A method and device for detecting rails on the track in operation

Through the combination of on-rail detection method and neural network model, real-time acquisition and analysis of rail ultrasonic signals is solved, and the problem that the existing technology cannot detect internal damage in the train is running is achieved, efficient and accurate rail damage detection is achieved, and the Cost and error.

CN113418987BActive Publication Date: 2024-12-03BEIJING RAILWAY ELECTRIFICATION SCHOOL
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110674395.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-06-17
Publication Date
2024-12-03
Estimated Expiration
2041-06-17

AI Technical Summary

Technical Problem

The existing rail detection methods cannot detect internal damage to the rails in real time when the train is running, resulting in large errors in the detection results and high costs, and manual interpretation has problems such as missed and missed inspections.

Method used

On-orbit detection method is adopted, train operation parameters and meteorological parameters are collected, and ultrasonic B-display image processing and neural network model (YOLO-v5) are automatically interpreted, and the multivariate linear regression model verification results are used to output rail damage information in real time.

Benefits of technology

Real-time detection under on-orbit operation is realized, detection accuracy and efficiency is improved, human error is reduced, manpower and material costs are saved, and the train central control system is adapted to meet personalized inspection needs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113418987B_ABST
    Figure CN113418987B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for detecting rails on-track in a running state, the detection method comprising the steps of information collection, image collection, image preprocessing, neural network model judgment and classification, multivariate linear regression model regression analysis, relative error calculation and information sending; the present invention realizes real-time detection of rail damage in a running state, improves the confidence and stability of the judgment by interpreting the rails using a neural network algorithm with deep learning capability, and uses a multivariate linear regression model to regress and analyze the judgment results, effectively avoiding the influence of train operation parameters and meteorological parameters on the judgment results, so that the detection results are closer to the actual situation; at the same time, the present invention also provides a detection device based on the method for detecting rails on-track, which effectively realizes compatibility, sharing and interconnection with a train control system through equipment mounting, wiring harness integration and program implantation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of rail detection, and in particular relates to an on-track detection method and detection equipment for a rail in a running state. Background Art

[0002] China's railway and subway transportation lines are busy, with short train running intervals and long operating mileage. The phenomenon of rails being overloaded and in service is serious, resulting in a high incidence of rail damage. In order to ensure the safety of railway transportation, nearly 10,000 full-time rail flaw detectors perform internal damage inspections on rails day and night every year, which consumes huge manpower and material resources. This has also become an important reason for the high cost of rail transportation operations.

[0003] At present, whether it is ordinary railway rails, high-speed railway rails, or urban underground tracks, the internal defects of the rails are diagnosed by combining daily manual inspections with regular inspections by local flaw detection vehicles. The inspection time is after the line completes daily transportation tasks and when there are no vehicles occupying the rail line. Due to the problem of differentiated professional levels in manual inspections, wrong inspections and missed inspections often occur. The inspection by local flaw detection vehicles has the problem of high cost of inspection equipment and high subsequent maintenance costs, which makes it difficult to meet the personalized needs of full inspection and on-time inspection.

[0004] However, as far as the internal damage of the rail is concerned, the width, length and depth of the damaged part can be fully detected by ultrasonic equipment, but the interpretation of the test results needs to rely on professional experience, and the results are often affected by various human factors and objective conditions such as weather, resulting in the inability to effectively control the interpretation error, which leads to distortion of the test results and a huge waste of time and cost. In recent years, with the rapid development of artificial intelligence and the substantial improvement of computer hyperthreading computing power, the use of artificial intelligence neural network algorithms to solve the intelligent interpretation of ultrasonic test results has become a direction that various research institutions in this field are competing to explore, and it is also the future development direction for solving the intelligent detection of in-service rails.

[0005] Chinese patent application No. 201720857163.0 discloses a fully automatic unmanned subway or urban rail track inspection device, including an inspection vehicle, which includes an inspection vehicle body with a running mechanism and an ultrasonic detection system, a video detection system, a track cross-section profile measurement system, a communication control system and a power supply system installed on the inspection vehicle body, and also includes a rescue vehicle connected to the inspection vehicle, the rescue vehicle at least includes a rescue vehicle body with a running mechanism, a communication control system installed on the rescue vehicle body and a power supply system installed on the rescue vehicle body, the inspection vehicle and the rescue vehicle are connected together through a quick docking and automatic detachment device, and also includes a track dust removal device, which is installed in front of the inspection vehicle. This device is still a type of bureau inspection vehicle, and can only be used for inspection when the railway is completely idle, and it is impossible to judge the track damage in the on-track running state at the first time. Summary of the invention

[0006] The purpose of the present invention is to provide a method and a detection device for detecting rails in operation, wherein the method comprises the following steps:

