Steel rail detection equipment and method based on deep fusion of steel rail and axle box vibration data

Abstract

The invention discloses a steel rail detection device and method based on deep fusion of steel rail and axle box vibration data, and relates to the technical field of steel rail detection. The first vibration detection assembly is installed on a steel rail with welded joint damage to collect steel rail vibration data corresponding to the welded joint damage position. The second steel rail disease information acquisition assembly comprises a second vibration detection assembly and a train position detection assembly; the second vibration detection assembly is installed on an axle box of the train so as to acquire axle box vibration data when the train runs. The train position detection assembly is installed on an axle box of a train to detect advancing distance data of the train. The control system is electrically connected with the first vibration detection assembly, the second vibration detection assembly and the train position detection assembly. According to the invention, convenient, efficient and intelligent detection of steel rail welding joint diseases can be realized, and the rail maintenance efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of rail detection, and in particular to a rail detection device and method based on deep fusion of rail and axle box vibration data. Background Art

[0002] In recent years, my country's rail transit has developed rapidly, but it also faces many engineering challenges. The large-scale rail transit network has put forward urgent demands for operational safety and efficiency. Tracks are the basis for safe and smooth train operation, and are also the focus of maintenance. As an important component of the track system, rails directly bear all the loads of train operation. Poor welding quality at the weld will cause geometric unevenness of the rails. Under the continuous impact of the wheels, rail wear and contact fatigue will increase, and the unevenness at the weld will deteriorate. Rail welding joint defects will deteriorate the wheel-rail contact relationship, greatly aggravate the dynamic interaction between the wheel and the rail, and lead to a series of problems such as loose fasteners, wheel-rail fatigue damage, and vibration and noise of the wheel-rail system. If minor rail defects are not treated in time, they may develop rapidly to a certain extent and endanger the safety of the track system. Severe rail defects will cause strong vibration and noise in the train track system, shorten the safe service life of train and track-related components, and reduce the safety, comfort and stability of train operation. Therefore, the control of the surface condition of the rail is crucial to train safety and operational comfort.

[0003] With the continuous expansion of the scale of my country's rail transit network, the increase in operating time and the increase in operating speed, track disease problems have gradually emerged in on-site operation practice, and the maintenance tasks of railways have become increasingly heavy. Among them, the monitoring of the health status of rails is a key link in maintenance work. Traditionally, rail disease detection and monitoring mainly rely on manual inspections, which is inefficient and labor-intensive, and is contrary to the current development trend of automation and intelligence in the rail transit industry. At present, there are relatively few studies on intelligent monitoring of rail diseases in rail transit. In order to effectively solve the problem of intelligent monitoring of rail welding joint diseases in rail transit systems, it is urgent to carry out relevant equipment and method research. Intelligent monitoring technology for rail status relies on various track inspection data, and different types of data have their own characteristics. The vibration acceleration of the train axle box can reflect the status of the entire track, while the vibration acceleration of the rail can directly reflect the impact of rail diseases. It can be seen that if the correlation between the vibration acceleration of the axle box and the vibration acceleration of the rail can be established, and the characteristics of the train axle box vibration acceleration signal corresponding to the rail disease can be extracted, the convenience and efficiency of rail disease detection will be greatly improved.

[0004] Therefore, how to propose a rail detection device and method based on the deep fusion of rail and axle box vibration data, which can be used to solve the problem of convenient, efficient and intelligent detection of rail welding joint defects, improve track maintenance efficiency, and effectively ensure the safety of track structure is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0005] In view of this, the present invention proposes a rail detection device and method based on deep fusion of rail and axle box vibration data, aiming to solve the technical problems of high labor intensity and low efficiency in the above-mentioned traditional rail weld joint disease detection.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] One aspect of the present invention provides a rail detection device based on deep fusion of rail and axle box vibration data, comprising a track and a train running on the track, the track comprising a rail assembly and a sleeper, the rail assembly comprising a plurality of rails welded and connected in sequence; and further comprising:

[0008] A rail defect information collection component 1, the rail defect information collection component 1 includes a vibration detection component 1, the vibration detection component 1 is installed on the rail with weld joint defects to collect rail vibration data corresponding to the weld joint defect position;

[0009] A rail defect information collection component 2, the rail defect information collection component 2 includes a vibration detection component 2 and a train position detection component; the vibration detection component 2 is installed on the axle box of the train to collect and obtain the axle box vibration data when the train is running; the train position detection component is installed on the axle box of the train to detect the travel distance data of the train;

[0010] A control system, wherein the control system is electrically connected to the vibration detection component 1, the vibration detection component 2 and the train position detection component.

[0011] The rail detection equipment based on the deep fusion of rail and axle box vibration data of the present invention collects rail vibration data by arranging a vibration detection component on the rail with welded joint defects. When the rail has welded joint defects, the wheel-rail interaction when the train passes is different from that of ordinary rails, and the wheel-rail impact excitation is more intense, causing the rail to vibrate more significantly. In addition, the rail vibration characteristics caused by different rail defects are different, and there are obvious differences in the spectrum characteristics, which are relatively easy to distinguish; by measuring the rail vibration acceleration in the section with rail defects, track defects can be directly identified. Axle box vibration data is collected by arranging a vibration detection component 2 on the axle box of the train. When the train passes through a defective rail, the wheel is subjected to abnormal impact excitation, and the vibration magnitude generated is larger. Different defects will cause different vibration response characteristics. In this way, the track defects can be indirectly identified by monitoring the axle box vibration acceleration through the vibration detection component 2. Through the axle box vibration acceleration detection, the rapid detection of the entire track line can be achieved, and the state of the rail welding joints of the line can be fully reflected. The above data are aligned on the time axis by the synchronous marking method. When the train equipped with the vibration detection component 2 travels to the position interval of the vibration detection component 1 installed on the rail, the axle box vibration acceleration and the rail vibration acceleration data are synchronously marked, and the acceleration data are aligned on the time axis with the marked point as the starting point, so as to establish the correlation between the rail and the axle box vibration acceleration corresponding to the defect position of the rail welding joint, and then the defect state of the rail can be detected by measuring the vibration acceleration of the train axle box, which greatly reduces the workload of rail defect detection and improves the efficiency of rail state detection.

