High-precision defect detection system for track inspection robot

Abstract

The invention discloses a high-precision defect detection system for a track inspection robot, which relates to the technical field of track inspection and comprises a multi-mode visual acquisition module, a time sequence synchronization module, a data fusion processing module and a defect identification output module, by adopting a multi-modal visual fusion detection technology, various modal visual technologies of visible light vision, infrared vision and ultraviolet vision are combined, fusion processing is carried out on all modal image data, visual defects such as cracks and abrasion on the surface of a track can be clearly recognized through visible light vision, and the visual performance of the track is improved. The infrared vision can detect temperature abnormity of a track component to find out potential poor contact and other problems, the ultraviolet vision can be used for detecting corona discharge and other phenomena, track information is comprehensively obtained, the accuracy and reliability of defect recognition are effectively improved, defects which are difficult to find in a single mode can be detected, and the detection efficiency is improved. And powerful support is provided for safe operation and maintenance of the track.

Description

Technical Field

[0001] The present invention relates to the technical field of track inspection, and in particular to a high-precision defect detection system for a track inspection robot. Background Art

[0002] Tracks are structural engineering facilities composed of rails, sleepers, connecting parts, and roadbeds. They are the core infrastructure of public transportation systems such as railways and urban rail transit, and play a key role in guiding train operations and transferring loads. Their safe and stable operation is directly related to transportation efficiency, the safety of passengers' lives and property, and social and economic order. They have irreplaceable strategic significance for ensuring the smooth operation of national infrastructure.

[0003] A rail inspection robot is an automated device equipped with multiple sensors. It conducts contactless inspections of tracks and ancillary facilities by autonomously traveling along the tracks or being mounted on inspection vehicles. Compared to traditional manual inspections, it overcomes the limitations of harsh environments, nighttime inspections, and long-distance inspections, significantly improving inspection efficiency and data collection continuity. As a core component of intelligent rail transit operations and maintenance systems, it plays a crucial role in promoting the evolution of rail inspection from manual labor to automated and intelligent processes.

[0004] Existing track inspection robots still have certain defects when performing track inspection. At present, track inspection robots often use single-modal visual inspection technology when detecting track defects, such as using only visible light vision for detection. However, single-modal vision technology has limitations. For example, visible light vision is difficult to detect temperature anomalies and corona discharges in track components, resulting in incomplete defect detection. The accuracy and reliability need to be improved. Therefore, it is of great significance to develop a high-precision defect detection system for track inspection robots. Summary of the Invention

[0005] The purpose of the present invention is to make up for the shortcomings of the existing technology and provide a high-precision defect detection system for track inspection robots. It can combine multiple modal vision technologies such as visible light vision, infrared vision, and ultraviolet vision, and fuse the image data of each modality through specific algorithms to obtain track information more comprehensively, effectively improve the accuracy and reliability of defect identification, and detect defects that are difficult to detect with a single modality, providing strong support for the safe operation and maintenance of the track.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: a high-precision defect detection system for a track inspection robot, the system comprising: a multimodal visual acquisition module, a timing synchronization module, a data fusion processing module and a defect recognition output module; The multimodal visual acquisition module includes a visible light camera, an infrared thermal imager, and an ultraviolet imager arranged in sequence along the robot's inspection direction, and is used to collect multimodal image data of the track, including visible light image data, infrared image data, and ultraviolet image data. The axis of the lens of each device forms a preset angle with the track detection surface; The timing synchronization module is electrically connected to the multimodal vision acquisition module and is used to trigger each device to collect image data of the corresponding detection area at the same inspection time; The data fusion processing module includes a preprocessing unit, a feature extraction unit and a fusion operation unit, which receives the multimodal image data after time sequence synchronization, and outputs the fusion result after preprocessing, feature extraction and fusion operation; The defect recognition output module has a built-in defect feature comparison library, and completes defect determination and output based on the fusion results.