[0007] Step 1: Collect and store train operation parameter information and meteorological parameter information;

[0008] Step 2: Collect and store the rail ultrasonic B-display images to form an image set;

[0009] Step 3: preprocess the ultrasonic B-display image to obtain a centralized B-display lesion image to be determined;

[0010] Step 4: Input the centralized undetermined B-display damage image into the neural network model and output the calculation result;

[0011] Step 5: Collect the respective variable values ​​of the multivariate linear regression model at the time point corresponding to the calculation results, and substitute the respective variable values ​​into the multivariate linear regression model to solve the regression analysis value under the condition that the significance level meets the threshold;

[0012] Step 6: Calculate the relative error between the calculation result output by the neural network model and the regression analysis value, and determine the calculation result output by the neural network model whose relative error reaches the threshold requirement as the data of the final B-display damage image;

[0013] Step 7: Save the data determined as the final B-display damage image, annotate the clock information and GPS information, and then send it to the track maintenance organization platform.

[0014] Furthermore, the collecting of train operation parameter information and meteorological parameter information in step 1 includes:

[0015] Train operation parameter information includes the train's real-time running speed, instantaneous speed change rate, load, wheel-rail contact surface information and locomotive GPS signal; meteorological parameter information includes the train's outside temperature, humidity, wind speed, wind direction and air pressure information.

[0016] Furthermore, the acquisition and storage of the rail ultrasonic B-display images to form an image set in step 2 includes the following steps:

[0017] Step 2.1. According to the real-time speed v of the train t and the instantaneous velocity change rate a t , calculate the operating frequency f of the ultrasonic generator t , and start the clock timer, using the time axis as the coordinate to mark the calculation information, f t The calculation formula is as follows (1):

[0018]

[0019] In formula (1), Represents the instantaneous rate of change of velocity a t The corresponding adjustment factor, F s Represents the initial operating frequency of the ultrasonic generator, V S Represents the train speed corresponding to the initial operating frequency of the ultrasonic generator;

[0020] Step 2.2. The ultrasonic generator receives the working frequency f t The instruction drives the ultrasonic excitation unit according to the working frequency f t Excite ultrasonic waves;

[0021] Step 2.3. After receiving the echo signal, the ultrasonic receiving unit generates a B-display image through an encoder and saves it together with the corresponding time information and GPS information to a storage unit to form an original ultrasonic B-display image set.

[0022] Furthermore, the preprocessing of the ultrasonic B-display image in step 3 includes the following steps:

[0023] Step 3.1. Clean the ultrasonic B-mode image set, determine the images containing gray points as pending B-mode damaged images and retain them, and determine the images without gray points as non-damaged B-mode images and delete them;

[0024] Step 3.2. For the pending B-display damage image, determine the offset direction and angle of the pending B-display damage image according to the train load and wheel-rail contact surface information, and perform rotation correction and centering processing on the pending B-display damage image according to the offset;

[0025] Step 3.3. Crop the area containing all grayscale points in the pending B-display damage image into a rectangular image with a size of 1920*1080 to obtain a centralized pending B-display damage image.

[0026] Furthermore, the neural network model in step 4 adopts the YOLO-v5 model, which includes the following steps:

[0027] Step 4.1. De-noise the centralized undetermined B-display damaged image. First, use a bilateral mean filter to reduce the noise interference in the spatial domain of the image while maintaining the gray point edge texture blur information. Then, use a median filter to eliminate the salt and pepper noise in the image while retaining the sharpness of the gray point edge.

[0028] Step 4.2. Automatically annotate the grayscale points in the centralized pending B-display damage image, where the annotated area contains the grayscale point edge texture fuzzy information and is tangent to the arc with the maximum curvature of the grayscale point edge;

[0029] Step 4.3. Perform histogram equalization on the marked area, scan all pixels in the marked area, and calculate the normalized histogram H of the marked area. The calculation process is to first list the differential histogram H(r k ) expression, and then the differential histogram H(r k ) integration, the process is expressed as follows (2) and (3):

[0030]

[0031] H′(i)=∑ 0≤j≤i H(j)……(3),

[0032] In formula (2) and formula (3), n is the total number of images, k is the kth gray level in the image; n k is the number of gray levels; i represents the highest gray level in the region, and j represents the gray level variable in the region;

[0033] Finally, the image is transformed and expressed as the following formula (4):

[0034] dst(x,y)=H′(src(x,y))……(4);

[0035] Step 4.4. Use the neighborhood block Gaussian weighted method to perform local adaptive binary threshold segmentation on the image of the marked area, segment the image of the marked area into the area of ​​interest and the remaining area, and extract the area of ​​interest for further analysis;

