Intelligent damage detection method for trackside equipment

Abstract

The invention relates to the technical field of intelligent damage detection of trackside equipment, and discloses an intelligent damage detection method for the trackside equipment, which comprises the following steps of: 1, acquiring trackside multi-angle images in real time through an anti-shake vehicle-mounted camera array; 2, constructing a track chain closed-loop matching system based on the positioning data; 3, performing damage detection by adopting a semi-supervised multi-field learning algorithm of a fused generative adversarial network (GAN); 4, performing model compression deployment; 5, damage judgment; and 6, transmitting the damaged data to a central server through a 5G private network. According to the intelligent damage detection method for the trackside equipment, through multi-source data synchronous acquisition of an anti-shake vehicle-mounted camera array, an RTK-GPS (with the precision of + / -0.5 cm) and an IMU (Inertial Measurement Unit) and in combination with a track chain closed-loop matching system (a dynamic window algorithm + RANSAC (Random Sample Consensus) denoising), accurate positioning of the equipment under 160 km / h high-speed operation is realized, and the efficiency is improved by 128% compared with manual inspection (less than or equal to 70 km / h); the invention relates to a calling mechanism of three-dimensional equipment template library 400,000 + labeled samples.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent damage detection of trackside equipment, and in particular to an intelligent damage detection method for trackside equipment. Background Art

[0002] Wayside equipment refers to all fixed infrastructure equipment installed on the ground on both sides of the railway tracks (including the main line, station lines, depots, parking lots and other areas) that serves the functions of train operation, control, safety, power supply and maintenance guarantee. It covers almost all fixed ground equipment on the railway line except the trains themselves and track-operated vehicles.

[0003] Minor damage or potential failures of trackside equipment (such as rails, switches, signals, contact lines, track circuits, etc.) are difficult to detect through manual inspections in the early stages. Intelligent detection systems (such as those based on computer vision, acoustic sensing, vibration sensing, lidar and other technologies) can identify defects such as cracks, wear, deformation, looseness, dirt, missing parts, etc. that are difficult to detect with the naked eye at high frequency and high precision. They can effectively prevent major traffic safety accidents such as derailment, collisions, signal errors, contact line short circuits / breaks caused by equipment failure, and maximize the safety of life and property of passengers and staff and the smooth operation of the lines.

[0004] Traditional inspection methods have significant defects:

[0005] (1) Low efficiency of manual inspection: The inspection speed is slow (≤70km / h), the labor cost is high (annual cost exceeds 2.6 billion yuan), and it is easy to miss inspections due to complex environments and diverse equipment (such as screw detachment, cable corrosion and other minor damage);

[0006] (2) Existing intelligent detection technologies are insufficient: Machine learning-based solutions are constrained by sample scarcity (track data is confidential), and training models are not robust enough to handle rainy, snowy weather or camera shake. Existing automated systems (such as Hangzhou Shenhao Technology's inspection robots) have high false alarm rates (missed detection rates > 30%) and are unable to process data in real time (detection delays are as high as minutes).

[0007] (3) Edge deployment bottleneck: Traditional algorithms are difficult to run efficiently on embedded devices. The model computational load is large and cannot meet the 160km / h high-speed inspection requirements. Therefore, it is necessary to propose an intelligent damage detection method for trackside equipment to solve the above problems. Summary of the Invention

[0008] (1) Technical problems solved

[0009] In view of the shortcomings of the existing technology, the present invention provides an intelligent damage detection method for trackside equipment, which has the advantages of improving the inspection efficiency and inspection quality of trackside equipment.

[0010] (2) Technical solution

[0011] To achieve the above objectives, the present invention provides the following technical solutions: a method for intelligent damage detection of trackside equipment, comprising:

[0012] Step 1: Use the anti-shake vehicle-mounted camera array to collect real-time multi-angle images of the trackside, and simultaneously obtain RTK-GPS positioning data (accuracy ±0.5cm) and inertial measurement unit (IMU) attitude data;

[0013] Step 2: Build a closed-loop tracking system based on the positioning data: Use the Gaussian mixture model to filter the positioning drift points and retrieve the 3D equipment template at the corresponding location from the trackside equipment defect sample library;

[0014] Step 3: Use a semi-supervised multi-domain learning algorithm that integrates a generative adversarial network (GAN) for damage detection:

[0015] Align the cross-domain feature distribution of the template and the real-time image through the domain adaptation module (including the gradient reversal layer GRL);