[0007] Furthermore, in the multimodal vision acquisition module, the focal length of the visible light camera lens is adjusted by an electric focusing assembly, and the focusing parameters are preset mapped to the robot's travel speed. The camera has a built-in polarizer, and the polarization angle of the polarizer is continuously adjusted in the range of 0-90° by a stepper motor to eliminate the interference of reflective interference on the track surface. The detection field of view of the infrared thermal imager has a 30%-50% overlap area with that of the visible light camera to ensure cross-modal feature alignment. Its sampling frequency is not less than 1 / 2 of the sampling frequency of the visible light camera, the detector resolution is not less than 640×512 pixels, and the temperature measurement range covers -20°C to 150°C.

[0008] Furthermore, the timing synchronization module includes a laser ranging sensor and a synchronization controller. The laser ranging sensor obtains the distance value between the robot and the track detection point in real time. The synchronization controller dynamically adjusts the trigger interval of each visual device according to the distance value. The trigger signal adopts a TTL level signal, and the rising edge triggers the acquisition action.

[0009] Furthermore, the preprocessing unit of the data fusion processing module performs denoising, edge enhancement and distortion correction on the visible light image, performs temperature calibration and pseudo-color coding conversion on the infrared image, and performs dark current elimination and image enhancement on the ultraviolet image. After preprocessing, the pixel resolution of each modality image remains consistent. The feature extraction unit extracts the crack edge gradient feature, the grayscale mean value of the wear area and the texture entropy feature from the visible light image, wherein the texture entropy feature calculation formula is: ,in, is the probability distribution of grayscale histogram of the wear area, is the number of gray levels, is a minimum positive constant, , which is used to avoid zero-value input in logarithmic operations, extract the temperature field distribution entropy, high-temperature point clustering density and temperature gradient change rate characteristics of infrared images, and extract the circularity, area ratio and grayscale distribution standard deviation characteristics of corona spots of ultraviolet images.

[0010] Furthermore, the fusion operation unit of the data fusion processing module adopts a weighted feature fusion algorithm, and the fusion feature vector calculation formula is: ,in, 、 、 They are the feature weights of visible light, infrared, and ultraviolet, respectively, and are allocated based on historical data or modal importance. 、 、 are the corresponding modal eigenvectors, is the mask matrix of the overlapping area of ​​visible light and infrared field of view, is the mask matrix of the overlapping area of ​​ultraviolet and visible light fields of view, Represents element-level multiplication operation. The weight correction is based on the confusion matrix calculation of the effective detection data in the historical dataset. The correction period is consistent with the time it takes for the robot to complete a full-line inspection.

[0011] Furthermore, when the defect recognition output module is working, the fused feature data is first divided into several feature subsets according to the defect type. Each feature subset is matched one by one with the threshold range of the corresponding sub-library in the defect feature comparison library. The matching process uses a fuzzy logic algorithm to calculate the feature similarity. The similarity calculation formula is: ,in, is the feature parameter to be matched, is the corresponding threshold in the comparison library, is the feature weight, is the feature dimension, similarity When the preset threshold is exceeded, it is marked as a suspected defect. The multimodal features of the suspected defect are then cross-validated. After the verification is passed, the defect type and level are determined, and an output file containing the defect coordinates, feature parameters and corresponding multimodal images is generated.

[0012] Furthermore, in the cross-validation link of the defect recognition output module, for suspected crack defects, it is necessary to simultaneously meet the requirements that there is no abnormal change in the crack characteristic parameters in the visible light image and the temperature gradient of the corresponding area in the infrared image; for suspected temperature abnormality defects, it is necessary to verify that there is no obstruction in the corresponding area in the visible light image and there is no abnormal discharge feature in the ultraviolet image; for suspected corona discharge defects, it is necessary to confirm that there is no sudden change in temperature in the corresponding area in the infrared image and there is no obvious physical damage in the visible light image.

[0013] Furthermore, the defect feature comparison library of the defect identification output module includes three sub-libraries: crack type, temperature anomaly type, and corona discharge type. The sub-libraries respectively store the characteristic parameter threshold range of the corresponding defects. Each threshold can be adjusted by receiving the update instruction sent by the external terminal through the Ethernet interface, and each update operation automatically records the modification time and operator information.