[0036] Step 4.5. Perform median filtering on the area of ​​interest again. After obtaining an interference-free image, perform 3×3 kernel erosion first, and then perform 12×12 kernel dilation to complete the morphological operation of the image;

[0037] Step 4.6. The image after morphological operation is displayed as a discontinuous polygon. After contour scanning, curvature fitting is performed to obtain a curve with closed boundaries;

[0038] Step 4.7. Calculate the closed and smooth curve frame through the 33-layer network structure of ResNet in the YOLO-v5 model, and output the five-dimensional vector x, y, width, height, confidence of the rectangular frame tangent to the curve frame, where x, y represent the coordinates of the upper left point of the rectangular frame, width represents the maximum width of the curve frame, height represents the maximum height of the curve frame, and confidence is the confidence of the result;

[0039] Step 4.8. By determining the grayscale points in the pending B-display damage image, the rail damage information contained in the output result is classified. The damage types include: single-point damage, multi-point damage, and continuous-point damage.

[0040] Furthermore, the step 5 of solving the regression analysis value under the condition that the significance level meets the threshold condition includes the following steps:

[0041] Step 5.1. Establish a multiple linear regression model;

[0042] Step 5.1.1. Take the height or width value in the five-dimensional vector output by the YOLO-v5 model as the dependent variable Y, and select the physical vector closely related to the dependent variable Y as the independent variable X. Then the multiple linear regression equation is as follows (5):

[0043] Y=β0+β1X1+β2X2+β3X3+β4X4+β5X5+β6X6+β7X7+ε……(5),

[0044] In formula (5), β0, β1, ..., β7 are regression parameters, X1 is the real-time speed of the train, X2 is the train load, X3 is the ambient temperature, X4 is the atmospheric humidity, X5 is the atmospheric pressure, X6 is the wind speed, X7 is the wind direction, and ε is the random error at each measurement, satisfying the Gauss-Markov hypothesis, that is, ε~N(0,σ 2 );

[0045] Step 5.1.2. Collect n sets of independent experimental data and form a matrix equation group, as shown in the following equation (6):

[0046] Y=Xβ+ε……(6),

[0047] in,

[0048] Step 5.1.3. Use the least squares method to find the regression parameters.

[0049] It is known that the expression of the sum of squares of n observation errors Q is as follows (7):

[0050] Q=ε'ε=(Y-Xβ)'(Y-Xβ)......(7),

[0051] The smaller the Q value, the smaller the error. In the ideal state, Q = 0. Let Q be differentiated with respect to β, and we get the following formula (8):

[0052]

[0053] Further solving equation (8), we get the following equation (9):

[0054]

[0055] In formula (9), is the regression parameter matrix when Q takes the minimum value;

[0056] Thus, the empirical regression equation is obtained:

[0057]

[0058] Step 5.1.4. Perform a significance test on the empirical regression equation. If the significance level reaches the threshold, the empirical equation is determined to be a multiple linear regression model;

[0059] Step 5.2. According to the time point coordinates, collect the X values ​​of each independent variable of the multivariate linear regression model corresponding to the time point of the YOLO-v5 model output value, substitute it into the model, and solve the regression analysis value when the significance level meets the threshold condition

[0060] Furthermore, the step 6 of determining the data of the final B-display damage image comprises the following steps:

[0061] Step 6.1. Calculate the relative error between the YOLO-v5 model output value and the regression analysis value, as shown in the following formula (11):

[0062]

[0063] Step 6.2. When the average value of the relative error is less than 2%, the YOLO-v5 model output value is determined to be credited; when the average value of the relative error is ≥2% and ≤5%, the YOLO-v5 model output value is determined to be pending; when the average value of the relative error is greater than 5%, the YOLO-v5 model output value is determined to be invalid;

[0064] Step 6.3. The B-display damaged image corresponding to the YOLO-v5 model output value determined to be trusted and to be checked is determined as the final B-display damaged image, and the B-display damaged image corresponding to the YOLO-v5 model output value determined to be invalid is deleted.

[0065] The present invention further provides an on-track detection device based on the rail, the detection device comprises: a locomotive state information acquisition module, an environmental parameter detection module, an image acquisition module, a laptop computer and an integrated wiring harness, wherein the locomotive state information acquisition module is arranged above the locomotive head steering knuckle arm, and is used to collect the real-time running speed, instantaneous speed change rate, load, wheel-rail contact surface information and locomotive GPS signal of the train, the environmental parameter detection module is arranged above the locomotive head steering mechanism, and is used to collect the temperature, humidity, wind speed, wind direction and air pressure information outside the train, the image acquisition module is composed of a timer, a frequency conversion controller unit, an ultrasonic signal excitation unit, an ultrasonic signal receiving unit and a signal encoder connected in series, and the image acquisition module is arranged at the axle box of the bogie; the laptop computer comprises a central processing unit, a memory and a wireless network modem, the laptop computer is arranged in the locomotive cab, and the memory is preset with an image preprocessing tool, a YOLO-v5 model and a multivariate linear regression model; the locomotive state information acquisition module, the environmental parameter detection module, the image acquisition module and the computer are connected through an integrated wiring harness.

