection method based on deep learning, comprising the following steps:

[0012] S1. Two high-speed, low-light industrial cameras are used for image acquisition. The cameras are installed on the windows of the train and collect images as the train moves. One horizontal camera captures images of the public network leaky cable, and the other tilted camera captures images of the private network leaky cable.

[0013] S2. Preprocess the collected images: filter the image data inside and outside the tunnel, and select the images that meet the training requirements of the ResNet50 network model; classify the filtered images into two categories, namely, inside the tunnel and outside the tunnel, numbered as 0 and 1 respectively; use the numbered images as the training sample set of the tunnel classification model;

[0014] S3. Build a ResNet50 network model to classify and identify the collected images and determine whether the images are inside or outside the tunnel; input the training sample set into the ResNet50 network model for model training;

[0015] S4, perform intelligent identification and positioning of tunnels: input the collected videos to be identified into the ResNet50 network model frame by frame in the time sequence of image acquisition, and determine whether each image is inside or outside the tunnel;

[0016] If an image is taken inside a tunnel, the first 50 consecutive images are combined to determine whether the image is the first image taken inside the tunnel. The first image taken inside the tunnel is taken as the tunnel entrance image, and the position information of the frame image in the acquired video is recorded.

[0017] Continue to judge the subsequent images of the tunnel entrance image. If a certain image is judged to be outside the tunnel, combine the first 50 consecutive adjacent images to determine whether the image is the first image outside the tunnel; take the first image outside the tunnel as the tunnel exit image, and record the position information of the frame image in the collected video;

[0018] The amount of image frame data contained in each tunnel is obtained by the position of the tunnel entrance image and the tunnel exit image in the video. The tunnel length is then obtained by combining the vehicle speed information obtained by GPS during the acquisition process and the camera frame rate.

[0019] Compare the length of each tunnel recorded in the tunnel kilometer mark file with the calculated tunnel length to confirm which tunnel the captured video image belongs to;

[0020] S5. Training target detection model: The Yolov8-Retinexformer target detection model is used to realize the accurate identification and location of the leaky cable fault fixture and fault; the Yolov8-Retinexformer target detection model adds a Retinexformer module after the input end of the traditional Yolov8 network model. The Retinexformer module includes an illumination estimator and a degradation repairer connected in sequence;

[0021] The specific processing process of the Retinexformer module is as follows:

[0022] Input image I and calculate the image's illumination prior value L p =main c (I), main c Indicates that the average value of each pixel is calculated along the channel dimension;

[0023] The input image I and the illumination prior value L p Input into the lighting estimation ε to generate the lighting map I lu and illuminate feature F lu ;

[0024] The degradation repairer adopts an illumination-guided transformer, which includes multiple illumination-guided attention blocks. The input illumination map I lu , it passes through a 3×3 convolution with a stride of 2, an IGAB, a 4×4 convolution with a stride of 2, two IGABs and a 4×4 convolution with a stride of 2 to generate the hierarchical feature F i ;

[0025] The illumination-guided attention block consists of two layer normalization LNs, an illumination-guided multi-head self-attention IG-MSA, and a feed-forward network FFN;

[0026] Select images with faulty fixtures from the image set in the tunnel, use LabelImage to standardize various fixtures in the image, and after standardization, add them to the training sample set; input the training sample set into Yolov8-Retinexformer for training;

[0027] S6. Use the target detection model to identify and locate the faulty fixture: The tunnel images identified by the ResNet50 network model are input into the target detection model in chronological order. The target detection model detects each frame of the tunnel image to determine whether there is a faulty fixture in the image. If so, the position of the faulty fixture detection frame in the image is recorded and the image information is stored. If not, it is skipped without any processing.

[0028] S7. Combining the vehicle speed information, the camera frame rate and the position information of the image in the video obtained by the GPS during the acquisition process, the kilometer mark position of the faulty fixture in the tunnel is inferred.

[0029] The beneficial effects of the present invention are:

[0030] 1. The present invention uses an image classification algorithm based on deep learning to classify images inside and outside the tunnel, extracts a set of all images inside the tunnel from the collected images, and then uses an improved Yolov8 to detect fixture faults in the tunnel, and then combines the vehicle speed information and the camera frame rate to obtain the location of the faulty fixture. This greatly reduces the workload of the inspection personnel, increases the accuracy of tunnel inspection, and shortens the time for data analysis.

