Double-track vehicle-mounted multi-source track disease detection system

Abstract

The invention relates to a double-track vehicle-mounted multi-source track disease detection system, which comprises two detection devices connected side by side, each detection device comprises: a frame body, the bottom of which is provided with a walking mechanism used for walking on a steel rail; the imaging assembly is arranged on the frame body and is used for imaging a steel rail to be detected and acquiring image data of the steel rail to be detected; the ultrasonic probe is arranged on the frame body and is used for scanning a to-be-detected steel rail and acquiring ultrasonic scanning data of the to-be-detected steel rail; the processing device is in signal connection with the imaging assembly and the ultrasonic probe, and the processing device obtains the image data and the ultrasonic scanning data and judges the disease state of the steel rail to be detected in combination with the image data and the ultrasonic scanning data. The steel rail detection efficiency and detection accuracy can be improved.

Description

A dual-track vehicle-mounted multi-source track defect detection system Technical Field

[0001] This application relates to the technical field of rail inspection, specifically to a dual-rail vehicle-mounted multi-source rail defect detection system. Background Technology

[0002] With the rapid development of my country's railway transportation network towards high speed, heavy load, and high density, the health of rails, as the core of rail transit infrastructure, directly affects operational safety. Under complex alternating loads and environmental factors, rails can develop various fatigue and damage defects such as core damage, cracks, and spalling. Therefore, regular, efficient, and accurate non-destructive testing of rails is the lifeline for ensuring railway transportation safety.

[0003] Currently, the rail flaw detection equipment widely used by domestic railway engineering departments mainly includes large flaw detection vehicles and hand-push monorail flaw detectors. Large flaw detection vehicles are expensive and lack flexibility. Hand-push monorail flaw detectors have the advantages of low cost and ease of use. They generally scan the internal structure of the rail with an ultrasonic probe or capture images of the rail surface with a camera. Inspectors judge the condition of the rail by observing the scanned or captured images. This manual judgment method is greatly affected by human subjectivity and has a high possibility of misjudgment and incorrect judgment, which needs to be improved. Summary of the Invention

[0004] To address the problems existing in the prior art, this application aims to provide a dual-track vehicle-mounted multi-source rail defect detection system. This application can improve the efficiency and accuracy of rail inspection.

[0005] This application discloses a dual-track vehicle-mounted multi-source track defect detection system, comprising two detection devices connected side-by-side. Each detection device includes: a frame with a traveling mechanism at its bottom for moving along the rails; an imaging component mounted on the frame for imaging the rail under test and acquiring image data; an ultrasonic probe mounted on the frame for scanning the rail under test and acquiring ultrasonic scanning data; and a processing device connected to the imaging component and the ultrasonic probe, respectively. The processing device acquires the image data and ultrasonic scanning data, and, by combining the image data and ultrasonic scanning data, determines the defect status of the rail under test.

[0006] Preferably, the imaging assembly includes at least a first imaging element, a second imaging element, and a third imaging element disposed at the front end of the frame, wherein the imaging end of the first imaging element faces the top surface of the rail, the imaging end of the second imaging element faces the inner vertical surface and the arc transition area of ​​the rail, and the imaging end of the third imaging element faces the outer vertical surface and the arc transition area of ​​the rail.

[0007] Preferably, when the processing device receives the image data, it performs image data preprocessing on the image data. The image data preprocessing includes the following steps: distortion correction, determining the radial distortion and tangential distortion of the image data respectively, and completing image correction based on the radial distortion and tangential distortion; and equalizing the corrected image data to eliminate tile boundary artifacts in the image data.

[0008] Preferably, the image data is corrected using the Brown-Conrad model, specifically including the following steps: obtaining the intrinsic parameter matrix of the imaging component through calibration. and distortion coefficient Obtain distortion-free coordinates With distorted coordinates The mapping relationship between radially undistorted coordinates and radially distorted coordinates is expressed as follows: ,in, , , and All are radial distortion coefficients; the mapping relationship between the tangential undistorted coordinates and the tangential distorted coordinates is expressed as: ;, ;in, , All are tangential distortion coefficients; image data correction is completed through the mapping relationship; equalization processing is performed on the corrected image data to eliminate tile boundary artifacts in the image data, specifically including: segmenting the image data into... The image data consists of several non-overlapping tiles. A histogram is calculated for each tile, and the histogram is cropped using a contrast limiting factor CL. The portion exceeding the cropping range is evenly distributed within the cropping range of the histogram. The cropped histogram is then equalized. The transformation functions of adjacent tiles are synthesized using bilinear interpolation to eliminate tile boundary artifacts in the image data.