[0014] Compared with existing technologies, this high-precision defect detection system for rail inspection robots has the following beneficial effects: 1. The present invention adopts multimodal vision fusion detection technology, combining multiple modal vision technologies such as visible light vision, infrared vision, and ultraviolet vision, and fuses the image data of each modality. Visible light vision can clearly identify surface defects such as cracks and wear on the track. Infrared vision can detect temperature anomalies of track components to discover potential problems such as poor contact. Ultraviolet vision can be used to detect phenomena such as corona discharge. Comprehensive track information is obtained, effectively improving the accuracy and reliability of defect identification. It can detect defects that are difficult to detect with a single modality, providing strong support for the safe operation and maintenance of the track.

[0015] 2. The present invention realizes the spatiotemporal alignment of multimodal images by setting a timing synchronization module, solves the problem of data misalignment, improves the accuracy of feature matching, adopts a fusion algorithm containing a field of view overlapping mask matrix, strengthens the correlation of cross-modal information, solves the problem of insufficient accuracy of fusion features, and uses a multi-dimensional cross-validation mechanism to combine multimodal features for defect confirmation, thereby reducing misjudgments caused by environmental interference.

[0016] Other advantages, objects and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be learned from the practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] 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. Those skilled in the art can also derive other drawings based on these drawings without inventive effort.

[0018] Figure 1 This is a schematic diagram of the structure of the high-precision defect detection system for rail inspection robots; Figure 2 This is the workflow diagram of the high-precision defect detection system for rail inspection robots. DETAILED DESCRIPTION

[0019] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0020] See also Figure 1 The high-precision defect detection system of this rail inspection robot consists of a multimodal vision acquisition module, a timing synchronization module, a data fusion processing module, and a defect recognition output module. The specific contents of each module are as follows: The multimodal vision acquisition module includes visible light cameras, infrared thermal imagers and ultraviolet imagers arranged in sequence along the robot's inspection direction. It is used to collect multimodal image data of the track, including visible light image data, infrared image data and ultraviolet image data. The axis of the lens of each device is at a preset angle to the track detection surface to ensure full coverage of the detection area.

[0021] Visible light camera: The lens focal length is adjusted by an electric focus assembly, and the focus parameters are pre-mapped to the robot's travel speed. The camera has a built-in polarizer, and the polarization angle of the polarizer is continuously adjusted within the range of 0-90° by a stepper motor to eliminate reflection interference from the track surface.

[0022] Infrared thermal imager: The detection field of view has a 30%-50% overlap with the visible light camera, the sampling frequency is not less than 1 / 2 of the visible light camera sampling frequency, the detector resolution is not less than 640×512 pixels, and the temperature measurement range covers -20℃ to 150℃.

[0023] Ultraviolet imager: used to collect ultraviolet image data of the track and assist in detecting phenomena such as corona discharge.

[0024] The timing synchronization module is electrically connected to the multimodal vision acquisition module, and includes a laser ranging sensor and a synchronization controller. The laser ranging sensor obtains the distance value between the robot and the track detection point in real time. The synchronization controller dynamically adjusts the trigger interval of each visual device according to the distance value. The trigger signal adopts a TTL level signal, and the rising edge triggers each device to collect image data of the corresponding detection area at the same inspection time, ensuring the consistency of the multimodal image data in the time dimension.

[0025] The data fusion processing module includes a preprocessing unit, a feature extraction unit, and a fusion operation unit, which is used to process the multimodal image data after time series synchronization. The specific contents are as follows: Preprocessing unit: performs denoising, edge enhancement and distortion correction on visible light images; performs temperature calibration and pseudo-color coding conversion on infrared images; performs dark current elimination and image enhancement on ultraviolet images. After preprocessing, the pixel resolution of each modality image remains consistent.

[0026] Feature extraction unit: extracts crack edge gradient features, grayscale mean and texture entropy features of the wear area from visible light images; extracts temperature field distribution entropy, high temperature point clustering density and temperature gradient change rate features from infrared images; extracts the circularity, area ratio and grayscale distribution standard deviation features of the corona spot from ultraviolet images.

[0027] Fusion operation unit: Using weighted feature fusion algorithm, based on the weights of visible light, infrared and ultraviolet features, corresponding modal feature vectors, mask matrices of the overlapping areas of visible light and infrared fields of view, and mask matrices of the overlapping areas of ultraviolet and visible light fields of view, perform fusion operations and output the fusion results.