[0012] As a further improvement of the above technical solution, the control system includes a data collector, an industrial computer and a remote workstation; the data collector is electrically connected to the vibration detection component 1, the vibration detection component 2 and the train position detection component to receive and store the rail vibration data, the axle box vibration data and the travel distance data; the industrial computer is electrically connected to the data collector; and the remote workstation is wirelessly connected to the industrial computer.

[0013] As a further improvement of the above technical solution, the vibration detection component 1 includes multiple acceleration sensors 1; the acceleration sensor 1 is installed on the bottom of the defective rail; the vibration detection component 2 includes acceleration sensor 2; the acceleration sensor 2 is installed on the axle box; the acceleration sensor 1 and the acceleration sensor 2 are both electrically connected to the data collector.

[0014] As a further improvement of the above technical solution, the acceleration sensor 1 is installed at the position of the rail fastener corresponding to the defective rail bottom and at the middle position corresponding to two adjacent rail fasteners; there are multiple acceleration sensors 2, and each axle box of the train is installed with the acceleration sensor 2.

[0015] As a further improvement of the above technical solution, the train position detection component includes a wheel speed sensor and a sleeper counting sensor; the wheel speed sensor is installed inside the axle box and corresponds to the wheel shaft of the train to detect the wheel speed; the sleeper counting sensor is installed on the axle box and corresponds to the track to measure the number of sleepers on the track passed by the train during its travel;

[0016] The data collector is electrically connected to the wheel speed sensor and the sleeper counting sensor.

[0017] As a further improvement of the above technical solution, the remote workstation includes a positioning module, which is used to analyze and process the data collected by the wheel speed sensor and the sleeper counting sensor and the full-line track laying map data of the track to accurately locate the defective position of the rail assembly at the sleeper level.

[0018] Another aspect of the present invention provides a rail detection method based on deep fusion of rail and axle box vibration data, including the rail detection device based on deep fusion of rail and axle box vibration data; the detection steps include:

[0019] S1. Establish a database corresponding to rail vibration time-frequency signals and rail welding joint defects, and use GRU neural network for feature capture;

[0020] S2, collect the axle box vibration acceleration signals of the train during the whole track operation, and use Alexnet convolutional neural network to screen the axle box vibration acceleration data corresponding to the rail welding joints of the whole track;

[0021] S3, input the axle box vibration acceleration data corresponding to the rail welding joint filtered in step S2 into the GRU neural network, perform feature capture, perform similarity analysis with the rail vibration characteristics in step S1, and output the corresponding rail welding joint status.

[0022] As a further improvement of the above technical solution, S1 specifically includes the following steps:

[0023] S11. Manually investigate the track status and find the location of rail welding joint defects;

[0024] S12, installing a vibration detection component 1 on the rail corresponding to the position of the rail welding joint defect in step S11, and collecting the rail vibration acceleration signal;

[0025] S13, using wavelet transform to extract the time-frequency distribution characteristics of the rail vibration acceleration signal collected in step S12, and establishing a rail vibration database under the condition of rail welding joint defects;

[0026] S14, using the rail vibration database in step S13 to train the GRU neural network to capture the rail vibration characteristics when the rail welding joint is in a defective state;

[0027] S2 specifically includes the following sub-steps:

[0028] S21, collecting axle box vibration acceleration signals using a second vibration detection component installed on the axle box of the train;

[0029] S22, using wavelet transform to extract the time-frequency distribution characteristics of the axle box vibration acceleration signal collected in step S21;

[0030] S23, manually intercepting and marking multiple welded joint and non-welded joint data segments from the data in steps S21 and S22, forming a sample library to perform migration training on the Alexnet convolutional neural network;

[0031] S24, using the Alexnet convolutional neural network in step S23 to screen the axle box vibration acceleration data corresponding to the rail welding joints of the entire track;

[0032] Among them, step S21 also includes: according to the synchronization mark, intercepting the axle box vibration acceleration signal corresponding to the time window of the train passing through the defective position of the rail welding joint aligned with the time axis; subsequently analyzing the axle box vibration acceleration signal in this time window to ensure the corresponding relationship between the rail vibration acceleration and the axle box vibration acceleration.

[0033] As a further improvement of the above technical solution, in S3, the captured features generate corresponding axle box vibration acceleration feature vectors; the similarity analysis specifically calculates the similarity score by combining the Euclidean distance and the cosine similarity, determines the similarity between the signals, and establishes the correlation between the vibration features of the rail and the axle box when there are rail defects, so as to perform deep fusion of the vibration data of the rail and the axle box;

[0034] The currently identified axle box vibration acceleration signal is positioned in the following way: a rough travel distance is obtained through the wheel speed sensor, and then the travel distance is corrected in combination with the sleeper counting sensor and the full-line track laying map data to obtain the precise travel distance at the sleeper level.