[0009] Preferably, when the processing device receives the ultrasonic scanning data, it performs ultrasonic scanning data preprocessing on the ultrasonic scanning data. The ultrasonic scanning data preprocessing includes the following steps: performing wavelet threshold noise reduction on the ultrasonic scanning data to obtain noise-reduced ultrasonic scanning data; and performing further processing on the noise-reduced ultrasonic scanning data. Layer wavelet decomposition to obtain wavelet coefficients of each layer. ,in, This represents the input variable of the threshold function. As a scale, For the location; perform unbiased risk estimation on the threshold, for the ... Layer wavelet decomposition results, threshold Represented as: ,in, For the first Number of coefficients in the layer The noise standard deviation is used; a soft thresholding function is used for thresholding; the ultrasonic scanning data is reconstructed using wavelet inverse using the thresholded coefficients to obtain the noise-reduced ultrasonic scanning data; a time-dependent gain function is used. Compensation for attenuation: , ,in, This is the initial gain. is the attenuation coefficient of the ultrasonic signal. This represents the propagation time of the ultrasonic signal.

[0010] Preferably, the imaging component acquires data synchronously with the ultrasonic probe through a synchronization mechanism. The synchronization mechanism includes: using a temperature-controlled crystal oscillator as a time reference, using an FPGA as a synchronous acquisition controller, and during the acquisition process, using an encoder pulse or timer signal as a trigger source input to the FPGA. The FPGA executes a trigger event, which generates a global trigger pulse for the FPGA and synchronously sends it to the imaging component and the ultrasonic probe. The FPGA timestamps the trigger event, so that both the image data acquired by the imaging component and the ultrasonic scanning data acquired by the ultrasonic probe carry the marked timestamp.

[0011] Preferably, the processing device acquires the image data and ultrasonic scan data, and determines the defect status of the rail under test by combining the image data and ultrasonic scan data, specifically including: coordinate normalization processing: establishing a coordinate system with the rail under test as the reference, calibrating the imaging component and ultrasonic probe, and associating the imaging component and the ultrasonic probe to the coordinate system; ultrasonic feature vector Construction: Extracting the echo signal from the ultrasonic scanning data Acquire the echo signal echo peak Echo energy echo centroid ; for the echo signal Perform FFT transform to obtain the echo frequency domain signal Acquire the echo frequency domain signal Spectral centroid and spectrum width Perform continuous wavelet transform on the echo signal and the echo frequency domain signal to obtain the time-frequency diagram. ,in, As a scale, For translation; extract within a specific scale range The maximum amplitude and energy are used as the ultrasonic feature vector. The specific scale range corresponds to the frequency of defect features; image feature vector Construction: A feature extraction model for extracting image features is constructed. Several preprocessed image data sets are input into the feature extraction model to extract feature vectors from the image data. The feature vectors are then weighted and summed to obtain the image feature vector. Feature fusion: For the ultrasonic feature vector... and image feature vectors Standardization was performed on each of the ultrasonic feature vectors. and image feature vectors Feature fusion is performed using a concatenation function to obtain a fused feature vector. Diagnostic output: The fused feature vector The results are input into a pre-trained diagnostic model to obtain the diagnostic results of the defects of the rail under test.

[0012] Preferably, in the feature fusion step, principal component analysis is used to process the image feature vector. Dimensionality reduction is performed.

[0013] Preferably, the diagnostic model includes an XGBoost model, and the training process of the diagnostic model includes the following steps: constructing a model containing... XGBoost model with regression trees: ;in, Indicates the first The original predicted scores for each sample. Indicates the first The fusion feature vector of each sample , Indicates the first A regression tree, Construct the objective function of the XGBoost model: , ,in, For loss function, For the regularization term of the regression tree, The number of leaf nodes in the regression tree. It is the fraction vector of the leaf nodes. and For hyperparameters; initialize constant prediction values, for the th In the next iteration, the loss function is calculated with respect to the current prediction result. First-order gradient and second-order gradient Based on dataset Greedily grow a new regression tree with the goal of minimizing the structure score. The structural fraction is represented as: , will the new regression tree Add to the XGBoost model: ,in, The learning rate is used; training is repeated until the XGBoost model meets the accuracy requirements or reaches the preset number of iterations.