[0028] The defect recognition output module has a built-in defect feature comparison library, which completes defect determination and output based on the fusion results. The specific contents are as follows: Defect feature comparison library: Contains three sub-libraries: crack, temperature anomaly, and corona discharge. Each sub-library stores the threshold range of characteristic parameters of the corresponding defects. Each threshold can be adjusted by receiving update instructions sent by an external terminal through the Ethernet interface, and each update operation automatically records the modification time and operator information.

[0029] Defect determination process: First, the fused feature data is divided into several feature subsets according to the defect type. Each feature subset is matched one by one with the threshold range of the corresponding sub-library in the defect feature comparison library. The feature similarity is calculated using a fuzzy logic algorithm. When the similarity exceeds the preset threshold, it is marked as a suspected defect; then the multimodal features of the suspected defects are cross-validated. After the verification is passed, the defect type and level are determined, and an output file containing the defect coordinates, feature parameters and corresponding multimodal images is generated.

[0030] Example 1

[0031] This embodiment is applied to the daily inspection scenarios of urban rail transit tracks. For various defects that may appear on subway tracks during long-term operation, such as rail surface cracks, abnormal fastener temperatures, and contact network corona discharge, a high-precision defect detection system of a track inspection robot is used to realize automated and comprehensive defect detection, replacing traditional manual inspection methods, improving inspection efficiency and defect identification accuracy, and ensuring the safe and stable operation of subway tracks.

[0032] During the actual inspection process, the rail inspection robot moves at a constant speed along the subway track. Figure 2 , each module of the system works together according to the following process: The visible light camera, infrared thermal imager and ultraviolet imager in the multimodal vision acquisition module are arranged in sequence along the robot's inspection direction. The lens axis of each device is at a preset angle to the track detection surface to ensure full coverage of the track detection area.

[0033] Visible light camera: Equipped with a built-in polarizer, the polarization angle is continuously adjusted within a range of 0-90° by a stepper motor to eliminate interference caused by light reflection on the track surface. At the same time, the focal length of the lens is adjusted by an electric focusing assembly according to the robot's travel speed. The focusing parameters are mapped to the travel speed in a preset relationship, ensuring that visible light images of the track surface can be clearly captured at different travel speeds, used to capture appearance defects such as cracks and wear.

[0034] Infrared thermal imager: The detection field of view overlaps with the visible light camera by 30%-50%. It is used to collect infrared image data of track components to detect temperature anomalies.

[0035] Ultraviolet imager: used to collect ultraviolet image data of relevant areas of the track and capture the ultraviolet signals generated by corona discharge phenomena.

[0036] The timing synchronization module is electrically connected to the multimodal vision acquisition module. The laser ranging sensor obtains the distance value between the robot and the track detection point in real time and transmits the distance information to the synchronization controller. The synchronization controller dynamically adjusts the trigger interval of each visual device according to the distance value. The trigger signal adopts a TTL level signal. The rising edge triggers the visible light camera, infrared thermal imager and ultraviolet imager to collect image data of the corresponding detection area at the same inspection time, ensuring the consistency of the multimodal image data in the time dimension and avoiding image misalignment problems caused by asynchronous acquisition time.

[0037] The data fusion processing module receives the time-series synchronized multimodal image data and performs preprocessing, feature extraction, and fusion operations in sequence: The preprocessing unit performs denoising, edge enhancement and distortion correction on visible light images, removes noise interference in the image, enhances the edge features of defects such as cracks, corrects image distortion caused by the camera optical system, performs temperature calibration and pseudo-color coding conversion on infrared images, converts the grayscale values ​​of infrared images into corresponding temperature values, and makes the temperature distribution more intuitive through pseudo-color coding. It performs dark current elimination and image enhancement on ultraviolet images, removes dark current noise generated by the detector itself, and enhances the characteristics of corona discharge spots. After preprocessing, the pixel resolution of each modal image remains consistent.