[0016] Coarse positioning stage: Use the improved RetinaNet network (introducing deformable convolution DCNv2) to locate the device and output the device bounding box and confidence level;

[0017] Fine detection stage: Generate pixel-level damage masks through the GhostNet-CA semantic segmentation network and integrate the DenseASPP module to enhance context perception;

[0018] Step 4: Model compression and deployment:

[0019] Layer pruning (pruning rate 38%) and channel pruning (pruning rate 45%) were performed on the object detection and segmentation networks respectively;

[0020] The category boundary knowledge of the teacher model is transferred to the student model through the PoKD technology, and the KL divergence is used to constrain the feature graph similarity;

[0021] Deploy to the Jetson Xavier edge computing platform and load the TensorRT engine to implement INT8 quantized reasoning;

[0022] Step 5: Damage determination:

[0023] Cable damage is analyzed using the aspect ratio analysis of the convex hull (threshold > 1:50) and cross-validation with the Hough line detection.

[0024] For screw falling off, SIFT feature point matching and convex hull extreme point distance comparison are used (error tolerance ±1.5px);

[0025] Step 6: Transmit damage data to the central server via the 5G private network, generate a three-dimensional disease heat map and trigger a real-time alarm.

[0026] Preferably, the semi-supervised multi-domain learning algorithm in step (3) specifically includes:

[0027] Build a cross-domain database containing normal and abnormal samples;

[0028] Independent multi-domain networks are trained for different device types to enhance the robustness of small sample disease recognition.

[0029] Preferably, the model compression method in step (4) includes:

[0030] Perform structured pruning on the target detection model and semantic segmentation network respectively;

[0031] The PoKD technology is used to transfer the knowledge of the unpruned teacher model to the pruned student model to compensate for the accuracy loss.

[0032] Preferably, the semantic segmentation network uses GhostNet-CA module to extract multi-scale features and integrates DenseASPP to enhance high-level semantic features. Preferably, the damage identification in step (3) uses customized algorithms for different devices:

[0033] Cable damage is identified through aspect ratio filtering and contour detection;

[0034] Screw loss is located using a pole comparison algorithm.

[0035] Preferably, the trackside equipment disease sample library is constructed by the "National Key Laboratory of Advanced Rail Transit Autonomous Operation" of Beijing Jiaotong University, containing more than 400,000 labeled data, and the edge server is an NVIDIA Jetson series, supporting real-time data transmission at a speed of 160km / h.

[0036] Preferably, the trajectory chain closed-loop matching system of step (2) adopts a sliding window dynamic matching algorithm:

[0037] The window size is dynamically adjusted with the vehicle speed (formula: v is the vehicle speed, t proc is the processing delay, δ is the redundancy coefficient);

[0038] The RANSAC algorithm is used to eliminate GPS jump point interference.

[0039] Preferably, the KL divergence constraint is specifically:

[0040] Minimize the distribution difference between the teacher model and the student model on the channel attention map (CA-Module). The loss function is: Among them (T i ||S i ) is the feature map of the i-th layer of the teacher / student model, and λ is the weight decay coefficient.

[0041] Preferably, the domain adaptation module of step (3) dynamically adjusts the domain discriminator weights through a meta-learning strategy: constructs a meta-training task set to simulate domain shift scenarios such as sudden illumination changes and rain and fog interference; uses the MAML algorithm to optimize the initial parameters of the discriminator to enhance the generalization ability of the model;

[0042] When deploying the edge computing platform, a dual-model redundant backup mechanism is adopted:

[0043] The main model is the pruned student network, and the backup model is a lightweight MobileNetv3 segmentation network;

[0044] When the confidence level of the main model is less than 85%, the backup model is automatically switched to ensure system availability.

[0045] Preferably, the screw falling off detection further includes:

[0046] Screw position verification based on prior knowledge: The screw hole is located through Hough circle detection, and the screw hole is judged to be detached when the detection point is offset by more than 2 pixels from the reference position.

[0047] (3) Beneficial effects

[0048] Compared with the existing technology, the present invention provides an intelligent damage detection method for trackside equipment, which has the following beneficial effects:

[0049] 1. This intelligent damage detection method for trackside equipment uses a shake-resistant onboard camera array, RTK-GPS (accuracy ±0.5cm), and IMU to synchronously collect multi-source data. Combined with a track chain closed-loop matching system (dynamic window algorithm + RANSAC noise reduction), it achieves precise positioning of equipment at high speeds of 160km / h, improving efficiency by 128% compared to manual inspections (≤70km / h). The invocation mechanism of a 3D equipment template library (400,000+ annotated samples) solves the problem of missed detection caused by the diverse equipment of traditional methods, covering all types of trackside equipment, including rails, signal machines, and contact lines.