[0014] Preferably, the diagnostic model further includes a Softmax function, which obtains raw scores about the rail under test through the XGBoost model, and inputs these raw scores into the Softmax function to obtain the probability distribution of defects about the rail under test. Based on the obtained disease probability distribution The results of the defect status detection of the rail under test are obtained.

[0015] The dual-track vehicle-mounted multi-source rail defect detection system described in this application has the following advantages: 1. This application combines ultrasonic scanning and surface visual recognition, and uses model algorithms to fuse ultrasonic scanning data and image data to comprehensively judge the defect status of the rail. It has the advantages of more comprehensive detection parameters and more accurate detection results, and can realize automated judgment, avoiding human interference with the detection results and improving the accuracy and objectivity of defect detection; 2. This application can realize simultaneous operation of dual tracks, which can significantly improve detection efficiency. Attached Figure Description

[0016] Figure 1 is a structural schematic diagram of a dual-track vehicle-mounted multi-source track defect detection system according to this application; Figure 2 is a structural schematic diagram of a single detection device according to this application; Figure 3 is a structural schematic diagram of a single detection device according to this application; Figure 4 is a front view of a single detection device according to this application; Figure 5 is a side view of a single detection device according to this application.

[0017] Explanation of reference numerals in the attached drawings: 1-Frame, 2-Walking mechanism, 21-First walking component, 22-Second walking component, 31-First imaging element, 32-Second imaging element, 33-Third imaging element, 4-Ultrasonic probe, 5-Processing device, 61-Storage tank, 62-First housing, 63-Second housing, 64-Third housing, 65-Side door, 651-Handle, 71-Dispensing tube, 72-Drip tube, 81-First connecting rod, 82-Second connecting rod, 83-Support rod, 84-Sleeve, 85-Push-pull rod. Detailed Implementation

[0018] As shown in Figures 1-5, the dual-track vehicle-mounted multi-source track defect detection system described in this application includes two detection devices connected side by side. The detection device specifically includes: a frame 1, and a walking mechanism 2 at the bottom of the frame 1. The walking mechanism 2 includes a first walking component 21 for walking on the rail and a second walking component 22 for walking on the ground. The walking mechanism 2 is used to drive the frame 1 as a whole to walk on the rail or on the ground.

[0019] The imaging assembly includes a first imaging element 31, a second imaging element 32, and a third imaging element 33 disposed at the front end of the frame 1. Specifically, the first imaging element 31, the second imaging element 32, and the third imaging element 33 are all industrial cameras. The imaging end of the first imaging element 31, i.e. the lens end, faces the top surface of the rail. Specifically, the first imaging element 31 faces the top surface of the rail vertically or at a slight angle, and is used to detect defects such as rail surface spalling, cracks, and wear.

[0020] The imaging end of the second imaging element 32 faces the inner vertical surface and arc transition area of ​​the rail, and is used to detect inner cracks and spalling.

[0021] The imaging end of the third imaging element 33 faces the outer vertical surface and arc transition area of ​​the rail, and is used to detect cracks and damage on the outer side.

[0022] During the inspection process, the three imaging devices are activated simultaneously, enabling simultaneous exposure and image acquisition of three feature surfaces of the rail at the same mileage, thus fundamentally eliminating blind spots.

[0023] In other alternative embodiments, the number of imaging elements can be increased according to detection requirements, and this embodiment does not limit this.

[0024] The ultrasonic probe 4 is located at the bottom of the frame 1, with its scanning end facing the rail, i.e., vertically downward, and is used to perform ultrasonic scanning on the rail to obtain an internal scanning image of the rail.

[0025] The processing device 5, specifically a computer or other device with computing capabilities, transmits images captured by the imaging component and scanned by the ultrasonic probe 4 to the computer for display. For example, on the same display interface, it simultaneously displays surface images or defect information acquired or processed by the multi-view visual inspection unit, as well as scanned image data from the ultrasonic flaw detection equipment. It can also provide interactive functionality through a software interface, automatically locating and highlighting the corresponding ultrasonic scan image at the selected defect location in the surface image in response to the user's selection.

[0026] The detection device also includes an illumination component, which is located above or beside the imaging component. The illumination component is specifically a searchlight, and its illumination range can cover the field of view of the three imaging components to provide illumination in nighttime or tunnel operation scenarios, thereby extending the suitable working time and applicable scenarios of the device.