[0038] The feature extraction unit extracts the crack edge gradient features, the grayscale mean value of the wear area and the texture entropy features from the visible light image. The texture entropy feature calculation formula is: , where is the probability distribution of grayscale histogram of the wear area, is the number of gray levels, is a minimum positive constant used to avoid zero-value input in logarithmic operations. The entropy of the temperature field distribution, the clustering density of high-temperature points, and the temperature gradient change rate are extracted from the infrared image to reflect the temperature distribution and changes of the track components. The circularity, area ratio, and grayscale distribution standard deviation of the corona spot are extracted from the ultraviolet image to characterize the characteristics of the corona discharge phenomenon.

[0039] The fusion operation unit adopts the weighted feature fusion algorithm, and the calculation formula of the fusion feature vector is: ,in 、 、 are the feature weights of visible light, infrared, and ultraviolet, with initial values ​​of 0.5, 0.3, and 0.2, respectively. 、 、 are the corresponding modal eigenvectors respectively; It is the mask matrix of the overlapping area of ​​visible light and infrared field of view; is the mask matrix of the overlapping area of ​​ultraviolet and visible light fields; Represents element-level multiplication operation. Through this fusion algorithm, multimodal features are effectively fused and the fusion result is output.

[0040] The defect recognition output module has a built-in defect feature comparison library, which contains three sub-libraries: crack, temperature anomaly, and corona discharge. Each library stores the threshold range of characteristic parameters of the corresponding defects. Each threshold can be adjusted by receiving update instructions sent by an external terminal through the Ethernet interface. Each update operation automatically records the modification time and operator information.

[0041] First, the fused feature data is divided into several feature subsets according to the defect type. Each feature subset is matched one by one with the threshold range of the corresponding sub-library in the defect feature comparison library. The matching process uses the fuzzy logic algorithm to calculate the feature similarity. The similarity calculation formula is: ,in is the feature parameter to be matched, is the corresponding threshold in the comparison library, is the feature weight, is the feature dimension.

[0042] When similarity When the preset threshold is exceeded, it is marked as a suspected defect, and then the multimodal features of the suspected defect are cross-validated: for suspected crack defects, it is necessary to simultaneously meet the requirements of no abnormal changes in the crack feature parameters in the visible light image and the temperature gradient in the corresponding area in the infrared image; for suspected temperature anomaly defects, it is necessary to verify that there are no obstructions in the corresponding area in the visible light image and no abnormal discharge features in the ultraviolet image; for suspected corona discharge defects, it is necessary to confirm that there is no sudden change in the temperature of the corresponding area in the infrared image and no obvious physical damage in the visible light image.

[0043] After verification, the defect type and level are determined, and an output file containing defect coordinates, feature parameters and corresponding multimodal images is generated, completing the entire defect detection process.

[0044] To sum up, through the application of this embodiment, the high-precision defect detection system of the track inspection robot can give full play to the advantages of multimodal visual fusion detection technology, comprehensively obtain various types of information on the track, and effectively improve the accuracy and reliability of defect identification. The timing synchronization module ensures the spatiotemporal alignment of multimodal images, the data fusion processing module realizes the effective fusion of multimodal features, and the defect identification output module reduces misjudgment through multi-dimensional cross-validation, providing strong technical support for the safe operation and maintenance of urban rail transit tracks, and significantly improving the intelligence level and work efficiency of track inspection.

[0045] Example 2

[0046] This embodiment is applied to the regular special inspection scenario of high-speed railway tracks. It aims to address hidden defects such as crack propagation at rail welds, temperature anomalies caused by loose sleeper bolts, and corona discharge of contact network insulators, which are easily caused by factors such as high-speed running loads and environmental erosion. With the help of the high-precision defect detection system of the track inspection robot, all-weather, high-precision automated inspection is achieved, solving the problems of low efficiency, high risk, and high defect missed detection rate of traditional manual inspection in high-speed track environments, thereby ensuring the safe operation of high-speed railway tracks.

[0047] In actual inspection operations, the rail inspection robot moves autonomously along the high-speed rail track according to the preset path. Figure 2 , the collaborative operation process of each module of the system is as follows: In the multimodal visual acquisition module, visible light cameras, infrared thermal imagers, and ultraviolet imagers are installed in sequence along the robot's inspection direction. The lens axis of each device is at a preset angle to the track detection surface to ensure accurate coverage of key parts of the track (such as rail joints, fasteners, contact lines, etc.).