[0050] 2. This intelligent damage detection method for trackside equipment uses a semi-supervised multi-domain learning algorithm (including GAN adversarial training + gradient reversal layer) to align cross-domain features through a domain adaptation module. Combined with meta-learning, it dynamically adjusts the discriminator weights (simulating rain, fog, and sudden changes in illumination), significantly improving the model's robustness in sample-scarce scenarios and reducing the false alarm rate to <10% (compared to >30% for traditional methods). A two-stage detection mechanism (improved RetinaNet coarse positioning + GhostNet-CA fine segmentation) fuses deformable convolution (DCNv2) with a contextual enhancement module (DenseASPP) to achieve pixel-level damage recognition (such as microcracks in cables with an aspect ratio >1:50 and 1.5px screw offsets).

[0051] 3. This intelligent damage detection method for trackside equipment uses model compression technology (38% pruning of the target detection layer and 45% pruning of the segmentation network channels) combined with PoKD knowledge distillation (KL divergence constrained channel attention map) to compress computing latency to milliseconds through INT8 quantization (TensorRT engine) on the Jetson Xavier platform. A dual-model redundant backup mechanism (switching to the MobileNetv3 backup model when the confidence level of the main model is <85%) ensures system availability in extreme scenarios such as high speed and low light, meeting the real-time processing requirements of 160 km / h. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a structural flow chart of the detection method of the present invention. DETAILED DESCRIPTION

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0054] See also Figure 1 , an intelligent damage detection method for trackside equipment, comprising:

[0055] Step 1: Use the anti-shake vehicle-mounted camera array to collect real-time multi-angle images of the trackside, and simultaneously obtain RTK-GPS positioning data (accuracy ±0.5cm) and inertial measurement unit (IMU) attitude data;

[0056] Step 2: Build a closed-loop tracking system based on the positioning data: Use the Gaussian mixture model to filter the positioning drift points and retrieve the 3D equipment template at the corresponding location from the trackside equipment defect sample library;

[0057] Step 3: Use a semi-supervised multi-domain learning algorithm that integrates a generative adversarial network (GAN) for damage detection:

[0058] Align the cross-domain feature distribution of the template and the real-time image through the domain adaptation module (including the gradient reversal layer GRL);

[0059] Coarse positioning stage: Use the improved RetinaNet network (introducing deformable convolution DCNv2) to locate the device and output the device bounding box and confidence level;

[0060] Fine detection stage: Generate pixel-level damage masks through the GhostNet-CA semantic segmentation network and integrate the DenseASPP module to enhance context perception;

[0061] Step 4: Model compression and deployment:

[0062] Layer pruning (pruning rate 38%) and channel pruning (pruning rate 45%) were performed on the object detection and segmentation networks respectively;

[0063] The category boundary knowledge of the teacher model is transferred to the student model through the PoKD technology, and the KL divergence is used to constrain the feature graph similarity;

[0064] Deploy to the Jetson Xavier edge computing platform and load the TensorRT engine to implement INT8 quantized reasoning;

[0065] Step 5: Damage determination:

[0066] Cable damage is analyzed using the aspect ratio analysis of the convex hull (threshold > 1:50) and cross-validation with the Hough line detection.

[0067] For screw falling off, SIFT feature point matching and convex hull extreme point distance comparison are used (error tolerance ±1.5px);

[0068] Step 6: Transmit damage data to the central server via the 5G private network, generate a three-dimensional disease heat map and trigger a real-time alarm.

[0069] In the case implementation, the system deployment configuration:

[0070] (1) Hardware Platform

[0071] Inspection vehicle: CRH380BJ-0503 high-speed comprehensive inspection train

[0072] Sensors: 8 Sony IMX585 anti-shake cameras (frame rate 60fps) + NovAtel RTK-GPS (positioning accuracy ±0.5cm) + ADIS16470 IMU

[0073] Edge device: NVIDIA Jetson Xavier NX (with TensorRT 8.4)

[0074] (2) Software environment

[0075] Algorithm framework: PyTorch1.10+OpenCV4.5

[0076] Communication protocol: 5G private network (uplink bandwidth ≥ 200Mbps).