[0027] The detection device also includes a liquid storage tank 61 mounted on the frame 1. The liquid storage tank 61 is used to store ultrasonic testing coupling agent or water and other liquids. It has an injection hole at the top for injecting liquid from the top. The bottom of the liquid storage tank 61 has an outlet connected to an outlet pipe, which is connected to a distributor pipe 71. The outlet pipe and the distributor pipe 71 are vertically arranged, and the distributor pipe 71 is horizontally arranged. The bottom wall of the distributor pipe 71 has a through hole connected to a dropper 72. A distributor valve is provided between the distributor pipe 71 and the dropper 72. The distributor valve can be an electric valve or a manual valve. When the distributor valve is open, the liquid in the liquid storage tank 61 flows through the outlet pipe to the distributor pipe 71, and then flows out through the dropper 72 to drip the coupling agent and other liquids onto the rail.

[0028] The detection device further includes: a communication component, which is signal-connected to the processing device 5; the communication component can be a wireless router, responsible for building an in-vehicle local area network and exchanging data with a remote server or cloud platform; and a power supply component, specifically a battery, which is electrically connected to the imaging component, ultrasonic probe 4, processing device 5, and lighting component for power supply.

[0029] The detection device also includes a first housing 62, a second housing 63, and a third housing 64. The liquid storage tank 61, the first housing 62, the second housing 63, and the third housing 64 are all square housings. The frame 1 includes two vertically arranged columns and a crossbeam connecting the two columns, forming an inverted "U" shape. A load-bearing space is formed between the columns and the crossbeam. The liquid storage tank 61 is located at the bottom of this load-bearing space. The first housing 62, the second housing 63, and the third housing 64 are respectively located beside, above, and opposite the liquid storage tank 61. The housings are bolted to the frame 1. The first housing 62 is used to store necessary maintenance tools and spare parts; the second housing 63 is used to store power components such as batteries; and the third housing 64 is used to store communication components such as wireless routers. This structure forms a balanced counterweight to ensure the overall weight of the equipment is balanced and its movement is smooth. A side door 65 is provided on the side of the housing for easy access to items, and a handle 651 is provided on the side door for easy pushing and pulling.

[0030] The upper part of the crossbeam is provided with a support for supporting the processing device 5. The support frame is specifically a U-shaped frame with elastic clips that can clamp the computer onto the support frame.

[0031] Furthermore, in this embodiment, the liquid storage tank 61 is made of iron and can store coupling agent, while the first tank 62, the second tank 63 and the third tank 64 are all made of PVC plastic to reduce the overall weight of the equipment.

[0032] The traveling component consists of rail wheels located in the middle of the frame 1 and ground wheels located on both sides of the frame 1, which can drive the equipment to travel on rails or the ground.

[0033] The two detection devices are connected by a first connecting rod 81 and a second connecting rod 82. Specifically, a sleeve 84 is provided on the side of the housing, with an internal thread inside the sleeve 84. Both ends of the first connecting rod 81 have external threads, and the first connecting rod 81 and the sleeve 84 are threaded together, thus connecting the two detection devices into one unit. Both ends of the second connecting rod 82 are provided with connecting sleeves, which are connected to the columns of the detection devices. A support rod 83 perpendicular to the second connecting rod 82 is provided on the second connecting rod 82, and a push-pull rod 85 is provided at the upper end of the support rod 83 to facilitate push-pull operation. Through the above structure, the two detection devices can move synchronously to achieve dual-track synchronous detection.

[0034] After receiving image data acquired by the imaging component and ultrasonic scanning data from the ultrasonic probe 4, the processing device 5 of this application sequentially performs the following steps: preprocessing, feature extraction, feature fusion, disease diagnosis, and output decision.

[0035] Specifically, when acquiring image data and ultrasonic scanning data, it is necessary to ensure the synchronization of the acquisition actions to improve detection accuracy. This embodiment achieves this as follows: the imaging component acquires data synchronously with the ultrasonic probe 4 through a synchronization mechanism. The synchronization mechanism includes: using a temperature-controlled crystal oscillator as a time reference and an FPGA (Field-Programmable Gate Array) as a synchronous acquisition controller. During the acquisition process, an encoder pulse or timer signal is used as a trigger source input to the FPGA. The FPGA executes a trigger event, which generates a global trigger pulse for the FPGA and synchronously sends it to the imaging component and the ultrasonic probe 4 through a dedicated digital I / O line. The FPGA timestamps the trigger event, so that both the image data acquired by the imaging component and the ultrasonic scanning data acquired by the ultrasonic probe 4 carry the marked timestamp.