[0035] As a further improvement of the above technical solution, the specific process of obtaining the rough travel distance is as follows: using a wheel speed sensor to measure and record the rotational angular velocity of the wheel rolling, and multiplying the rotational angular velocity by the radius of the wheel rolling circle to obtain the vehicle speed; by integrating the vehicle speed, the train running distance at any time is obtained, and combined with the initial travel distance, the travel distance information at any time is obtained;

[0036] The specific process of obtaining the precise travel distance at the sleeper level is as follows: the sleepers are scanned and counted by the sleeper counting sensor installed on the axle box; finally, the precise travel distance information is obtained by combining the track laying map data of the entire track line, thereby realizing the precise sleeper-level positioning of the defective position of the rail welding joint.

[0037] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a rail detection device and method based on deep fusion of rail and axle box vibration data, which has the following advantages and beneficial effects:

[0038] The present invention adopts a method for deep fusion of train and track vibration data, establishes the correlation between rail and axle box vibration acceleration signals through a GRU neural network, and has high accuracy in identifying different types of rail defects. After establishing a rail defect feature database and a rail-axle box vibration correlation, the present invention can realize rapid detection of rail defects on the entire line through axle box acceleration data, and the equipment is easy to install and operate, and can timely and automatically reflect the status of line rail defects. The present invention adopts a combination of wheel speed sensors, sleeper counting sensors and track laying maps to measure vehicle speed and integrate to obtain mileage information, and corrects it through sleeper scanning and track laying maps to achieve sleeper-level positioning accuracy of rail defect locations, and no GPS wireless signal transmission is required during the positioning process, thereby expanding the application scenario and application scope. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0040] Figure 1 Schematic diagram of the acceleration acquisition of rail defect position of the rail detection device based on the deep fusion of rail and axle box vibration data of the present invention;

[0041] Figure 2 Schematic diagram of full-line axle box acceleration acquisition of rail detection equipment based on deep fusion of rail and axle box vibration data of the present invention;

[0042] Figure 3 Schematic diagram of the time domain characteristics of the axle box acceleration signal corresponding to the rail welding joint disease of the present invention;

[0043] Figure 4 Schematic diagram of wavelet time-frequency characteristics of axle box acceleration signal corresponding to rail welding joint defects of the present invention;

[0044] Figure 5A flowchart of deep learning feature extraction and similarity analysis of axle box and rail vibration acceleration of the present invention;

[0045] Figure 6 A schematic diagram of the process of detecting and locating rail weld joint defects according to the present invention;

[0046] In the figure: 1, track; 11, rail assembly; 111, rail; 1111, rail disease; 12, sleeper; 2, train; 21, bogie; 211, axle box; 212, wheel; 2121, shaft; 3, rail disease information collection component 1; 31, vibration detection component 1; 311, acceleration sensor 1; 4, rail disease information collection component 2; 41, vibration detection component 2; 411, acceleration sensor 2; 42, train position detection component; 421, wheel speed sensor; 422, sleeper counting sensor device; 5, control system; 51, data acquisition device; 511, shielded wire; 512, wire; 52, signal display; 521, signal transmission line; 601, axle box vibration acceleration; 602, rail vibration acceleration; 603, GRU neural network; 604, feature capture; 605, feature vector; 606, similarity analysis; 607, input layer; 608, GRU layer one; 609, GRU layer two; 610, fully connected layer (FC) one; 611, fully connected layer (FC) two; 612, output layer. DETAILED DESCRIPTION

[0047] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0048] In the description of the present invention, it is necessary to understand that the terms "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship are based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0049] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0050] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0051] When there are defects in the rail weld joints, the wheel-rail impact excitation is more intense when the train passes, causing a larger amplitude of characteristic vibration of the rail. Wheel 212 is also subjected to abnormal impact excitation, and the vibration generated has specific characteristics. The train axle box vibration acceleration can reflect the state of the entire track, while the rail vibration acceleration can directly reflect the impact of rail defects. Therefore, through the deep fusion of rail and axle box vibration data, rail weld joint detection can be carried out; by monitoring the axle box vibration acceleration to indirectly identify track defects, the surface state of the rails of the entire line can be quickly detected, which can reduce the workload of rail disease detection and improve the efficiency of rail state detection.

[0052] According to an embodiment of the present invention, Figure 1 and Figure 2 As shown, the rail detection equipment based on the deep fusion of rail and axle box vibration data includes a track 1 and a train 2 running on the track 1, the track 1 includes a rail assembly 11 and a sleeper 12, the rail assembly 11 includes a plurality of rails 111 welded in sequence; it also includes: a rail disease information collection component 1 3, a rail disease information collection component 2 4 and a control system 5.

[0053] The rail defect information collection component 3 includes a vibration detection component 31, which is installed on the rail 111 with weld joint defects to collect rail vibration data corresponding to the weld joint defect position.

[0054] The rail disease information collection component 24 includes a vibration detection component 241 and a train position detection component 42; the vibration detection component 241 is installed on the axle box 211 of the train 2 to collect the axle box vibration data when the train 2 is running; the train position detection component 42 is installed on the axle box 211 of the train 2 to detect the travel distance data of the train 2.

[0055] The control system 5 is electrically connected to the vibration detection component 1 31 , the vibration detection component 2 41 and the train position detection component 42 .