[0066] Compared with the prior art, the on-track rail detection method and detection equipment of the present invention have the following beneficial effects:

[0067] 1. The on-track rail detection method and detection equipment of the present invention collects real-time ultrasonic signals of rails when the train is running, and converts the ultrasonic signals into visual B-display images for analysis. Such image information effectively covers the influence of train speed and load on internal damage of rails, can detect internal defects of rails in locomotive and vehicle loaded state in real time, reflect the real state of rails in real time, and effectively make up for the defect that the local flaw detection vehicle cannot simulate the aggravation effect of internal damage of rails under the running state of trains.

[0068] 2. The on-track rail detection method and detection equipment described in the present invention adopt a neural network model with deep self-learning ability to analyze and judge the damage of the rail B-display image. With the continuous accumulation of data, the confidence of the judgment will become higher and higher, thereby effectively avoiding the problem of missed detection and false detection caused by human errors, and the work efficiency is high and the output results are stable. At the same time, the neural network model can meet the efficiency requirements of work intensity in a real-time environment, and there is no need to choose to stop the train before implementing the detection, which provides technical support for saving a lot of manpower, material resources and time costs.

[0069] 3. The on-track rail detection method and detection equipment described in the present invention fully consider the physical vectors closely related to the influence of rail damage, including train speed, load, temperature, humidity, air pressure, wind speed and wind direction. Since these factors are not correlated with each other and a linear relationship is formed between a single variable and the internal damage of the rail, the multivariate linear regression model thus formed can further verify the confidence of the output results of the neural network model, thereby making the on-track rail detection method described in the present invention more scientific and comprehensive.

[0070] 4. The on-track rail detection method and detection equipment described in the present invention rely on the train central control system. Through equipment installation and program implantation, they effectively achieve compatibility with the train central control system. There is no need to install them separately from the train control system. It only needs to standardize the integrated wiring harness and the train vehicle CAN communication bus for compatibility. In this way, data sharing and interconnection between the detection equipment information and the train central control system information can be achieved. Therefore, the present invention has strong adaptability and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 The present invention is a flow chart of the on-track rail detection method.

[0072] Figure 2 It is a schematic structural diagram of the on-track rail detection equipment of the present invention.

[0073] In the figure, 1-locomotive status information acquisition module, 2-environmental parameter detection module, 3-image acquisition module, 3.1-timer, 3.2-frequency controller unit, 3.3-ultrasonic signal excitation unit, 3.4-ultrasonic signal receiving unit, 3.5-signal encoder, 4-laptop computer, 4.1-central processing unit, 4.2-memory, 4.3-wireless network modem, 5-integrated wiring harness. DETAILED DESCRIPTION

[0074] Now, in conjunction with the accompanying drawings of the specification, the specific implementation methods of the on-track rail detection method and detection equipment of the present invention are described in detail.

[0075] like Figure 1 As shown, a method for detecting rail on-track in a running state and a detection device thereof, the method for detecting rail on-track comprises the following steps:

[0076] Step 1: Collect and store train operation parameter information and meteorological parameter information;

[0077] Step 2: Collect and store the rail ultrasonic B-display images to form an image set;

[0078] Step 3: preprocess the ultrasonic B-display image to obtain a centralized B-display lesion image to be determined;

[0079] Step 4: Input the centralized undetermined B-display damage image into the neural network model and output the calculation result;

[0080] Step 5: Collect the respective variable values ​​of the multivariate linear regression model at the time point corresponding to the calculation results, and substitute the respective variable values ​​into the multivariate linear regression model to solve the regression analysis value under the condition that the significance level meets the threshold;

[0081] Step 6: Calculate the relative error between the calculation result output by the neural network model and the regression analysis value, and determine the calculation result output by the neural network model whose relative error reaches the threshold requirement as the data of the final B-display damage image;

[0082] Step 7: Save the data determined as the final B-display damage image, annotate the clock information and GPS information, and then send it to the track maintenance organization platform.

[0083] Furthermore, the collecting of train operation parameter information and meteorological parameter information in step 1 includes:

[0084] Train operation parameter information includes the train's real-time running speed, instantaneous speed change rate, load, wheel-rail contact surface information and locomotive GPS signal; meteorological parameter information includes the train's external temperature, humidity, wind speed, wind direction and air pressure information.