[0036] For example, a timestamp can be represented as After acquiring data, all imaging components and ultrasonic probes 4 transmit the data to the processing device 5 via a high-speed bus. Each set of data carries the timestamp information to verify whether the acquired data is synchronized. Only when the timestamp error between each set of data is not greater than the preset time error threshold, the acquired data can be used in the subsequent disease detection process to avoid the impact of acquisition time error on the detection results.

[0037] After simultaneously acquiring image data and ultrasonic scanning data through the above steps, the processing device 5 receives the image data and performs image data preprocessing. The image data preprocessing includes the following steps: distortion correction, determining the radial and tangential distortions of the image data respectively, and completing image correction based on the radial and tangential distortions; and correcting the image data using the Brown-Conrad model, specifically including the following steps: obtaining the intrinsic parameter matrix of the imaging component through calibration. and distortion coefficient Obtain distortion-free coordinates With distorted coordinates The mapping relationship between radially undistorted coordinates and radially distorted coordinates is expressed as follows: ,in, , , and All are radial distortion coefficients; the mapping relationship between the tangential undistorted coordinates and the tangential distorted coordinates is expressed as: ;, ;in, , All are tangential distortion coefficients; image data correction is performed through the mapping relationship; the corrected image data is then subjected to equalization processing to eliminate tile boundary artifacts in the image data; the image data is then segmented into... The image is divided into several non-overlapping tiles. A histogram is calculated for each tile, and the histogram is cropped using a contrast limiting factor (CL). The portion exceeding the cropping range is evenly distributed within the cropping range of the histogram. The cropped histogram is then equalized. The transformation functions of adjacent tiles are synthesized using bilinear interpolation to eliminate tile boundary artifacts in the image data.

[0038] When the processing device 5 receives the ultrasonic scanning data, it performs ultrasonic scanning data preprocessing on the ultrasonic scanning data. The ultrasonic scanning data preprocessing includes the following steps: performing wavelet threshold noise reduction processing on the ultrasonic scanning data to obtain noise-reduced ultrasonic scanning data; specifically, Symlets (Sym8) wavelet is selected because it is similar in shape to ultrasonic signals.

[0039] The noise-reduced ultrasonic scanning data is then processed. Layer wavelet decomposition to obtain wavelet coefficients of each layer. ,in, This represents the input variable of the threshold function. is a scale used to characterize the frequency components of ultrasonic signals; The position is used to characterize the signal's location on the time axis, corresponding to the rail's internal depth information; an unbiased risk estimate is performed on the threshold, for the ... Layer wavelet decomposition results, threshold Represented as: ,in, For the first Number of coefficients in the layer The noise standard deviation can be obtained through From the highest frequency subband ( Robust estimation is obtained from the wavelet coefficients.

[0040] Thresholding is performed using a soft thresholding function, which is expressed as follows: To maintain a smooth signal line.

[0041] The ultrasonic scanning data is reconstructed using wavelet inverse reconstruction with the thresholded coefficients to obtain denoised ultrasonic scanning data; a time-dependent gain function is then used. Compensation for attenuation: , ,in, This is the initial gain. is the attenuation coefficient of the ultrasonic signal. This represents the propagation time of the ultrasonic signal.

[0042] The aforementioned steps complete the preprocessing of image data and ultrasonic scanning data. Then, feature extraction and feature fusion are performed, specifically as follows: Coordinate normalization: A world coordinate system is established with the rail under test as the reference; specifically, a point on the rail web is taken as the origin. The direction of rail extension is Vertical direction is Horizontal direction is .

[0043] The imaging component and the ultrasonic probe 4 are calibrated, and the imaging component and the ultrasonic probe 4 are associated with the coordinate system.

[0044] Specifically, through offline calibration, the accuracy of each ultrasonic probe's 4 parameters is determined. , The position coordinates on the plane are as follows Each column in the B-Scan image corresponds to a rail. A slice in the direction of depth The echo information on the screen, and so on.

[0045] For imaging component calibration, the extrinsic parameters of each imaging element are known. (Rotation and translation from the world coordinate system to the imaging coordinate system) and intrinsic parameters A world point Projected to pixel coordinates The process is as follows: .