[0056] The rail detection equipment of this embodiment based on the deep fusion of rail and axle box vibration data collects rail vibration data by arranging a vibration detection component 31 on the rail 111 with weld joint defects. When the rail has weld joint defects, the wheel-rail interaction when the train passes is different from that of ordinary rails, and the wheel-rail impact excitation is more intense, causing greater vibration of the rail. In addition, the rail vibration characteristics caused by different rail defects are different, and there are obvious differences in the spectrum characteristics, which are relatively easy to distinguish; by measuring the rail vibration acceleration in the section with rail defects, track defects can be directly identified. By arranging a vibration detection component 241 on the axle box 211 of the train 2 to collect axle box vibration data, when the train passes through a defective rail, the wheel 212 is subjected to abnormal impact excitation, the vibration magnitude generated is larger, and different defects will cause different vibration response characteristics, so that the track defects can be indirectly identified by monitoring the axle box vibration acceleration through the vibration detection component 241. Through the axle box vibration acceleration detection, the rapid detection of the entire track line can be achieved, and the state of the line rail welding joints can be fully reflected; the above data are aligned on the time axis by the synchronous marking method. When the train equipped with the vibration detection component 241 travels to the position interval of the vibration detection component 131 installed on the rail, the axle box vibration acceleration and the rail vibration acceleration data are synchronously marked, and the acceleration data are aligned on the time axis with the marked point as the starting point, and then the correlation between the rail and the axle box vibration acceleration corresponding to the defect position of the rail welding joint can be established, and then the defect state of the rail can be detected by measuring the train axle box vibration acceleration, which greatly reduces the workload of rail defect detection and improves the efficiency of rail state detection.

[0057] Specifically, train 2 may be a test train.

[0058] In some embodiments, the control system 5 includes a data collector 51, an industrial computer and a remote workstation; the data collector 51 is electrically connected to the vibration detection component 1 31, the vibration detection component 2 41 and the train position detection component 42 to receive and store rail vibration data, axle box vibration data and travel distance data; the industrial computer is electrically connected to the data collector 51; and the remote workstation is wirelessly connected to the industrial computer.

[0059] In some embodiments, the vibration detection component 31 includes multiple acceleration sensors 311; acceleration sensor 1 311 is installed on the bottom of the defective rail 111; the vibration detection component 41 includes acceleration sensor 2 411; acceleration sensor 2 411 is installed on the axle box 211; acceleration sensor 1 311 and acceleration sensor 2 411 are both electrically connected to the data collector 51.

[0060] In some embodiments, an acceleration sensor 1 311 is installed at the corresponding rail fastener position on the rail bottom of the defective rail 111 and at the middle position between two adjacent rail fasteners; there are multiple acceleration sensors 2 411, and an acceleration sensor 2 411 is installed on each axle box 211 of the train 2.

[0061] In some embodiments, an insulator is placed between the acceleration sensor 311 and the defective rail 111 to prevent the charged rail 111 from interfering with the acceleration sensor 311 and affecting its measurement accuracy.

[0062] Specifically, Figure 1 As shown in the figure, when installing the signal acquisition system on the rails with rail welding joint defects in the line, it is necessary to first select a suitable acceleration sensor-311 according to the vibration acceleration and vibration frequency range of the rail to be measured, and the sensitivity should reach 1.000mv / (m·s -2 ), the maximum permissible acceleration should be 5000m·s -2 , the measurement frequency range is 1 to 10000 Hz. Both the acceleration sensor 1 311 and the acceleration sensor 2 411 can be selected as a unidirectional piezoelectric acceleration sensor. Select a rail with specific defects on the line, and paste the unidirectional piezoelectric acceleration sensor on the outer side of the rail bottom through AB glue. The acceleration sensor 1 311 is installed in the direction perpendicular to the track plane to detect the vertical vibration acceleration signal of the rail; the acceleration sensor 1 311 and the rail are insulated by laying a glass sheet, and the acceleration sensor 1 311 is connected to the charge amplifier or dynamic strain gauge through a coaxial cable for signal filtering and amplification. A unidirectional piezoelectric acceleration sensor is installed at the corresponding position of the rail fastener and the mid-span position of the fastener to obtain the vibration acceleration signal of the rail 111 at different support positions. The acceleration sensor 1 311 is connected to the data collector 51 through a shielded wire 511, and then connected to the signal display 52 through a signal transmission line 521, so as to realize the real-time observation of the rail vibration acceleration signal. During the test preparation, the data collector 51 is placed in a suitable position, the shielded wire 511 is connected to the data collector 51, the joint is waterproof and insulated, the various parameters of the data collector 51 are set, and the acceleration sensor 311 is manually triggered or the test train is temporarily passed to perform the test, and the system debugging is completed; the site is organized to ensure the orderly progress of the test work. During the formal test, the data collector 51 and the signal display 52 and other collection instrument boxes need to be fixed to the pipe adjacent to the tunnel wall by iron wire, and fixed to the pipe with AB glue, iron wire, and ensure that they are fixed safely without affecting the limit requirements.

[0063] In order to construct a vibration characteristic database of rail weld joint defects, it is necessary to select multiple rail weld joint defect sections and conduct multiple tests on the rails with defects to generate sufficient rail vibration acceleration data corresponding to the rail weld joint defects.

[0064] In some embodiments, the train position detection component 42 includes a wheel speed sensor 421 and a sleeper counting sensor 422; the wheel speed sensor 421 is installed inside the axle box 211 and corresponds to the rotating shaft 2121 of the wheel 212 of the train 2 to detect the wheel speed; the sleeper counting sensor 422 is installed on the axle box 211 and corresponds to the track 1 to measure the number of sleepers 12 of the track 1 passed by the train 2 during its travel;

[0065] The data collector 51 is electrically connected to the wheel speed sensor 421 and the sleeper counting sensor 422 via a wire.

[0066] In some embodiments, the remote workstation includes a positioning module, which is used to analyze and process the data collected by the wheel speed sensor 421 and the sleeper count sensor 422 and the full-line track laying map data of the track 1 to accurately locate the defective position of the rail assembly 11 at the sleeper level.