[0048] Visible light camera: The focal length of its lens is dynamically adjusted by the electric focus assembly according to the robot's travel speed. The focusing parameters and travel speed have a preset mapping relationship, ensuring that the surface details of the rail can still be clearly captured even at high speeds. The camera's built-in polarizer is continuously adjusted in the range of 0-90° by a stepper motor, effectively eliminating reflections on the rail surface caused by direct sunlight or night-time supplementary lighting, and accurately capturing visible light images of appearance defects such as cracks in welds and loose bolts.

[0049] Infrared thermal imager: The detection field of view maintains a 30%-50% overlap with the visible light camera to ensure that the temperature information and appearance information of the same detection area are complementary. Its sampling frequency is no less than 1 / 2 of the visible light camera sampling frequency. The detector resolution meets the detection requirements, and the temperature measurement range covers the temperature range that may occur in high-speed rail track components. It is used to collect infrared image data of track components and capture temperature anomalies caused by poor contact, loose bolts, etc.

[0050] Ultraviolet imager: focuses on high-voltage components such as the contact network, collects ultraviolet image data, and captures ultraviolet signals generated by corona discharge phenomena, providing a basis for judging problems such as insulator aging and poor contact.

[0051] The timing synchronization module is electrically connected to the multimodal vision acquisition module. The laser ranging sensor monitors the distance between the robot and the track detection point in real time, and transmits the data to the synchronization controller in real time. The synchronization controller dynamically calculates and adjusts the trigger interval of each visual device according to the distance value to ensure that during the movement of the robot, the visible light camera, infrared thermal imager, and ultraviolet imager accurately collect image data of the corresponding detection area at the same inspection time. The trigger signal adopts a TTL level signal, and the acquisition action is triggered by the rising edge to ensure the consistency of the multimodal image data in time and space, and avoid the dislocation of defect information caused by asynchronous acquisition.

[0052] The data fusion processing module receives the time-series synchronized multimodal image data and processes it in the order of preprocessing, feature extraction, and fusion operation: The preprocessing unit denoises the visible light image to remove random noise in the image; uses an edge enhancement algorithm to highlight the edge contours of defects such as cracks and loose bolts; performs distortion correction to eliminate the image geometric distortion caused by the camera optical system, performs temperature calibration on the infrared image, and converts the image grayscale value into the actual temperature value; uses pseudo-color coding conversion to present the temperature distribution in an intuitive color form to facilitate subsequent feature extraction, and performs dark current elimination on the ultraviolet image to remove the dark current noise generated by the detector itself; uses an image enhancement algorithm to improve the contrast between the corona spot and the background and highlight the discharge characteristics. After preprocessing, the pixel resolution of the three modal images remains consistent, laying the foundation for subsequent fusion processing.

[0053] The feature extraction unit extracts the crack edge gradient feature (reflecting the direction and depth of the crack), the grayscale mean of the wear area (reflecting the degree of wear), and the texture entropy feature from the visible light image. The texture entropy feature calculation formula is: , where is the probability distribution of the grayscale histogram of the wear area, is the number of gray levels, is a minimum positive constant used to avoid zero-value input in logarithmic operations. This formula can be used to quantify the texture complexity of the wear area and assist in determining the type of wear. The temperature field distribution entropy (reflecting the uniformity of temperature distribution), high-temperature point clustering density (reflecting the concentration of abnormal temperature areas), and temperature gradient change rate characteristics (reflecting the severity of temperature changes) are extracted from infrared images to provide a basis for determining the cause of temperature anomalies. The circularity (determining the shape regularity of the discharge area), area ratio (reflecting the scale of discharge), and grayscale distribution standard deviation (reflecting the uniformity of discharge intensity) characteristics of the corona spot are extracted from the ultraviolet image to accurately characterize the corona discharge phenomenon.