[0085] Furthermore, the acquisition and storage of the rail ultrasonic B-display images to form an image set in step 2 includes the following steps:

[0086] Step 2.1. According to the real-time speed v of the train t and the instantaneous velocity change rate a t , calculate the operating frequency f of the ultrasonic generator t , and start the clock timer, using the time axis as the coordinate to mark the calculation information, f t The calculation formula is as follows (1):

[0087]

[0088] In formula (1), Represents the instantaneous rate of change of speed a t The corresponding adjustment factor, F s Represents the initial operating frequency of the ultrasonic generator, V S Represents the train speed corresponding to the initial operating frequency of the ultrasonic generator;

[0089] Step 2.2. The ultrasonic generator receives the working frequency f t The instruction drives the ultrasonic excitation unit according to the working frequency f t Excite ultrasonic waves;

[0090] Step 2.3. After receiving the echo signal, the ultrasonic receiving unit generates a B-display image through an encoder and saves it together with the corresponding time information and GPS information to a storage unit to form an original ultrasonic B-display image set.

[0091] Furthermore, the preprocessing of the ultrasonic B-display image in step 3 includes the following steps:

[0092] Step 3.1. Clean the ultrasonic B-mode image set, determine the images containing gray points as pending B-mode damaged images and retain them, and determine the images without gray points as non-damaged B-mode images and delete them;

[0093] Step 3.2. For the pending B-display damage image, determine the offset direction and angle of the pending B-display damage image according to the train load and wheel-rail contact surface information, and perform rotation correction and centering processing on the pending B-display damage image according to the offset;

[0094] Step 3.3. Crop the area containing all grayscale points in the pending B-display damage image into a rectangular image with a size of 1920*1080 to obtain a centralized pending B-display damage image.

[0095] Furthermore, the neural network model in step 4 adopts the YOLO-v5 model, which includes the following steps:

[0096] Step 4.1. De-noise the centralized undetermined B-display damaged image. First, use a bilateral mean filter to reduce the noise interference in the spatial domain of the image while maintaining the gray point edge texture blur information. Then, use a median filter to eliminate the salt and pepper noise in the image while retaining the sharpness of the gray point edge.

[0097] Step 4.2. Automatically annotate the grayscale points in the centralized pending B-display damage image, where the annotated area contains the grayscale point edge texture fuzzy information and is tangent to the arc with the maximum curvature of the grayscale point edge;

[0098] Step 4.3. Perform histogram equalization on the marked area, scan all pixels in the marked area, and calculate the normalized histogram H of the marked area. The calculation process is to first list the differential histogram H(r k ) expression, and then the differential histogram H(r k ) integration, the process is expressed as follows (2) and (3):

[0099]

[0100] H′(i)=∑ 0≤j≤i H(j)……(3),

[0101] In formula (2) and formula (3), n is the total number of images, k is the kth gray level in the image; n k is the number of gray levels; i represents the highest gray level in the region, and j represents the gray level variable in the region;

[0102] Finally, the image is transformed and expressed as the following formula (4):

[0103] dst(x,y)=H′(src(x,y))……(4);

[0104] Step 4.4. Use the neighborhood block Gaussian weighted method to perform local adaptive binary threshold segmentation on the image of the marked area, segment the image of the marked area into the area of ​​interest and the remaining area, and extract the area of ​​interest for further analysis;

[0105] Step 4.5. Perform median filtering on the area of ​​interest again. After obtaining an interference-free image, perform 3×3 kernel erosion first, and then perform 12×12 kernel dilation to complete the morphological operation of the image;

[0106] Step 4.6. The image after morphological operation is displayed as a discontinuous polygon. After contour scanning, curvature fitting is performed to obtain a curve with closed boundaries;

[0107] Step 4.7. Calculate the closed and smooth curve frame through the 33-layer network structure of ResNet in the YOLO-v5 model, and output the five-dimensional vector x, y, width, height, confidence of the rectangular frame tangent to the curve frame, where x, y represent the coordinates of the upper left point of the rectangular frame, width represents the maximum width of the curve frame, height represents the maximum height of the curve frame, and confidence is the confidence of the result;

[0108] Step 4.8. By determining the grayscale points in the pending B-display damage image, the rail damage information contained in the output result is classified. The damage types include: single-point damage, multi-point damage, and continuous-point damage.