[0067] Specifically, Figure 2 As shown, a signal acquisition system is installed on the train 2 under test. Four acceleration sensors 411 are installed one by one on the bodies of the four axle boxes 211 of the bogie 21. The bodies of the axle boxes 211 are rigid, and the specific installation positions of the acceleration sensors 411 can be flexibly adjusted as needed; the acceleration sensors 411 are installed in a direction perpendicular to the plane of the track to detect the vertical vibration acceleration signal of the axle box; the acceleration sensors 411 are electrically connected to the data collector 51 through a wire 512. To avoid missing minor track defects, acceleration sensors 411 can be installed at both ends of the same wheelset and the front and rear axle boxes on the same side. To obtain the wheel speed signal, a wheel speed sensor 421 is installed in the axle box, and the wheel speed sensor 421 can use a pulse rotation speed sensor. In order to realize sleeper counting, a sleeper counting sensor 422 is installed on the outer wall of the axle box. The sleeper counting sensor 422 can be a displacement sensor. The displacement sensor is installed in a direction perpendicular to the track 1 and its detection end corresponds to the track 1 to detect the distance displacement signal between the detection end of the sleeper counting sensor 422 on the axle box and the track, and record the change of the displacement signal when passing through the sleeper 12. The sleeper counting problem is converted into a change record of the displacement signal, and the accurate counting of the sleepers can be realized by processing the displacement signal.

[0068] Specifically, in order to realize remote acquisition control, the data collector 51 of the control system 5 can select a digital signal collector and be equipped with an A / D analog-to-digital conversion circuit, a wireless network component and a power supply device.

[0069] The acceleration sensor 1 311 and the acceleration sensor 2 411 collect the vertical vibration acceleration signal of the rail and the vertical vibration acceleration signal of the axle box respectively, the wheel speed sensor 421 collects the pulse signal generated by one rotation of the wheel, and the sleeper counting sensor 422 collects the distance signal between the axle box and the track. The acceleration sensor 1 311, the acceleration sensor 2 411, the wheel speed sensor 421, and the sleeper counting sensor 422 are all electrically connected to the A / D analog-to-digital conversion circuit through wires, the A / D analog-to-digital conversion circuit is electrically connected to the digital signal collector through wires, the digital signal collector is electrically connected to the industrial computer through wires, the industrial computer is wirelessly connected to the remote workstation through wireless network components, the power supply equipment and its supporting components supply power to the digital signal collector and the industrial computer; the power supply equipment acts as a power source to provide electrical energy to the equipment.

[0070] The power supply equipment and the industrial computer are kept in a continuous standby state. Once the data collection demand arises, the remote workstation can remotely control the industrial computer through the wireless network, and then command the digital signal collector to start data collection. Under the action of the A / D analog-to-digital conversion circuit, the vertical vibration acceleration signal of the axle box, the speed signal of the wheel shaft, and the displacement signal of the displacement sensor are all converted into digital signals and transmitted to the digital signal collector for storage. After the data collection is completed, the remote workstation sends a stop command to the industrial computer through the wireless network. The experimental data is first transmitted from the memory card of the data collector 51 to the hard disk of the industrial computer through the data cable, and then the industrial computer sends the data to the remote control workstation through the wireless network.

[0071] Another embodiment of the present invention provides a rail detection method based on deep fusion of rail and axle box vibration data, including a rail detection device based on deep fusion of rail and axle box vibration data; the detection steps include:

[0072] S1. Establish a database corresponding to rail vibration time-frequency signals and rail welding joint defects, and use GRU neural network for feature capture;

[0073] S2, collecting the axle box vibration acceleration signal of train 2 during the whole track operation, and using Alexnet convolutional neural network to screen the axle box vibration acceleration data corresponding to the rail welding joints of the whole track;

[0074] S3, input the axle box vibration acceleration data corresponding to the rail welding joint filtered in step S2 into the GRU neural network, perform feature capture, perform similarity analysis with the rail vibration characteristics in step S1, and output the corresponding rail welding joint status.

[0075] In some embodiments, S1 specifically includes the following steps:

[0076] S11. Manually investigate the track status and find the location of rail welding joint defects;

[0077] S12, installing a vibration detection component 31 on the rail 111 at the defective position of the rail welding joint corresponding to step S11, and collecting the rail vibration acceleration signal;

[0078] S13, using wavelet transform to extract the time-frequency distribution characteristics of the rail vibration acceleration signal collected in step S12, and establishing a rail vibration database under the condition of rail welding joint defects;

[0079] S14, using the rail vibration database in step S13 to train the GRU neural network to capture the rail vibration characteristics when the rail welding joint is in a defective state;

[0080] S2 specifically includes the following sub-steps:

[0081] S21, collecting the axle box vibration acceleration signal using the vibration detection component 2 41 installed on the axle box 211 of the train 2;

[0082] S22, using wavelet transform to extract the time-frequency distribution characteristics of the axle box vibration acceleration signal collected in step S21;

[0083] S23, manually intercepting and marking multiple welded joint and non-welded joint data segments from the data in steps S21 and S22, forming a sample library to perform migration training on the Alexnet convolutional neural network;

[0084] S24, using the Alexnet convolutional neural network in step S23 to screen the axle box vibration acceleration data corresponding to the rail welding joints of the entire track;

[0085] Among them, step S21 also includes: according to the synchronization mark, intercepting the axle box vibration acceleration signal corresponding to the time window of the train passing through the defective position of the rail welding joint aligned with the time axis; subsequently analyzing the axle box vibration acceleration signal in this time window to ensure the corresponding relationship between the rail vibration acceleration and the axle box vibration acceleration.