[0046] By solving the above equations, defective pixels in the image data can be identified. Three-dimensional coordinates projected onto the rail surface superior.

[0047] This allows ultrasonic scanning data and image data to be unified into the same coordinate system.

[0048] Ultrasonic feature vector Construction: Extracting the echo signal from the ultrasonic scanning data Acquire the echo signal echo peak Echo energy echo centroid ; for the echo signal Perform FFT transform to obtain the echo frequency domain signal Acquire the echo frequency domain signal Spectral centroid and spectrum width Perform continuous wavelet transform on the echo signal and the echo frequency domain signal to obtain the time-frequency diagram. ,in, It serves as a scale to reflect changes in the frequency components of ultrasonic signals; This is the translation amount, used to reflect the position of the signal on the time axis, corresponding to the internal depth of the rail or the scanning position.

[0049] Extraction within a specific scale range The maximum amplitude and energy are used as the ultrasonic feature vector. The specific scale range corresponds to the defect feature frequency, which is determined by frequency analysis of the ultrasonic signals of known defect samples within the specific scale range.

[0050] Image feature vector Construction: Construct a feature extraction model for extracting image features. Specifically, use a ResNet-50 pre-trained on ImageNet and fine-tuned on a large defect dataset as the feature extraction model.

[0051] Several preprocessed image data sets are input into the feature extraction model to extract feature vectors from the image data; specifically, three corrected image data sets are used... , , ,in, , , To capture images of the rail surface from different perspectives at the same time, several consecutive frames of images are used. , , The inputs are fed into ResNet-50, and feature maps from the activation layers preceding the Global Average Pooling (GAP) layer are extracted and denoted as follows: , , Each feature map has a dimension of 1. .

[0052] Perform GAP processing on each feature map to obtain three feature vectors, denoted as follows: , , .

[0053] Calculate attention weights using a fully connected network: The eigenvectors are then summed using weighted methods. ,final That is, to obtain the image feature vector. Feature fusion: For the ultrasonic feature vector... and image feature vectors Z-score normalization was performed on each ultrasonic feature vector. and image feature vectors Feature fusion is performed using a concatenation function to obtain a fused feature vector. Due to image feature vectors The feature vectors have high dimensionality, requiring principal component analysis to be used to analyze them. Dimensionality reduction is performed, preserving 99% of the variance.

[0054] PCA is implemented through singular value decomposition (SVD) to obtain the projection matrix. , making .

[0055] Standardized ultrasonic feature vectors and image feature vectors Feature fusion is performed using a concatenation function to obtain a fused feature vector. The concatenation function is expressed as: .

[0056] Diagnostic output: The fused feature vector The data is input into a pre-trained diagnostic model to obtain the diagnostic results of the rail under test. Specifically, the diagnostic model includes an XGBoost model, and the training process of the diagnostic model includes the following steps: constructing a model containing... XGBoost model with regression trees: ;in, Indicates the first The original predicted scores for each sample. Indicates the first The fusion feature vector of each sample , Indicates the first A regression tree, F is the hypothesis space of the regression tree.

[0057] Construct the objective function of the XGBoost model: , ,in, For loss function, For the regularization term of the regression tree, The number of leaf nodes in the regression tree. It is the fraction vector of the leaf nodes. and For hyperparameters; initialize constant prediction values, such as 0 for all samples, for the th... In the next iteration, the loss function is calculated with respect to the current prediction result. First-order gradient and second-order gradient Based on dataset Greedily grow a new regression tree with the goal of minimizing the structure score. The structural fraction is represented as: , will the new regression tree Add to the XGBoost model: ,in, The learning rate is used; training is repeated until the XGBoost model meets the accuracy requirements or reaches the preset number of iterations; the raw score of the rail under test is obtained through the XGBoost model, and the raw score is input into the Softmax function to obtain the defect probability distribution of the rail under test. Based on the obtained disease probability distribution Specifically, regarding sample data Model output dimensional vectors ( (where is the number of categories), and the value for each dimension is a raw score, which is then converted into a probability using Softmax, as follows: Based on the detection results of the defect status of the rail under test, it is specifically expressed as follows: , , ,like or all If the result is "Uncertain", then mark it as "Uncertain" and save all the original data and intermediate features for expert review and model retraining.