[0109] Furthermore, the step 5 of solving the regression analysis value under the condition that the significance level meets the threshold condition includes the following steps:

[0110] Step 5.1. Establish a multiple linear regression model;

[0111] Step 5.1.1. Take the height or width value in the five-dimensional vector output by the YOLO-v5 model as the dependent variable Y, and select the physical vector closely related to the dependent variable Y as the independent variable X. Then the multiple linear regression equation is as follows (5):

[0112] Y=β0+β1X1+β2X2+β3X3+β4X4+β5X5+β6X6+β7X7+ε……(5),

[0113] In formula (5), β0, β1, ..., β7 are regression parameters, X1 is the real-time speed of the train, X2 is the train load, X3 is the ambient temperature, X4 is the atmospheric humidity, X5 is the atmospheric pressure, X6 is the wind speed, X7 is the wind direction, and ε is the random error at each measurement, satisfying the Gauss-Markov hypothesis, that is, ε~N(0,σ 2 );

[0114] Step 5.1.2. Collect n sets of independent experimental data and form a matrix equation group, as shown in the following equation (6):

[0115] Y=Xβ+ε……(6),

[0116] in,

[0117] Step 5.1.3. Use the least squares method to find the regression parameters.

[0118] It is known that the expression of the sum of squares of n observation errors Q is as follows (7):

[0119] Q=ε'ε=(Y-Xβ)'(Y-Xβ)......(7),

[0120] The smaller the Q value, the smaller the error. In the ideal state, Q = 0. Let Q be differentiated with respect to β, and we get the following formula (8):

[0121]

[0122] Further solving equation (8), we get the following equation (9):

[0123]

[0124] In formula (9), is the regression parameter matrix when Q takes the minimum value;

[0125] Thus, the empirical regression equation is obtained:

[0126]

[0127] Step 5.1.4. Perform a significance test on the empirical regression equation. If the significance level reaches the threshold, the empirical equation is determined to be a multiple linear regression model;

[0128] Step 5.2. According to the time point coordinates, collect the X values ​​of each independent variable of the multivariate linear regression model corresponding to the time point of the YOLO-v5 model output value, substitute it into the model, and solve the regression analysis value when the significance level meets the threshold condition

[0129] Furthermore, the step 6 of determining the data of the final B-display damage image comprises the following steps:

[0130] Step 6.1. Calculate the relative error between the YOLO-v5 model output value and the regression analysis value, as shown in the following formula (11):

[0131]

[0132] Step 6.2. When the average value of the relative error is less than 2%, the YOLO-v5 model output value is determined to be credited; when the average value of the relative error is ≥2% and ≤5%, the YOLO-v5 model output value is determined to be pending; when the average value of the relative error is greater than 5%, the YOLO-v5 model output value is determined to be invalid;

[0133] Step 6.3. The B-display damaged image corresponding to the YOLO-v5 model output value determined to be trusted and to be checked is determined as the final B-display damaged image, and the B-display damaged image corresponding to the YOLO-v5 model output value determined to be invalid is deleted.

[0134] like Figure 2 As shown, the present invention further provides an on-track rail detection device based on the above-mentioned rail, and the detection device comprises: a locomotive state information acquisition module 1, an environmental parameter detection module 2, an image acquisition module 3, a laptop computer 4 and an integrated wiring harness 5, wherein the locomotive state information acquisition module 1 is arranged above the locomotive head steering knuckle arm to collect the real-time running speed, instantaneous speed change rate, load, wheel-rail contact surface information and locomotive GPS signal of the train, the environmental parameter detection module 2 is arranged above the locomotive head steering mechanism to collect the temperature, humidity, wind speed, wind direction and air pressure information outside the train, and the image acquisition module 3 is sequentially controlled by the timer 3.1, The frequency conversion controller unit 3.2, the ultrasonic signal excitation unit 3.3, the ultrasonic signal receiving unit 3.4 and the signal encoder 3.5 are connected in series, and the image acquisition module 3 is arranged at the axle box of the bogie; the laptop computer 4 includes a central processing unit 4.1, a memory 4.2 and a wireless network modem 4.3, and the laptop computer 4 is arranged in the locomotive cab. The memory 4.2 is preset with an image preprocessing tool, a YOLO-v5 model and a multivariate linear regression model; the locomotive state information acquisition module 1, the environmental parameter detection module 2, the image acquisition module 3 and the laptop computer 4 are connected by an integrated wiring harness 5.

[0135] The present invention is not limited to the above-mentioned embodiments. Without departing from the essential content of the present invention, any deformation, improvement and substitution that can be thought of by those skilled in the art shall fall into the protection scope of the present invention.