[0086] In some embodiments, in S3, the captured features generate corresponding axle box vibration acceleration feature vectors; the similarity analysis specifically calculates a similarity score by combining Euclidean distance and cosine similarity, determines the similarity between signals, and establishes a correlation between the vibration features of the rail and the axle box when there is a rail defect, so as to perform deep fusion of the vibration data of the rail and the axle box;

[0087] The currently identified axle box vibration acceleration signal is positioned in the following manner: a rough travel distance is obtained through the wheel speed sensor 421, and then the travel distance is corrected in combination with the sleeper counting sensor 422 and the full-line track laying map data of the track to obtain the precise travel distance at the sleeper level.

[0088] In some embodiments, the specific process of obtaining the rough travel distance is as follows: using the wheel speed sensor 421 to measure and record the rotational angular velocity of the wheel rolling, and multiplying the rotational angular velocity by the radius of the wheel rolling circle to obtain the vehicle speed; by integrating the vehicle speed, the train running distance at any time is obtained, and combined with the initial travel distance to obtain the travel distance information at any time;

[0089] The specific process of obtaining the precise travel distance at the sleeper level is as follows: the sleepers 12 are scanned and counted by the sleeper counting sensor 422 installed on the axle box; finally, the precise travel distance information is obtained by combining the track laying map data of the entire track line, thereby realizing the precise sleeper level positioning of the defective position of the rail welding joint.

[0090] Specifically, Figure 3 The signal time domain characteristics corresponding to the rail welding joint defects are shown. Figure 3 (b) Figure 3 (a) is a partial enlarged schematic diagram; Figure 3 In (a), the interval between each two adjacent impulse signals is L 1 =25m, the length of a single rail, Figure 3 In (b), the distance between each two adjacent impacts, one large and one small, is L. 2 =2.2m, which is the wheelbase of the bogie.

[0091] The wavelet time-frequency characteristics of rail welding joint defect signal are as follows: Figure 4 The welding joint signal has obvious two-point characteristics in the time-frequency domain, with one bright point and one dark point; that is, Figure 4 Between the horizontal coordinates 529.4 and 529.5s, there are two bright spots near the vertical coordinates 500Hz and 750Hz, one corresponding to the bright spot at the vertical coordinate 750Hz, and the other corresponding to the bright spot at the vertical coordinate 500Hz, showing one bright and one dark. The bright spot corresponding to the vertical coordinate 750Hz is brighter than the bright spot corresponding to the vertical coordinate 500Hz, and the corresponding energy is more intense.

[0092] Specifically, Figure 5 The implementation steps of similarity analysis based on deep fusion of train and track vibration data are shown, including:

[0093] Step 1: Preprocessing the vibration acceleration data of the rail with defects and the vibration acceleration data of the axle box of the entire line.

[0094] The rail and axle box vibration acceleration signal data collected by continuous synchronous marking when the train passes is divided into a series of time windows or fixed-length sequences to achieve signal segmentation. Meaningful frequency domain features (Fourier transform coefficients), time domain features (mean, variance, etc.), time-frequency features (wavelet transform coefficients), etc. are extracted from each segmented signal. In order to make the model easier to learn, the features are standardized and normalized.

[0095] Step 2: Build and train a GRU neural network model suitable for vibration acceleration signal feature learning and similarity comparison.

[0096] Design a gated recurrent unit (GRU) architecture to learn long-term dependencies and avoid overfitting. Input the preprocessed signal feature sequence into the GRU model. Set two GRU hidden layers and two fully connected layers to extract patterns and dependencies in time series data. Define the mean square error (MSE) loss function to measure the difference between the model output and the true value. Select the Adam optimization algorithm to adjust the network weights to minimize the loss function. Update the weights in the network through the backpropagation algorithm.

[0097] Step 3: Capture the features and perform similarity comparison of rail and axle box vibration data.

[0098] During the training process, the GRU model learns the characteristic representation of the signal and captures the key time series characteristics and frequency domain patterns of the signal. Using the trained GRU model, the signal is converted into a feature vector. Distance metrics such as Euclidean distance and cosine similarity are used to compare the feature vectors of different signals to evaluate their similarity. The similarity score calculated by the distance metric is used to determine the similarity between the signals.

[0099] Specifically, Figure 6 The three main processes of rail weld joint inspection are shown:

[0100] Process 1: Establish a database of time-frequency characteristics corresponding to rail welding joint defects and a database of time-frequency characteristics corresponding to axle box vibration of the entire line of trains. The specific implementation method is: investigate the status of rail welding joints, select rails with welding joint defects for vertical acceleration testing, install acceleration sensors-311 and rail vertical acceleration acquisition system such as Figure 1 As shown. Continuous wavelet transform can convert the non-stationary one-dimensional train axle box acceleration time series signal into a two-dimensional image signal of time and frequency, thereby realizing the identification of short-wave defects with the help of a high-efficiency and high-accuracy two-dimensional image recognition tool. The rail acceleration signal is processed by continuous wavelet transform to obtain the rail acceleration wavelet time-frequency signal, thereby establishing a time-frequency feature database corresponding to the rail welding joint defects. In addition, an acceleration sensor 411 is fixedly installed on the outer wall of the train axle box to collect the vertical acceleration signal of the axle box. The installation of the acceleration sensor 411 and the axle box vertical acceleration collection system are shown in the figure. Figure 2 Similarly, the axle box acceleration signal is processed by continuous wavelet transform to obtain the axle box acceleration wavelet time-frequency signal, thereby establishing a time-frequency feature database corresponding to the axle box vibration of the entire line of trains.