[0058] The disease classification can be divided into multiple categories, such as cracks, core damage, spalling, and abrasion. These categories can be selected and set according to actual needs, and the probability distribution of each disease type is calculated separately during prediction. Take probability distribution The disease type corresponding to the maximum value in the value is used as the predicted disease type. The maximum value is also used as the confidence level.

[0059] If the confidence level is lower than the set threshold, or the predicted probability for all disease types... If all values ​​are less than 0.5, the test object is marked as "uncertain" and the data is saved for manual verification.

[0060] In the formula, Representing input features Category The probability, Represents probability distribution The index of the disease type corresponding to the maximum value. Represents probability distribution The maximum value, i.e., the confidence level , This represents the preset confidence threshold, used to determine whether the prediction result is reliable.

[0061] In the description of this application, it should be understood that the orientation or positional relationship indicated by directional terms such as "front, back, up, down, left, right", "horizontal, vertical, horizontal" and "top, bottom" is usually based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing this application and simplifying the description. Unless otherwise stated, these directional terms do not indicate or imply that the device or element referred to must have a specific orientation or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on the scope of protection of this application.

[0062] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this application.

Claims

1. A dual-track vehicle-mounted multi-source track defect detection system, characterized in that, The system includes two parallel detection devices, each comprising: a frame with a traveling mechanism at its bottom for moving along the rail; an imaging component mounted on the frame for imaging the rail under test and acquiring image data of the rail; an ultrasonic probe mounted on the frame for scanning the rail under test and acquiring ultrasonic scan data of the rail; and a processing device connected to the imaging component and the ultrasonic probe, respectively. The processing device acquires the image data and the ultrasonic scan data, and, by combining the image data and the ultrasonic scan data, determines the defect status of the rail under test.

2. The dual-track vehicle-mounted multi-source track defect detection system according to claim 1, characterized in that, The imaging assembly includes at least a first imaging element, a second imaging element, and a third imaging element disposed at the front end of the frame, wherein the imaging end of the first imaging element faces the top surface of the rail, the imaging end of the second imaging element faces the inner vertical surface and arc transition area of ​​the rail, and the imaging end of the third imaging element faces the outer vertical surface and arc transition area of ​​the rail.

3. The dual-track vehicle-mounted multi-source track defect detection system according to claim 1 or 2, characterized in that, When the processing device receives the image data, it performs image data preprocessing on the image data. The image data preprocessing includes the following steps: distortion correction, determining the radial distortion and tangential distortion of the image data respectively, and completing image correction based on the radial distortion and tangential distortion; and equalization processing on the corrected image data to eliminate tile boundary artifacts in the image data.

4. The dual-track vehicle-mounted multi-source track defect detection system according to claim 3, characterized in that, The image data is corrected using the Brown-Conrad model, specifically including the following steps: obtaining the intrinsic parameter matrix of the imaging component through calibration. and distortion coefficient Obtain distortion-free coordinates With distorted coordinates The mapping relationship between radially undistorted coordinates and radially distorted coordinates is expressed as follows: ,in, ， 、 and All are radial distortion coefficients; the mapping relationship between the tangential undistorted coordinates and the tangential distorted coordinates is expressed as: ；， ;in, 、 All are tangential distortion coefficients; image data correction is completed through the mapping relationship; equalization processing is performed on the corrected image data to eliminate tile boundary artifacts in the image data, specifically including: segmenting the image data into... The image data consists of several non-overlapping tiles. A histogram is calculated for each tile, and the histogram is cropped using a contrast limiting factor CL. The portion exceeding the cropping range is evenly distributed within the cropping range of the histogram. The cropped histogram is then equalized. The transformation functions of adjacent tiles are synthesized using bilinear interpolation to eliminate tile boundary artifacts in the image data.

5. The dual-track vehicle-mounted multi-source track defect detection system according to claim 4, characterized in that, When the processing device receives the ultrasonic scanning data, it performs ultrasonic scanning data preprocessing on the ultrasonic scanning data. The ultrasonic scanning data preprocessing includes the following steps: performing wavelet threshold noise reduction on the ultrasonic scanning data to obtain noise-reduced ultrasonic scanning data; and performing further processing on the noise-reduced ultrasonic scanning data. Layer wavelet decomposition to obtain wavelet coefficients of each layer. ,in, This represents the input variable of the threshold function. As a scale, For the location; perform unbiased risk estimation on the threshold, for the ... Layer wavelet decomposition results, threshold Represented as: ,in, For the first Number of coefficients in the layer The noise standard deviation is used; a soft thresholding function is used for thresholding; the ultrasonic scanning data is reconstructed using wavelet inverse using the thresholded coefficients to obtain the noise-reduced ultrasonic scanning data; a time-dependent gain function is used. Compensation for attenuation: ， ,in, This is the initial gain. is the attenuation coefficient of the ultrasonic signal. This represents the propagation time of the ultrasonic signal.