[0077] In the case implementation, the anti-shake camera array was innovatively deployed: forward array: 4 Sony IMX585 cameras (35mm focal length, covering rails / turnouts); side array: 4 Basler ac A4112 cameras (50mm focal length, covering contact network / signals); anti-shake mechanism: 6-axis mechanical gimbal + electronic image stabilization (EIS), compensating for ±3° vibration.

[0078] In the case implementation, through the synchronous collection of multi-source data from the anti-shake vehicle-mounted camera array, RTK-GPS (accuracy ±0.5cm), and IMU, combined with the trajectory chain closed-loop matching system (dynamic window algorithm + RANSAC noise reduction), precise positioning of equipment at a high speed of 160km / h was achieved, which is 128% more efficient than manual inspection (≤70km / h); the calling mechanism of the 3D equipment template library (400,000 + annotated samples) solves the problem of missed detection caused by the diversity of equipment in traditional methods, covering all categories of trackside equipment such as rails, signal machines, and contact lines.

[0079] In the case implementation, the semi-supervised multi-domain learning algorithm (including GAN adversarial training + gradient reversal layer) aligns cross-domain features through the domain adaptation module, and combines meta-learning to dynamically adjust the discriminator weights (simulating rain and fog / light mutations), significantly improving the robustness of the model in sample-scarce scenarios, and reducing the false alarm rate to <10% (traditional methods >30%); the two-stage detection mechanism (improved RetinaNet coarse positioning + GhostNet-CA fine segmentation) integrates deformable convolution (DCNv2) and context enhancement module (DenseASPP) to achieve pixel-level damage recognition (such as micro cracks with cable aspect ratio >1:50 and 1.5px screw offset).

[0080] In the case implementation, through model compression technology (38% pruning of the target detection layer + 45% pruning of the segmentation network channel) combined with PoKD knowledge distillation (KL divergence constrained channel attention map), the computing delay was compressed to milliseconds through INT8 quantization (TensorRT engine) on the Jetson Xavier platform; the dual-model redundant backup mechanism (switching to the MobileNetv3 backup model when the confidence of the main model is <85%) ensures the system's availability in extreme scenarios such as high speed and low light, and meets the real-time processing requirements of 160km / h.

[0081] In summary, this intelligent damage detection method for trackside equipment achieves precise positioning of equipment at high speeds of 160 km / h through the simultaneous acquisition of multi-source data from an anti-shake on-board camera array, RTK-GPS (accuracy ±0.5 cm), and IMU, combined with a trajectory chain closed-loop matching system (dynamic window algorithm + RANSAC noise reduction). This improves efficiency by 128% compared to manual inspections (≤70 km / h). The invocation mechanism of a 3D equipment template library (400,000 + annotated samples) solves the problem of missed detection caused by the diversity of equipment in traditional methods, covering all types of trackside equipment, including rails, signal machines, and contact lines.

[0082] In addition, the semi-supervised multi-domain learning algorithm (including GAN adversarial training + gradient reversal layer) aligns cross-domain features through a domain adaptation module, and combines meta-learning to dynamically adjust the discriminator weights (simulating rain and fog / illumination mutations), significantly improving the robustness of the model in sample-scarce scenarios, and reducing the false alarm rate to <10% (traditional methods >30%); the two-stage detection mechanism (improved RetinaNet coarse positioning + GhostNet-CA fine segmentation) integrates deformable convolution (DCNv2) and context enhancement module (DenseASPP) to achieve pixel-level damage recognition (such as microcracks with cable aspect ratio >1:50 and 1.5px screw offset).

[0083] In addition, through model compression technology (38% pruning of the target detection layer + 45% pruning of the segmentation network channel) combined with PoKD knowledge distillation (KL divergence constrained channel attention map), the computing delay is compressed to the millisecond level through INT8 quantization (TensorRT engine) on the Jetson Xavier platform; the dual-model redundant backup mechanism (switching to the MobileNetv3 backup model when the confidence of the main model is <85%) ensures the system's availability in extreme scenarios such as high speed and low light, and meets the real-time processing requirements of 160km / h.

