A high-speed railway fagging defect detection method combined with an improved SSD algorithm

Abstract

The patent discloses a high-speed railway pendant defect detection method combined with an improved SSD algorithm, which can determine the defect condition with better detection accuracy and speed while the train is running, and is beneficial to the smooth operation of the high-speed railway contact network. First, the contact network picture taken by the railway 4C detection vehicle is preprocessed, such as denoising, so that the image can be better operated subsequently. Then, the positioning and identification of the pendant are carried out, and the improved Faster R-CNN algorithm can have higher positioning and identification accuracy. Finally, the fault of the contact network pendant is judged, and the improved SSD algorithm is combined to automatically determine the fault type. The patent reduces the loss of manpower and material resources, makes the detection data clearer, and improves the accuracy and pertinence of the contact network detection and monitoring.

Description

Technical Field

[0001] This invention belongs to the field of high-speed railway catenary technology, and in particular relates to a method for detecting defects in high-speed railway droppers by combining an improved SSD algorithm. Background Technology

[0002] The traction power supply system determines the safe and stable operation of electrified railways. The overhead contact line suspension status detection and monitoring device (4C system) developed in my country is truly important, serving as a crucial link in ensuring the quality of current collection from the pantograph to the contact line. With the rapid development of my country's high-speed railways and the continuous expansion of the railway network, the requirements for railway operation and management are even higher to ensure operational safety. The widespread adoption of overhead contact line suspension status detection and monitoring devices on high-speed railways can improve image acquisition efficiency.

[0003] In overhead contact line suspension systems, the dropper mechanism is a critical structural component. A failure in this mechanism can severely impact the current-collecting quality of the pantograph system, and may even lead to other major mechanical accidents such as pantograph failure and wire breakage. The dropper system is a key monitoring and inspection target in the railway 4C system. The 4C system primarily uses sensors to detect the main components of the train's overhead contact line, their internal structural data, and the positional parameters of related components. This allows for the detection of design defects in the train's contact and suspension structures, providing reliable guidance for daily inspections and maintenance. Therefore, a new generation of high-speed railway overhead contact line dropper defect analysis and detection system technology incorporating 3D image processing is of paramount importance. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a method for detecting defects in overhead contact line droppers in high-speed railways, incorporating an improved SSD algorithm. This method effectively detects defects in overhead contact line droppers in high-speed railways, ensuring the safe and reliable operation of trains.

[0005] The technical solution adopted in this invention is as follows: First, a railway 4C inspection vehicle takes pictures of the overhead contact line. Then, preprocessing operations such as noise reduction are performed on the images to make them clearer, thus facilitating subsequent operations. Next, the Faster R-CNN algorithm is used to locate and identify the droppers, which has higher accuracy in location identification. Finally, faults in the overhead contact line droppers are determined, and the fault type is automatically identified using an improved SSD algorithm. The process includes the following steps:

[0006] The S100 uses a high-definition camera on a railway 4C inspection vehicle to automatically capture images of the overhead contact line support suspension device. Then, it cleans and denoises the image data, filters out some irrelevant data, and removes environmental noise interference from the image, thereby improving the image clarity.

[0007] S200 performs positioning and identification of catenary droppers, and uses an improved Faster R-CNN target localization algorithm to generate more accurate regions, which can better improve the target detection accuracy.

[0008] The S300 can classify the status of the dropper wire into four types: normal, missing, broken strand, and slack. With the improved SSD algorithm, the status of the dropper wire of the contact network can be determined after the positioning is completed.

[0009] In the image denoising stage, the RR-DCT denoising acceleration method is adopted. Its specific steps are as follows: Let Ω be the set of all patches in the input image I, and convert the i-th patch into a frequency domain signal:

[0010] F i =ψ DCT (f i )

[0011] In the above formula, ψ DCT (.) represents a positive DCT function. If the DCT coefficients are less than a threshold m, they are considered noise and should be discarded.

[0012]

[0013] In the above formula, F i ' represents the processed coefficients, u represents the row vector in the processed coefficient matrix, and v represents the column vector in the processed coefficient matrix. Then, the coefficients F... i Convert to spatial domain signal:

[0014] f i '=ψ iDCT (F i ')

[0015] RR-DCT performs patch subsampling, and the value of pixel q in the output image is as follows:

[0016]

[0017] In the above formula, I'(q) represents the value at pixel q in the output image, w(q) represents a set of patches including pixel q, and map i (q) represents the mapping function between the corresponding center pixel q and the i-th patch at the corresponding position. This represents the summation of pixel values ​​for all patches;

[0018] The improved Faster R-CNN target localization algorithm is as follows: First, the input image is processed by CNN to obtain activation feature maps; second, feature maps at different levels are fused to form a unified feature space; next, a lightweight region proposal network is used to generate a certain number of candidate regions; finally, the candidate regions are classified and their boundaries are adjusted by the detection module to output the final detection result.