[0031] 2. The present invention adopts an improved Yolov8 algorithm, which combines the estimation and loss recovery of lighting information. On the basis of the Retinex-based framework (ORF), an illumination-guided transformer (IGT) is designed. The illumination information is used to guide the modeling of non-local interactions in regions under different lighting conditions, which effectively enhances low-light images and improves the quality of low-light images while maintaining the natural appearance and quality of the images. The recognition accuracy and processing speed of the algorithm are better than those of other similar target detection algorithms, which makes the present invention suitable for the identification and positioning of faulty fixtures in low-light environments, and improves the target recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flow chart of a leaky coaxial cable clamp fault detection method based on deep learning of the present invention;

[0033] Figure 2 It is a structural diagram of the ResNet50 network of the present invention;

[0034] Figure 3 Schematic diagram of the structure of the Retinexformer module of the present invention. DETAILED DESCRIPTION

[0035] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0036] like Figure 1As shown, a leaky coaxial cable clamp fault detection method based on deep learning of the present invention comprises the following steps:

[0037] S1. Two high-speed low-light industrial cameras are used for image acquisition. The cameras are installed on the windows of the train and acquire images as the train moves. One horizontal camera captures images of the public network leaky cable, and the other tilted camera captures images of the private network leaky cable. The cameras start to acquire images from the departure station and will not pause or interrupt in the middle, so the acquired images include images inside the tunnel and images outside the tunnel. Since the present invention only needs to analyze the images inside the tunnel, the intelligent tunnel positioning algorithm is subsequently used to classify the images inside the tunnel and images outside the tunnel, and the images inside the tunnel are extracted for subsequent processing.

[0038] There are two types of leaky coaxial cables in the tunnel, public network and private network. The public network is installed 2 meters below and the private network is installed 4.7 meters above.

[0039] The two cameras in the device cannot move left and right, but support up and down rotation to fine-tune the shooting angle. The present invention supports the modification of camera parameters (exposure time, gain, brightness, saturation, gamma transformation, frame rate, etc.) without pausing the acquisition program and takes effect in real time, so that the brightness of the collected data is balanced and the clarity is high under various environments and lighting conditions, meeting the requirements of later image processing and analysis. Due to the deviation in the standard height of the cable installation in each tunnel in different lines, and the limitation of the field of view (horizontal and vertical directions) of the acquisition camera lens itself, the private network cable of some tunnels is missed; considering that the horizontal viewing angle of a general camera is greater than the vertical viewing angle, when the horizontal field of view meets the shooting requirements, try to choose a lens with a large vertical viewing angle, which solves the problem of insufficient vertical viewing angle. The acquisition camera selects a high-speed, low-light industrial camera with a maximum frame rate of 80 frames, which meets the train scene with a speed of 120 to 350 km / h.

[0040] S2. Preprocess the collected images: filter the image data inside and outside the tunnel, and select images that meet the training requirements of the ResNet50 network model (clear images, appropriate brightness, no overexposure, no darkening; clear outline of leaky cables; rich image texture features); classify the filtered images into two categories, namely, inside the tunnel and outside the tunnel, numbered 0 and 1 respectively; use the numbered images as the training sample set of the tunnel classification model;

[0041] S3. Build a ResNet50 network model to classify and identify the collected images and determine whether the images are inside or outside the tunnel. Input the training sample set into the ResNet50 network model for model training. The structure of the ResNet50 (image classification algorithm based on deep learning) network model is as follows: Figure 2As shown in the figure, during the model training process, it is necessary to tune the parameters according to the training results, optimize the training results, and finally obtain the most effective model.

[0042] S4, perform intelligent identification and positioning of tunnels: input the collected videos to be identified into the ResNet50 network model frame by frame in the time sequence of image acquisition, and determine whether each image is inside or outside the tunnel;

[0043] If an image is taken inside a tunnel, the first 50 consecutive images are combined to determine whether the image is the first image taken inside the tunnel. The first image taken inside the tunnel is taken as the tunnel entrance image, and the position information of the frame image in the acquired video is recorded.

[0044] Continue to judge the subsequent images of the tunnel entrance image. If an image is judged to be outside the tunnel, combine the first 50 adjacent continuous images to determine whether the image is the first image outside the tunnel; take the first image outside the tunnel as the tunnel exit image, and record the position information of the frame image in the captured video.