[0054] The fusion operation unit adopts the weighted feature fusion algorithm, and the calculation formula of the fusion feature vector is: ,in 、 、 are the feature weights of visible light, infrared, and ultraviolet, respectively (the initial values ​​are 0.5, 0.3, and 0.2, respectively, and can be dynamically modified based on the confusion matrix of historical detection data). 、 are the eigenvectors of the corresponding modes respectively; is the mask matrix of the overlapping area of ​​visible light and infrared field of view (used to enhance the feature correlation of the overlapping area), is the mask matrix of the overlapping area of ​​ultraviolet and visible light fields; Represents element-level multiplication operation. Through this algorithm, multimodal features are deeply fused to output a fusion result that comprehensively reflects the track status.

[0055] The defect feature comparison library built into the defect identification output module contains three sub-libraries: crack, temperature anomaly, and corona discharge. The sub-libraries store the threshold ranges of characteristic parameters for various defects. These thresholds can be adjusted by receiving update instructions from a remote terminal via the Ethernet interface. Each update automatically records the modification time and operator information to ensure the timeliness and traceability of the comparison library.

[0056] The defect recognition process is as follows: First, the fused feature data is divided into several feature subsets according to the defect type (such as rail crack subset, temperature anomaly subset, corona discharge subset, etc.). Each feature subset is matched one by one with the threshold range of the corresponding sub-library in the comparison library. The fuzzy logic algorithm is used to calculate the feature similarity during matching. The formula is: ,in is the feature parameter to be matched, is the corresponding threshold in the comparison library, is the feature weight, is the feature dimension, when the similarity When the preset threshold is exceeded, it is marked as a suspected defect.

[0057] For suspected defects such as cracks, it is necessary to verify that there is no abnormal change in the crack characteristic parameters in the visible light image and the temperature gradient in the corresponding area of ​​the infrared image to avoid misjudging non-crack surface scratches as cracks; for suspected defects such as temperature anomalies, it is necessary to confirm that there are no obstructions in the corresponding area of ​​the visible light image and that the ultraviolet image has no abnormal discharge characteristics to eliminate temperature misjudgment caused by foreign object obstruction; for suspected defects such as corona discharge, it is necessary to verify that there is no sudden change in temperature in the corresponding area of ​​the infrared image and that there is no obvious physical damage in the visible light image to ensure that the discharge phenomenon is caused by problems with the component itself.

[0058] After verification, the defect type and level are determined, and a detection report containing the precise coordinates of the defect, detailed characteristic parameters (such as crack length, temperature difference, discharge intensity, etc.) and corresponding multimodal images is generated and sent to the ground control center via the wireless transmission module.

[0059] To sum up, this embodiment realizes all-round and high-precision detection of various defects in high-speed rail tracks through the application of the high-precision defect detection system of the track inspection robot. The multimodal visual fusion technology makes up for the limitations of single-modality detection. The timing synchronization module ensures the temporal and spatial consistency of the data. The data fusion processing module improves the integrity of the feature information. The defect identification output module reduces the misjudgment rate through cross-validation. The system significantly improves the efficiency and accuracy of high-speed rail track inspection, provides reliable data support for the preventive maintenance of high-speed rail tracks, and ensures the safety and stability of high-speed rail operation.

[0060] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can, without departing from the scope of the technical solution of the present invention, make some changes or modifications to equivalent embodiments using the technical contents disclosed above. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. High-precision defect detection system for rail inspection robots, characterized by: The system includes: multimodal vision acquisition module, time sequence synchronization module, data fusion processing module and defect recognition output module; The multimodal visual acquisition module includes a visible light camera, an infrared thermal imager, and an ultraviolet imager arranged in sequence along the robot's inspection direction, and is used to collect multimodal image data of the track, including visible light image data, infrared image data, and ultraviolet image data. The axis of the lens of each device forms a preset angle with the track detection surface; The timing synchronization module is electrically connected to the multimodal vision acquisition module and is used to trigger each device to collect image data of the corresponding detection area at the same inspection time; The data fusion processing module includes a preprocessing unit, a feature extraction unit and a fusion operation unit, which receives the multimodal image data after time sequence synchronization, and outputs the fusion result after preprocessing, feature extraction and fusion operation; The defect recognition output module has a built-in defect feature comparison library, and completes defect determination and output based on the fusion results.