[0019] In the improved SSD algorithm, each prediction layer incorporates a residual module before classification and regression, and introduces deconvolutional layers whose size matches that of the auxiliary convolutional layers, so as to pass the fused feature map to the prediction module for object classification and boundary regression.

[0020] The beneficial effects of adopting this technical solution are:

[0021] 1. This invention uses an improved Faster R-CNN algorithm and an improved SSD algorithm to detect defects in droppers, which significantly improves the detection accuracy and efficiency of railway power supply systems and provides an important support for the safe and efficient operation of railways;

[0022] 2. This invention can effectively preprocess images, ensuring that details in the image do not become blurred after processing; moreover, it does not damage the image's outline and edges, guaranteeing clear and reliable edge features;

[0023] 3. Our company has invented and designed a set of detection algorithms for the dropper problem testing system, which enables relatively fast, accurate, and reliable intelligent identification of various problems such as dropper issues. The identification and detection accuracy is high, and many of the technologies have good versatility. Furthermore, it has the potential for real-time image processing, improving the efficiency and intelligence level of railway system inspections. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a high-speed railway dropper defect detection method combining an improved SSD algorithm according to the present invention.

[0025] Figure 2 This is a schematic diagram of the patch for the RR-DCT denoising method;

[0026] Figure 3 This is the overall framework of the improved Faster R-CNN;

[0027] Figure 4 This is a diagram of the improved SSD algorithm detection framework. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described below with reference to the accompanying drawings.

[0029] In this embodiment, see Figure 1 As shown, this invention proposes a method for detecting defects in overhead contact line droppers in high-speed railways, incorporating an improved SSD algorithm. The technical solution employed is as follows: a camera mounted on the top of a work vehicle captures images of the dropper section; the images are then pre-processed, including denoising, to improve image clarity and facilitate subsequent operations. Next, the dropper is located and identified using the Faster R-CNN algorithm, which offers higher accuracy. Finally, the faults in the overhead contact line droppers are assessed, and the improved SSD algorithm automatically identifies the fault type. The method includes the following steps:

[0030] The S100 uses a high-definition camera on a railway 4C inspection vehicle to automatically capture images of the overhead contact line support suspension device. Then, it cleans and denoises the image data, filters out some irrelevant data, and removes environmental noise interference from the image, thereby improving the image clarity.

[0031] S200 performs positioning and identification of catenary droppers, and uses an improved Faster R-CNN target localization algorithm to generate more accurate regions, which can better improve the target detection accuracy.

[0032] The S300 can classify the status of the dropper wire into four types: normal, missing, broken strand, and slack. With the improved SSD algorithm, the status of the dropper wire of the contact network can be determined after the positioning is completed.

[0033] During image preprocessing, the RR-DCT denoising acceleration method is used. Its specific steps are as follows: Let Ω be the set of all patches in the input image I, and convert the i-th patch into a frequency domain signal:

[0034] F i =ψ DCT (f i )

[0035] In the above formula, ψ DCT (.) represents a positive DCT function. If the DCT coefficients are less than a threshold m, they are considered noise and should be discarded.

[0036]

[0037] In the above formula, F i ' represents the processed coefficients, u represents the row vector in the processed coefficient matrix, and v represents the column vector in the processed coefficient matrix. Then, the coefficients F... i Convert to spatial domain signal:

[0038] f i '=ψ iDCT (F i ')

[0039] RR-DCT sampling of patch subsamples, such as Figure 2 The image shown is a schematic diagram of the patch. The value of pixel q in the output image is as follows:

[0040]

[0041] In the above formula, I'(q) represents the value at pixel q in the output image, w(q) represents a set of patches including pixel q, and map i (q) represents the mapping function between the corresponding center pixel q and the i-th patch at the corresponding position. This represents the summation of pixel values ​​for all patches.

[0042] In the string recognition and localization technology, RPN and Faster R-CNN techniques are integrated into the same network. Since the proposals generated by RPN during forward propagation are relatively fixed in each iteration of stochastic gradient descent, the accuracy of recognition can be greatly improved by pre-calculating them.