[0101] Process 2: Train the Alexnet convolutional neural network to screen abnormal axle box vibration acceleration, and train the GRU neural network to perform similarity analysis of rail and axle box vibration data. The time domain characteristics of the axle box acceleration signal corresponding to the rail welding joint disease are as follows: Figure 3 As shown. Among them, Figure 3 (a) shows the collected data of 80m mileage. Figure 3 (b) Figure 3 (a) is a partial enlargement. Figure 3 (b) shows the data of 9m mileage; Figure 3 In (a), the interval between each two impulse signals is L 1 =25m, the length of a single rail 111, Figure 3 In (b), the distance between the two shocks, one large and one small, is L. 2 =2.2m, which is the wheelbase of bogie 21; the wavelet time-frequency characteristics of the axle box acceleration signal corresponding to the rail welding joint disease are as follows: Figure 4 As shown. The welding joint signal is an obvious two-point feature in the time-frequency domain, and the two points appear one bright and one dark. According to the above time domain and time-frequency features, the AlexNet convolutional neural network model is used for transfer learning to process the vibration acceleration of the axle box, so as to identify the rail welding joint defects. Transfer learning is a machine learning method, which means that a pre-trained model is reused in another task. Transfer learning can be used to quickly train a neural network model that meets the requirements for a specific task under the condition of a small number of training samples. Using the pre-trained AlexNet model, modify the output parameter settings, and use the new axle box vibration data to retrain the model, so as to obtain a new model that can identify the axle box vibration signal corresponding to the rail welding joint defect. Then, a GRU deep learning neural network model is established to perform feature learning extraction and similarity analysis on the rail acceleration signal with welding joint defects and the screened axle box acceleration signal. The specific implementation method is as follows. Figure 5 As shown in the figure: the preprocessed signal feature sequence is input into the GRU model, two GRU hidden layers and two fully connected layers are set to extract patterns and dependencies in the time series data, the mean square error (MSE) loss function is defined to measure the difference between the model output and the true value, the Adam optimization algorithm is selected to adjust the network weights to minimize the loss function, and the weights in the network are updated through the back propagation algorithm to train the model. Using the trained GRU model, the signal is converted into a feature vector, and the feature vectors of different signals are compared using distance metrics such as Euclidean distance and cosine similarity to evaluate their similarity. According to the similarity score calculated by the distance metric, the similarity between the signals is judged, the deep fusion of train and track vibration data is achieved, and the correlation between the two is established.

[0102] Process three: Use the trained GRU neural network model to automatically identify the status of rail welding joints and achieve sleeper-level positioning. Establish a rail defect rapid identification module based on axle box vibration acceleration to detect rail defects in a timely and automatic manner and achieve sleeper-level precision positioning. The specific implementation method is: obtain the axle box vibration acceleration signal data of the entire line, and screen the abnormal axle box vibration acceleration through the AlexNet model. Use the GRU model to capture the features of the abnormal axle box vibration acceleration data, and combine the correlation between the rail vibration and the axle box vibration to determine the type of rail disease corresponding to the abnormal axle box vibration, so as to achieve efficient and timely identification of rail diseases. After identifying the rail disease, the sleeper-level precision positioning of the rail disease is achieved through the joint positioning method of the rotary pulse tachometer sensor installed on the wheel axle end and the displacement sensor installed on the outer wall of the axle box. Figure 2 As shown, the rotary pulse speed sensor is installed in the axle box to obtain the distance traveled by the train, and the displacement sensor is installed on the lower part of the outer wall of the axle box, and the detection end scans downward to obtain the distance from the track. The specific process is that the wheel rolls one circle to generate a pulse signal. The interval time of each pulse signal can be recorded to calculate the rotational angular velocity of the wheel, and then multiplied by the radius of the wheel rolling circle to obtain the vehicle speed. The train running distance at any time can be obtained by integrating the vehicle speed. Then, the travel distance (mileage) information at any time is obtained in combination with the initial travel distance (mileage), and then corrected by the sleeper data collected by the displacement sensor. The specific correction process is that the displacement sensor scans the sleepers, records the number of sleepers passed, and then corresponds to the number of sleepers in the track laying map of the entire line, so as to determine the real and accurate mileage information at any time and achieve sleeper-level positioning accuracy.

[0103] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification.

[0104] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A rail detection device based on deep fusion of rail and axle box vibration data, comprising a rail (1) and a train (2) running on the rail (1), the rail (1) comprising a rail assembly (11) and a sleeper (12), the rail assembly (11) comprising a plurality of rails (111) welded in sequence; characterized in that: Also includes: A rail defect information collection component (3), the rail defect information collection component (3) comprising a vibration detection component (31), the vibration detection component (31) being installed on a rail (111) having a weld joint defect to collect rail vibration data corresponding to the weld joint defect position; A rail defect information collection component 2 (4), the rail defect information collection component 2 (4) comprising a vibration detection component 2 (41) and a train position detection component (42); the vibration detection component 2 (41) is installed on the axle box (211) of the train (2) to collect and obtain axle box vibration data when the train (2) is running; the train position detection component (42) is installed on the axle box (211) of the train (2) to detect travel distance data of the train (2); A control system (5), wherein the control system (5) is electrically connected to the vibration detection component 1 (31), the vibration detection component 2 (41) and the train position detection component (42).

2. The rail detection device based on deep fusion of rail and axle box vibration data according to claim 1 is characterized in that: The control system (5) comprises a data collector (51), an industrial computer and a remote workstation; the data collector (51) is electrically connected to the vibration detection component 1 (31), the vibration detection component 2 (41) and the train position detection component (42) to receive and store the rail vibration data, the axle box vibration data and the travel distance data; the industrial computer is electrically connected to the data collector (51); and the remote workstation is wirelessly connected to the industrial computer.