2. The high-precision defect detection system for rail inspection robots according to claim 1 is characterized in that: In the multimodal vision acquisition module, the focal length of the visible light camera's lens is adjusted by an electric focusing assembly, and the focusing parameters are mapped to the robot's travel speed in a preset relationship. The camera also has a built-in polarizer, and the polarization angle of the polarizer is continuously adjusted within the range of 0-90° by a stepper motor. The detection field of view of the infrared thermal imager overlaps with that of the visible light camera by 30%-50%.

3. The high-precision defect detection system for rail inspection robots according to claim 1, characterized in that: The timing synchronization module includes a laser ranging sensor and a synchronization controller. The laser ranging sensor obtains the distance value between the robot and the track detection point in real time. The synchronization controller dynamically adjusts the trigger interval of each visual device according to the distance value. The trigger signal adopts a TTL level signal, and the rising edge triggers the acquisition action.

4. The high-precision defect detection system for rail inspection robots according to claim 1, characterized in that: The preprocessing unit of the data fusion processing module performs denoising, edge enhancement, and distortion correction on the visible light image, performs temperature calibration and pseudo-color coding conversion on the infrared image, and performs dark current elimination and image enhancement on the ultraviolet image. After preprocessing, the pixel resolution of each modality image remains consistent. The feature extraction unit extracts the crack edge gradient feature, the grayscale mean value of the wear area, and the texture entropy feature from the visible light image. The texture entropy feature calculation formula is: ,in, is the probability distribution of grayscale histogram of the wear area, is the number of gray levels, is a minimum positive constant. The entropy of temperature field distribution, high temperature point clustering density and temperature gradient change rate characteristics are extracted from infrared images. The circularity, area ratio and grayscale distribution standard deviation characteristics of corona spots are extracted from ultraviolet images.

5. The high-precision defect detection system for rail inspection robots according to claim 1, characterized in that: The fusion operation unit of the data fusion processing module adopts a weighted feature fusion algorithm, and the fusion feature vector calculation formula is: ,in, 、 、 are the feature weights of visible light, infrared, and ultraviolet, respectively. 、 、 are the corresponding modal eigenvectors, is the mask matrix of the overlapping area of ​​visible light and infrared field of view, is the mask matrix of the overlapping area of ​​ultraviolet and visible light fields of view, Represents element-wise multiplication operation.

6. The high-precision defect detection system for rail inspection robots according to claim 1, characterized in that: When the defect recognition output module is working, the fused feature data is first divided into several feature subsets according to the defect type. Each feature subset is matched one by one with the threshold range of the corresponding sub-library in the defect feature comparison library. The matching process uses a fuzzy logic algorithm to calculate the feature similarity. The similarity calculation formula is: ,in, is the feature parameter to be matched, is the corresponding threshold in the comparison library, is the feature weight, is the feature dimension, similarity When the preset threshold is exceeded, it is marked as a suspected defect. The multimodal features of the suspected defect are then cross-validated. After the verification is passed, the defect type and level are determined, and an output file containing the defect coordinates, feature parameters and corresponding multimodal images is generated.

7. The high-precision defect detection system for rail inspection robots according to claim 6, characterized in that: The cross-validation link of the defect recognition output module requires that, for suspected crack defects, there must be no abnormal changes in the crack characteristic parameters in the visible light image and the temperature gradient in the corresponding area in the infrared image. For suspected temperature abnormality defects, it must be verified that there are no obstructions in the corresponding area in the visible light image and that the ultraviolet image has no abnormal discharge characteristics. For suspected corona discharge defects, it must be confirmed that there is no sudden change in the temperature of the corresponding area in the infrared image and that there is no obvious physical damage in the visible light image.

8. The high-precision defect detection system for rail inspection robots according to claim 1, characterized in that: The defect feature comparison library of the defect identification output module includes three sub-libraries: crack type, temperature anomaly type, and corona discharge type. The sub-libraries respectively store the characteristic parameter threshold range of the corresponding defects. Each threshold can be adjusted by receiving update instructions sent by an external terminal through the Ethernet interface, and each update operation automatically records the modification time and operator information.

Images