[0043] Faster R-CNN is a milestone in object detection technology. It significantly improves detection efficiency by integrating multiple steps, including candidate region generation, feature extraction, classification, and boundary regression, into a single deep learning network. Specifically, it employs a shared convolutional neural network to construct the region proposal network in the candidate region generation stage, which greatly accelerates object detection. Figure 3 As shown, the improved Faster R-CNN target localization algorithm can be summarized as follows: First, the input image is processed by a CNN to obtain activation feature maps; second, feature maps at different levels are fused to form a unified feature space; next, a lightweight region proposal network (SPR) is used to generate a certain number of candidate regions; finally, the candidate regions are classified and their boundaries are adjusted by the detection module to output the final detection result. To further improve the detection efficiency, a cascaded structure is used for state discrimination. The first step uses the improved Faster R-CNN algorithm to accurately locate the droppers in the acquired catenary image. The second step, based on the first step, uses the result of the first step as input to the improved SSD algorithm to determine the dropper fault state.

[0044] like Figure 4As shown, the improved SSD network adopts an "hourglass" structure, where each prediction layer incorporates a residual module before classification and regression, and introduces deconvolutional layers. The size of these deconvolutional layers matches the size of the auxiliary convolutional layers to pass the fused feature map to the prediction module for object classification and boundary regression. This change positively impacts the network's detection accuracy. The introduction of the residual prediction module aims to extract deeper-dimensional features for subsequent classification and regression tasks, thereby improving detection accuracy. This module is positioned after the deconvolutional structure and before classification and regression. Overall, the improved SSD network, by fusing high-level and low-level semantic information, fully utilizes contextual information and increases the resolution of the feature map, thus improving object detection accuracy, especially for small objects.

[0045] The above examples demonstrate the basic scientific principles, main methodological features, and various advantages of the basic invention method. All those skilled in the art should clearly understand that this invention is completely unrestricted by technical limitations. The descriptions in the above implementation diagrams and product specification drawings are merely for illustrating some principles of the invention. Without deviating from the scope of this invention, various variations and technical improvements will inevitably occur, all of which will ultimately benefit this invention and require protection. The appended claims and their equivalent boundary diagrams clearly define the effective protection and scope of this invention.

Claims

1. A method for detecting defects in high-speed railway droppers by combining an improved SSD algorithm, characterized in that, Including the following steps: The S100 uses a high-definition camera on a railway 4C inspection vehicle to automatically capture images of the overhead contact line support suspension device. Then, it cleans and denoises the image data, filters out irrelevant data, removes environmental noise interference from the image, and improves the image clarity. S200 performs positioning and identification of catenary droppers, and uses an improved Faster R-CNN target localization algorithm to generate more accurate regions and improve the target detection accuracy. The S300 classifies the status of the dropper wire into four types: normal, missing, broken strand, and slack. The improved SSD algorithm can determine the status of the dropper wire after positioning is completed. In the image denoising stage, the RR-DCT denoising acceleration method is adopted. Its specific steps are as follows: Let Ω be the set of all patches in the input image I, and convert the i-th patch into a frequency domain signal: F i =ψ DCT (f i ) In the above formula, ψ DCT (.) represents a positive DCT function. If the DCT coefficients are less than a threshold m, they are considered noise and should be discarded. In the above formula, F i ' represents the processed coefficients, u represents the row vector in the processed coefficient matrix, and v represents the column vector in the processed coefficient matrix. Then, the coefficients F... i Convert to spatial domain signal: f i '=ψ iDCT (F i ') RR-DCT performs patch subsampling, and the value of pixel q in the output image is as follows: In the above formula, I'(q) represents the value at pixel q in the output image, w(q) represents a set of patches including pixel q, and map i (q) represents the mapping function between the corresponding center pixel q and the i-th patch at the corresponding position. This represents the summation of pixel values ​​for all patches; The improved Faster R-CNN target localization algorithm is as follows: First, the input image is processed by CNN to obtain activation feature maps; second, feature maps at different levels are fused to form a unified feature space; next, a lightweight region proposal network is used to generate a certain number of candidate regions; finally, the candidate regions are classified and their boundaries are adjusted by the detection module to output the final detection result. In the improved SSD algorithm, each prediction layer incorporates a residual module before classification and regression, and introduces deconvolutional layers whose size matches that of the auxiliary convolutional layers, so as to pass the fused feature map to the prediction module for object classification and boundary regression.

Images