Claims

1. A method for detecting rails on the track in a running state, characterized in that: The rail on-track detection method comprises the following steps: Step 1: Collecting and storing train operation parameter information and meteorological parameter information. Collecting train operation parameter information and meteorological parameter information includes: Train operation parameter information includes real-time train speed, instantaneous speed change rate, train load, wheel-rail contact surface information and locomotive GPS signal; Meteorological parameter information includes train exterior temperature, humidity, wind speed, wind direction and air pressure information; Step 2: Collect and store the rail ultrasonic B-display images to form an image set, including the following steps: Step 2.

1. According to the real-time speed v of the train t and the instantaneous velocity change rate a t , calculate the operating frequency f of the ultrasonic generator t , and start the clock timer, using the time axis as the coordinate to mark the calculation information, f t The calculation formula is as follows (1): In formula (1), Represents the instantaneous rate of change of speed a t The corresponding adjustment factor, F s Represents the initial operating frequency of the ultrasonic generator, V S Represents the train speed corresponding to the initial operating frequency of the ultrasonic generator; Step 2.

2. The ultrasonic generator receives the working frequency f t The instruction drives the ultrasonic excitation unit according to the working frequency f t Excite ultrasonic waves; Step 2.

3. After receiving the echo signal, the ultrasonic receiving unit generates a B-display image through an encoder and saves it together with the corresponding time information and GPS information to a storage unit to form an original ultrasonic B-display image set; Step 3: preprocess the ultrasonic B-display image to obtain a centralized B-display lesion image to be determined; Step 4: Input the centralized undetermined B-display damage image into the neural network model, and use the YOLO-v5 neural network model to output the calculation results; Step 5: Collect the respective variable values ​​of the multivariate linear regression model at the time point corresponding to the calculation results, and substitute the respective variable values ​​into the multivariate linear regression model to solve the regression analysis value under the condition that the significance level meets the threshold; Step 6: Calculate the relative error between the calculation result output by the neural network model and the regression analysis value, and determine the calculation result output by the neural network model whose relative error reaches the threshold requirement as the data of the final B-display damage image; Step 7: Save the data determined as the final B-display damage image, annotate the clock information and GPS information, and then send it to the track maintenance organization platform.

2. The on-track rail detection method according to claim 1, characterized in that: The preprocessing of the ultrasonic B-display image in step 3 includes the following steps: Step 3.

1. The original ultrasound B-mode image set occupies a large storage space. In order to compress the space and reduce the amount of calculation, the original ultrasound B-mode image set needs to be cleaned. Since the ultrasound echo signal reflecting the damaged part is shown as a gray point in the B-mode image, the image containing the gray point is determined as a pending B-mode damaged image and is retained, and the image not containing the gray point is determined as a non-damaged B-mode image and is deleted; Step 3.

2. For the pending B-display damage image, determine the offset direction and angle of the pending B-display damage image according to the train load and wheel-rail contact surface information, and perform rotation correction and centering processing on the pending B-display damage image according to the offset; Step 3.

3. Crop the area containing all grayscale points in the pending B-display damage image into a rectangular image with a size of 1920*1080, thereby obtaining a centralized pending B-display damage image.

3. The on-track rail detection method according to claim 2, characterized in that: The neural network model used in step 4 is the YOLO-v5 model, which includes the following steps: Step 4.

1. De-noise the centralized undetermined B-display damaged image. First, use a bilateral mean filter to reduce the noise interference in the spatial domain of the image while maintaining the gray point edge texture blur information. Then, use a median filter to eliminate the salt and pepper noise in the image while retaining the sharpness of the gray point edge. Step 4.

2. Automatically annotate the grayscale points in the centralized pending B-display damage image, where the annotated area contains the grayscale point edge texture fuzzy information and is tangent to the arc with the maximum curvature of the grayscale point edge; Step 4.

3. Perform histogram equalization on the marked area, scan all pixels in the marked area, and calculate the normalized histogram H of the marked area. The calculation process is to first list the differential histogram H(r k ) expression, and then the differential histogram H(r k ) integration, the process is expressed as follows (2) and (3): H′(i)=∑ 0≤j≤i H(j)……(3), In formula (2) and formula (3), n is the total number of images, k is the kth gray level in the image; n k is the number of gray levels; i represents the highest gray level in the region, and j represents the gray level variable in the region; Finally, the image is transformed and expressed as the following formula (4): dst(x,y)=H′(src(x,y))……(4); Step 4.

4. Use the neighborhood block Gaussian weighted method to perform local adaptive binary threshold segmentation on the image of the marked area, segment the image of the marked area into the area of ​​interest and the remaining area, and extract the area of ​​interest for further analysis; Step 4.

5. Perform median filtering on the area of ​​interest again. After obtaining an interference-free image, perform 3×3 kernel erosion first, and then perform 12×12 kernel dilation to complete the morphological operation of the image; Step 4.