[0045] The amount of image frame data contained in each tunnel is obtained by the position of the tunnel entrance image and the tunnel exit image in the video. The tunnel length is then obtained by combining the vehicle speed information obtained by GPS during the acquisition process and the camera frame rate.

[0046] The length of each tunnel recorded in the tunnel kilometer mark file is compared with the calculated tunnel length to confirm which tunnel the captured video image belongs to.

[0047] S5. Training target detection model: The Yolov8-Retinexformer (target detection algorithm based on low-light image enhancement) target detection model is used to realize the accurate identification and positioning of leaky cable fault fixtures and faults; the Yolov8-Retinexformer target detection model adds a Retinexformer (low-light image enhancement) module after the input end of the traditional Yolov8 network model; the Retinexformer module includes an illumination estimator and a degradation repairer connected in sequence, and the specific structure is as follows Figure 3 As shown. The algorithm combines the estimation and loss recovery of lighting information. Based on the Retinex-based framework (ORF), an illumination-guided transformer (IGT) is designed. The illumination information is used to guide the modeling of non-local interactions of regions under different lighting conditions, which effectively enhances low-light images and improves the quality of low-light images while maintaining the natural appearance and seagate of the image. The recognition accuracy and processing speed of this algorithm are better than other similar target detection algorithms. The specific processing process of the Retinexformer module is:

[0048] Input image: Input a low-light image I∈RH×W×3 , and calculate the image's illumination prior value L p =main c (I), L p ∈R H×W , main c Indicates that the average value of each pixel is calculated along the channel dimension;

[0049] The input image I and the illumination prior value L p Input into the lighting estimation ε to generate the lighting map I lu ∈R H×W×3 and illuminate feature F lu ; The calculation formula is (I lu , F lu )=ε(I,L p );

[0050] The degradation repairer adopts an illumination-guided transformer (IGT), which includes multiple illumination-guided attention blocks (IGABs). The illumination-guided attention blocks use illumination features to guide the attention mechanism and gradually restore the image quality. The input of IGT is the illumination map I lu ∈R H×W×3 , it passes through a 3×3 convolution with a stride of 2, an IGAB, a 4×4 convolution with a stride of 2, two IGABs and a 4×4 convolution layer with a stride of 2 to generate hierarchical features

[0051] The illumination-guided attention block (IGAB) consists of two layer normalizations (LN), an illumination-guided multi-head self-attention (IG-MSA), and a feed-forward network (FFN). IG-MSA is a key component of IGAB, which guides self-attention calculations through illumination information to capture complex image details, and can use illumination information as a key clue to guide the capture of long-range dependencies.

[0052] Final image output: After layer-by-layer processing, the final output is an enhanced image, which has significantly improved quality, and color distortion and noise are effectively controlled.

[0053] First, build the Yolov8-Retinexformer development environment: a. Get the YOLOv8 source code: Visit the YOLOv8 official GitHub repository and download the source code. b. Install CUDA and CUDNN: Download and install CUDA and CUDNN according to the highest CUDA version supported by the graphics card driver. c. Install Anaconda: Download and install Anaconda to create and manage Python virtual environments. d. Create a virtual environment: Use Anaconda to create a Python virtual environment named YOLOv8 and install dependent libraries such as PyTorch. e. Verify the environment: Verify that CUDA, PyTorch and other environments are installed successfully by running simple test code.

[0054] Then, images with faulty fixtures are selected from the image set in the tunnel, and LabelImage is used to standardize various fixtures in the image. After standardization, they are added to the training sample set. The training sample set is input into the Yolov8-Retinexformer algorithm program for training; during the model training process, parameters need to be tuned according to the training results to optimize the training results and finally obtain the most effective model.

[0055] S6. Use the target detection model to identify and locate the faulty fixture: The tunnel images identified by the ResNet50 network model are input into the target detection model in chronological order. The target detection model detects each frame of the tunnel image to determine whether there is a faulty fixture in the image. If so, the position of the faulty fixture detection frame in the image is recorded and the image information is stored. If not, it is skipped without any processing.

[0056] S7. Combining the vehicle speed information, the camera frame rate and the position information of the image in the video obtained by the GPS during the acquisition process, the kilometer mark position of the faulty fixture in the tunnel is inferred.