3. The rail detection device based on deep fusion of rail and axle box vibration data according to claim 2 is characterized in that: The vibration detection component 1 (31) includes a plurality of acceleration sensors 1 (311); the acceleration sensors 1 (311) are installed on the rail bottom of the defective rail (111); the vibration detection component 2 (41) includes acceleration sensor 2 (411); the acceleration sensor 2 (411) is installed on the axle box (211); the acceleration sensor 1 (311) and the acceleration sensor 2 (411) are both electrically connected to the data collector (51).

4. The rail detection device based on deep fusion of rail and axle box vibration data according to claim 3 is characterized in that: The acceleration sensor 1 (311) is installed at the position of the rail fastener corresponding to the defective rail (111) and at the position between two adjacent rail fasteners; there are a plurality of acceleration sensors 2 (411), and each axle box (211) of the train (2) is installed with the acceleration sensor 2 (411).

5. The rail detection device based on deep fusion of rail and axle box vibration data according to claim 2 is characterized in that: The train position detection component (42) comprises a wheel speed sensor (421) and a sleeper counting sensor (422); the wheel speed sensor (421) is installed inside the axle box (211) and corresponds to the wheel shaft (2121) of the train (2) to detect the wheel speed; the sleeper counting sensor (422) is installed on the axle box (211) and corresponds to the track (1) to measure the number of sleepers (12) on the track (1) passed by the train (2) during its travel; The data collector (51) is electrically connected to the wheel speed sensor (421) and the sleeper counting sensor (422).

6. The rail detection device based on deep fusion of rail and axle box vibration data according to claim 5 is characterized in that: The remote workstation comprises a positioning module, which is used to analyze and process the data collected by the wheel speed sensor (421) and the sleeper counting sensor (422) and the full-line track laying map data of the track (1) to accurately locate the defective position of the rail assembly (11) at the sleeper level.

7. A rail detection method based on deep fusion of rail and axle box vibration data, characterized in that: The invention comprises a rail detection device based on deep fusion of rail and axle box vibration data as claimed in any one of claims 1 to 6; the detection step comprises: S1. Establish a database corresponding to rail vibration time-frequency signals and rail welding joint defects, and use GRU neural network for feature capture; S2, collecting axle box vibration acceleration signals of train (2) during the whole track operation, and using Alexnet convolutional neural network to screen axle box vibration acceleration data corresponding to the rail welding joints of the whole track; S3, input the axle box vibration acceleration data corresponding to the rail welding joint filtered in step S2 into the GRU neural network, perform feature capture, perform similarity analysis with the rail vibration characteristics in step S1, and output the corresponding rail welding joint status.

8. The rail detection method based on deep fusion of rail and axle box vibration data according to claim 7 is characterized in that: S1 specifically includes the following steps: S11. Manually investigate the track status and find the location of rail welding joint defects; S12, installing a vibration detection component 1 (31) on the rail (111) at the position of the rail weld joint defect corresponding to step S11, and collecting the rail vibration acceleration signal; S13, using wavelet transform to extract the time-frequency distribution characteristics of the rail vibration acceleration signal collected in step S12, and establishing a rail vibration database under the condition of rail welding joint defects; S14, using the rail vibration database in step S13 to train the GRU neural network to capture the rail vibration characteristics when the rail welding joint is in a defective state; S2 specifically includes the following sub-steps: S21, collecting axle box vibration acceleration signals using a second vibration detection component (41) installed on an axle box (211) of a train (2); S22, using wavelet transform to extract the time-frequency distribution characteristics of the axle box vibration acceleration signal collected in step S21; S23, manually intercepting and marking multiple welded joint and non-welded joint data segments from the data in steps S21 and S22, forming a sample library to perform migration training on the Alexnet convolutional neural network; S24, using the Alexnet convolutional neural network in step S23 to screen the axle box vibration acceleration data corresponding to the rail welding joints of the entire track; Among them, step S21 also includes: according to the synchronization mark, intercepting the axle box vibration acceleration signal corresponding to the time window of the train passing through the defective position of the rail welding joint aligned with the time axis; subsequently analyzing the axle box vibration acceleration signal in this time window to ensure the corresponding relationship between the rail vibration acceleration and the axle box vibration acceleration.

9. The rail detection method based on deep fusion of rail and axle box vibration data according to claim 7, characterized in that: In S3, the captured features generate the corresponding axle box vibration acceleration feature vector; the similarity analysis specifically calculates the similarity score by combining the Euclidean distance and the cosine similarity, determines the similarity between the signals, and establishes the correlation between the rail and axle box vibration features when the rail disease exists, so as to perform deep fusion of the rail and axle box vibration data; The currently identified axle box vibration acceleration signal is positioned in the following manner: a rough travel distance is obtained through a wheel speed sensor (421), and then the travel distance is corrected in combination with a sleeper counting sensor (422) and the full-line track laying map data of the track to obtain an accurate travel distance at the sleeper level.

10. The rail detection method based on deep fusion of rail and axle box vibration data according to claim 9, characterized in that: The specific process of obtaining the rough travel distance is as follows: using a wheel speed sensor (421) to measure and record the rotational angular velocity of the wheel rolling, and multiplying the rotational angular velocity by the radius of the wheel rolling circle to obtain the vehicle speed; integrating the vehicle speed to obtain the train running distance at any time, and combining the initial travel distance to obtain the travel distance information at any time; The specific process of obtaining the precise travel distance at the sleeper level is as follows: the sleepers (12) are scanned and counted by the sleeper counting sensor (422) installed on the axle box; finally, the precise travel distance information is obtained by combining the track laying map data of the entire track line, so as to realize the precise positioning of the defect position of the rail welding joint at the sleeper level.

Images