6. The image after morphological operation is displayed as a discontinuous polygon. After contour scanning, curvature fitting is performed to obtain a curve with closed boundaries; Step 4.

7. Calculate the closed and smooth curve frame through the 33-layer network structure of ResNet in the YOLO-v5 model, and output the five-dimensional vector x, y, width, height, confidence of the rectangular frame tangent to the curve frame, where x, y represent the coordinates of the upper left point of the rectangular frame, width represents the maximum width of the curve frame, height represents the maximum height of the curve frame, and confidence is the confidence of the result; Step 4.

8. By determining the grayscale points in the pending B-display damage image, the rail damage information contained in the output result is classified. The damage types include: single-point damage, multi-point damage, and continuous-point damage.

4. The on-track rail detection method according to claim 3, characterized in that: Step 5 of solving the regression analysis value under the condition that the significance level meets the threshold includes the following steps: Step 5.

1. Build a multiple linear regression model Step 5.1.

1. Take the height or width value in the five-dimensional vector output by the YOLO-v5 model as the dependent variable Y, and select the physical vector closely related to the dependent variable Y as the independent variable X. Then the multiple linear regression equation is as follows (5): Y=β0+β1X1+β2X2+β3X3+β4X4+β5X5+β6X6+β7X7+ε……(5), In formula (5), β0, β1, …, β7 are regression parameters, X1 is the real-time speed of the train, X2 is the train load, X3 is the ambient temperature, X4 is the atmospheric humidity, X5 is the atmospheric pressure, X6 is the wind speed, X7 is the wind direction, and ε is the random error at each measurement, which satisfies the Gauss-Markov hypothesis, that is, ε~N(0,σ 2 ); Step 5.1.

2. Collect n sets of independent experimental data and form a matrix equation group, as shown in the following equation (6): Y=Xβ+ε……(6), in, Step 5.1.

3. Use the least squares method to find the regression parameters. It is known that the expression of the sum of squares of n observation errors Q is as follows (7): Q=ε'ε=(Y-Xβ)'(Y-Xβ)......(7), In the formula, the smaller the Q value is, the smaller the error is. In the ideal state, Q=0. Let Q be differentiated with respect to β, and we get the following formula (8): Further solving, we get the following formula (9): In the formula, is the regression parameter matrix when Q takes the minimum value; Thus, the empirical regression equation is obtained: In the formula, is the regression analysis value; Step 5.1.

4. Perform a significance test on the empirical regression equation. If the significance level reaches the threshold, the empirical equation is determined to be a multiple linear regression model; Step 5.

2. According to the time point coordinates, collect the X values ​​of each independent variable of the multivariate linear regression model and substitute them into the model to solve the regression analysis value when the significance level meets the threshold condition.

5. The on-track rail detection method according to claim 1, characterized in that: Determining the data of the final B-display damage image in step 6 includes the following steps: Step 6.

1. Calculate the relative error between the YOLO-v5 model output value and the regression analysis value, as shown in the following formula (11): Step 6.

2. When the average value of the relative error is less than 2%, the YOLO-v5 model output value is determined to be credited; when the average value of the relative error is ≥2% and ≤5%, the YOLO-v5 model output value is determined to be pending; when the average value of the relative error is greater than 5%, the YOLO-v5 model output value is determined to be invalid; Step 6.

3. The B-display damaged image corresponding to the YOLO-v5 model output value determined to be trusted and to be checked is determined as the final B-display damaged image, and the B-display damaged image corresponding to the YOLO-v5 model output value determined to be invalid is deleted.

6. A detection device based on the on-track rail detection method according to any one of claims 1 to 5, the detection device comprising: A locomotive state information acquisition module (1), an environmental parameter detection module (2), an image acquisition module (3), a laptop computer (4) and an integrated wiring harness (5), wherein the locomotive state information acquisition module (1) is arranged above the locomotive head steering knuckle arm, the environmental parameter detection module (2) is arranged above the locomotive head steering mechanism, the image acquisition module (3) is sequentially composed of a timer (3.1), a frequency conversion controller unit (3.2), an ultrasonic signal excitation unit (3.3), an ultrasonic signal receiving unit (3.4) and a signal encoder (3.5) connected in series, and the image acquisition module (3) is arranged at the axle box of the bogie; the laptop computer (4) includes a central processing unit (4.1), a memory (4.2) and a wireless network modem (4.3), the laptop computer (4) is arranged in the locomotive cab, and the memory (4.2) is preset with an image processing tool, a YOLO-v5 model and a multivariate linear regression model; the locomotive state information acquisition module (1), the environmental parameter detection module (2), the image acquisition module (3) and the laptop computer (4) are connected via an integrated wiring harness (5).