6. The dual-track vehicle-mounted multi-source track defect detection system according to claim 5, characterized in that, The imaging component acquires data synchronously with the ultrasonic probe through a synchronization mechanism. The synchronization mechanism includes: using a temperature-controlled crystal oscillator as a time reference, using an FPGA as a synchronous acquisition controller, and during the acquisition process, using encoder pulses or timer signals as trigger sources input to the FPGA; the FPGA executes a trigger event, which generates a global trigger pulse for the FPGA and synchronously sends it to the imaging component and the ultrasonic probe; the FPGA timestamps the trigger event, so that both the image data acquired by the imaging component and the ultrasonic scanning data acquired by the ultrasonic probe carry the marked timestamp.

7. The dual-track vehicle-mounted multi-source track defect detection system according to claim 6, characterized in that, The processing device acquires the image data and ultrasonic scan data, and determines the defect status of the rail under test by combining the image data and ultrasonic scan data, specifically including: coordinate normalization processing: establishing a coordinate system with the rail under test as the reference, calibrating the imaging component and ultrasonic probe, and associating the imaging component and the ultrasonic probe to the coordinate system; ultrasonic feature vector. Construction: Extracting the echo signal from the ultrasonic scanning data Acquire the echo signal echo peak Echo energy echo centroid ; for the echo signal Perform FFT transform to obtain the echo frequency domain signal Acquire the echo frequency domain signal Spectral centroid and spectrum width Perform continuous wavelet transform on the echo signal and the echo frequency domain signal to obtain the time-frequency diagram. ,in, As a scale, For translation; extract within a specific scale range The maximum amplitude and energy are used as the ultrasonic feature vector. The specific scale range corresponds to the frequency of defect features; image feature vector Construction: A feature extraction model for extracting image features is constructed. Several preprocessed image data sets are input into the feature extraction model to extract feature vectors from the image data. The feature vectors are then weighted and summed to obtain the image feature vector. Feature fusion: For the ultrasonic feature vector... and image feature vectors Standardization was performed on each of the ultrasonic feature vectors. and image feature vectors Feature fusion is performed using a concatenation function to obtain a fused feature vector. Diagnostic output: The fused feature vector The results are input into a pre-trained diagnostic model to obtain the diagnostic results of the defects of the rail under test.

8. The dual-track vehicle-mounted multi-source track defect detection system according to claim 7, characterized in that, In the feature fusion step, principal component analysis is used to analyze the image feature vector. Dimensionality reduction is performed.

9. The dual-track vehicle-mounted multi-source track defect detection system according to claim 7 or 8, characterized in that, The diagnostic model includes an XGBoost model, and the training process of the diagnostic model includes the following steps: constructing a model containing... XGBoost model with regression trees: ;in, Indicates the first The original predicted scores for each sample. Indicates the first The fusion feature vector of each sample ， Indicates the first A regression tree, ， Represent the hypothesis space of the regression tree; construct the objective function of the XGBoost model: ， ,in, For loss function, For the regularization term of the regression tree, The number of leaf nodes in the regression tree. It is the fraction vector of the leaf nodes. and For hyperparameters; initialize constant prediction values, for the th In the next iteration, the loss function is calculated with respect to the current prediction result. First-order gradient and second-order gradient Based on dataset Greedily grow a new regression tree with the goal of minimizing the structure score. The structural fraction is represented as: , will the new regression tree Add to the XGBoost model: ,in, The learning rate is used; training is repeated until the XGBoost model meets the accuracy requirements or reaches the preset number of iterations.

10. The dual-track vehicle-mounted multi-source track defect detection system according to claim 9, characterized in that, The diagnostic model also includes a Softmax function, which obtains raw scores for the rail under test through the XGBoost model, and inputs these raw scores into the Softmax function to obtain the probability distribution of defects for the rail under test. Based on the obtained disease probability distribution The results of the defect status detection of the rail under test are obtained.

Images