[0057] The intelligent tunnel positioning algorithm introduced above is used to extract the image set of all the tunnels that the collected line passes through; the intelligently identified tunnel group is matched with the pre-prepared kilometer mark file (provided by the railway customer) to know the kilometer mark positions of the entrance and exit of each tunnel in the collected data; after matching the tunnels, the faulty wire clamps are detected for the images in each tunnel. After finding the faulty wire clamps, the kilometer mark position of the current frame image is calculated based on the frame rate of the camera, the average vehicle speed during the acquisition, and the position information of the image in the video.

[0058] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A leaky coaxial cable fixture fault detection method based on deep learning, characterized in that: The following steps are involved: S1. Two high-speed, low-light industrial cameras are used for image acquisition. The cameras are installed on the windows of the train and collect images as the train moves. One horizontal camera captures images of the public network leaky cable, and the other tilted camera captures images of the private network leaky cable. S2. Preprocess the collected images: filter the image data inside and outside the tunnel, and select the images that meet the training requirements of the ResNet50 network model; classify the filtered images into two categories, namely, inside the tunnel and outside the tunnel, numbered as 0 and 1 respectively; use the numbered images as the training sample set of the tunnel classification model; S3. Build a ResNet50 network model to classify and identify the collected images and determine whether the images are inside or outside the tunnel; Input the training sample set into the ResNet50 network model for model training; S4, perform intelligent identification and positioning of tunnels: input the collected videos to be identified into the ResNet50 network model frame by frame in the time sequence of image acquisition, and determine whether each image is inside or outside the tunnel; If an image is taken inside a tunnel, the previous 50 consecutive images are combined to determine whether the image is the first image taken inside a tunnel; The first tunnel image is used as the tunnel entrance image, and the position information of the frame image in the acquired video is recorded; Continue to judge the subsequent images of the tunnel entrance image. If a certain image is judged to be outside the tunnel, combine the first 50 consecutive adjacent images to determine whether the image is the first image outside the tunnel; take the first image outside the tunnel as the tunnel exit image, and record the position information of the frame image in the collected video; The amount of image frame data contained in each tunnel is obtained by the position of the tunnel entrance image and the tunnel exit image in the video. The tunnel length is then obtained by combining the vehicle speed information obtained by GPS during the acquisition process and the camera frame rate. Compare the length of each tunnel recorded in the tunnel kilometer mark file with the calculated tunnel length to confirm which tunnel the captured video image belongs to; S5. Training target detection model: The Yolov8-Retinexformer target detection model is used to realize the accurate identification and location of the leaky cable fault fixture and fault; the Yolov8-Retinexformer target detection model adds a Retinexformer module after the input end of the traditional Yolov8 network model. The Retinexformer module includes an illumination estimator and a degradation repairer connected in sequence; Select images with faulty fixtures from the image set in the tunnel, use LabelImage to standardize various fixtures in the image, and after standardization, add them to the training sample set; input the training sample set into Yolov8-Retinexformer for training; S6. Use the target detection model to identify and locate the faulty fixture: The tunnel images identified by the ResNet50 network model are input into the target detection model in chronological order. The target detection model detects each frame of the tunnel image to determine whether there is a faulty fixture in the image. If so, the position of the faulty fixture detection frame in the image is recorded and the image information is stored. If not, it is skipped without any processing. S7. Combining the vehicle speed information, the camera frame rate and the position information of the image in the video obtained by the GPS during the acquisition process, the kilometer mark position of the faulty fixture in the tunnel is inferred.

2. A leaky coaxial cable fixture fault detection method based on deep learning according to claim 1, characterized in that: The specific processing process of the Retinexformer module is as follows: Input image I and calculate the image's illumination prior value L p =main c (I), main c Indicates that the average value of each pixel is calculated along the channel dimension; The input image I and the illumination prior value L p Input into the lighting estimation ε to generate the lighting map I lu and illuminate feature F lu ; The degradation repairer adopts an illumination-guided transformer, which includes multiple illumination-guided attention blocks. The input illumination map I lu , it passes through a 3×3 convolution with a stride of 2, an IGAB, a 4×4 convolution with a stride of 2, two IGABs and a 4×4 convolution with a stride of 2 to generate the hierarchical feature F i ; The illumination-guided attention block consists of two layer normalization LNs, an illumination-guided multi-head self-attention IG-MSA, and a feed-forward network FFN.

Images