[0084] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0085] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent damage detection of trackside equipment, characterized in that: include: Step 1: Use the anti-shake vehicle-mounted camera array to collect real-time multi-angle images of the trackside, and simultaneously obtain RTK-GPS positioning data (accuracy ±0.5cm) and inertial measurement unit (IMU) attitude data; Step 2: Build a closed-loop tracking system based on the positioning data: Use the Gaussian mixture model to filter the positioning drift points and retrieve the 3D equipment template at the corresponding location from the trackside equipment defect sample library; Step 3: Use a semi-supervised multi-domain learning algorithm that integrates a generative adversarial network (GAN) for damage detection: Align the cross-domain feature distribution of the template and the real-time image through the domain adaptation module (including the gradient reversal layer GRL); Coarse positioning stage: Use the improved RetinaNet network (introducing deformable convolution DCNv2) to locate the device and output the device bounding box and confidence level; Fine detection stage: Generate pixel-level damage masks through the GhostNet-CA semantic segmentation network and integrate the DenseASPP module to enhance context perception; Step 4: Model compression and deployment: Layer pruning (pruning rate 38%) and channel pruning (pruning rate 45%) were performed on the object detection and segmentation networks respectively; The category boundary knowledge of the teacher model is transferred to the student model through the PoKD technology, and the KL divergence is used to constrain the feature graph similarity; Deploy to the Jetson Xavier edge computing platform and load the TensorRT engine to implement INT8 quantized reasoning; Step 5: Damage determination: Cable damage is analyzed using the aspect ratio analysis of the convex hull (threshold > 1:50) and cross-validation with the Hough line detection. For screw falling off, SIFT feature point matching and convex hull extreme point distance comparison are used (error tolerance ±1.5px); Step 6: Transmit damage data to the central server via the 5G private network, generate a three-dimensional disease heat map and trigger a real-time alarm.

2. The intelligent damage detection method for trackside equipment according to claim 1, characterized in that: The semi-supervised multi-domain learning algorithm in step (3) specifically includes: Build a cross-domain database containing normal and abnormal samples; Independent multi-domain networks are trained for different device types to enhance the robustness of small sample disease recognition.

3. The intelligent damage detection method for trackside equipment according to claim 1, characterized in that: The model compression method in step (4) includes: Perform structured pruning on the target detection model and semantic segmentation network respectively; The PoKD technology is used to transfer the knowledge of the unpruned teacher model to the pruned student model to compensate for the accuracy loss.

4. The intelligent damage detection method for trackside equipment according to claim 1, characterized in that: The semantic segmentation network uses the GhostNet-CA module to extract multi-scale features and integrates DenseASPP to enhance high-level semantic features.

5. The intelligent damage detection method for trackside equipment according to claim 1, characterized in that: The damage identification in step (3) uses customized algorithms for different devices: Cable damage is identified through aspect ratio filtering and contour detection; Screw loss is located using a pole comparison algorithm.

6. The intelligent damage detection method for trackside equipment according to claim 1, characterized in that: The trackside equipment disease sample library was built by the "National Key Laboratory of Advanced Rail Transit Autonomous Operation" of Beijing Jiaotong University, containing more than 400,000 labeled data. The edge server is the NVIDIA Jetson series, which supports real-time data transmission at a speed of 160km / h.

7. The intelligent damage detection method for trackside equipment according to claim 1, characterized in that: The trajectory chain closed-loop matching system of step (2) adopts a sliding window dynamic matching algorithm: The window size is dynamically adjusted with the vehicle speed (formula: v is the vehicle speed, t proc is the processing delay, δ is the redundancy coefficient); The RANSAC algorithm is used to eliminate GPS jump point interference.

8. The intelligent damage detection method for trackside equipment according to claim 3 is characterized in that: The KL divergence constraint is specifically: Minimize the distribution difference between the teacher model and the student model on the channel attention map (CA-Module). The loss function is: Among them (T i ||S i ) is the feature map of the i-th layer of the teacher / student model, and λ is the weight decay coefficient.

9. The intelligent damage detection method for trackside equipment according to claim 1, characterized in that: The domain adaptation module of step (3) dynamically adjusts the domain discriminator weights through a meta-learning strategy: constructs a meta-training task set to simulate domain shift scenarios such as sudden illumination changes and rain and fog interference; uses the MAML algorithm to optimize the initial parameters of the discriminator to enhance the generalization ability of the model; When deploying the edge computing platform, a dual-model redundant backup mechanism is adopted: The main model is the pruned student network, and the backup model is a lightweight MobileNetv3 segmentation network; When the confidence level of the main model is less than 85%, the backup model is automatically switched to ensure system availability.

10. The intelligent damage detection method for trackside equipment according to claim 5, characterized in that: The screw falling off detection further includes: Screw position verification based on prior knowledge: The screw hole is located through Hough circle detection, and the screw hole is judged to be detached when the detection point is offset by more than 2 pixels from the reference position.

Images