Wheel-rail roughness separation method and application of wheel-rail roughness separation method in quantitative detection of steel rail corrugation

Abstract

The invention discloses a wheel track roughness separation method. The method comprises the following steps: resampling an original time domain ABA signal and interpolating the original time domain ABA signal into an angular domain signal with equal angle intervals; according to a wheel rotation total period and a synchronous average window, an angular domain signal of the ABA signal is segmented into N segments by using a synchronous average thought; carrying out downsampling on the segmented segments by using a measurement matrix based on an independent identically distributed Gaussian matrix; adding data points corresponding to each random down-sampling signal and calculating an average value to obtain a sequence of sparse angular domain synchronous average SADSA; reconstructing a sequence based on an orthogonal matching pursuit (OMP) algorithm in a compressed sensing basic theory (CS) to obtain a periodic component extracted from the ABA signal; the invention also discloses an application of the wheel-rail roughness separation method in quantitative detection of rail corrugation, and the invention proposes that ABA signals are subjected to random downsampling and accurate reconstruction by using a compressed sensing principle under the condition of not changing original data attributes.

Description

Technical Field

[0001] The invention relates to the technical field of rail roughness corrugation, in particular to a wheel-rail roughness separation method and its application in the quantitative detection of rail corrugation. Background Art

[0002] Compared with the existing wheel-rail roughness detection methods, monitoring wheel-rail roughness based on axlebox acceleration (ABA) collected from daily operating trains is a more economical and effective method, which can guide the operation department to formulate wheel-rail operation and maintenance strategies more efficiently. However, typical wheel-rail roughness (such as wheel polygon and rail corrugation) often occurs simultaneously on the same transportation line, which has a significant coupling effect on the vehicle dynamic response, especially the ABA signal, which poses a huge challenge to the accurate monitoring of wheel-rail roughness based on ABA.

[0003] Due to the periodic rolling contact between wheels and rails, the wheel-rail system will inevitably be damaged during long-term service, resulting in various forms of wheel-rail roughness, i.e. damage. The most typical and most frequently reported cases are wheel polygons and rail corrugation. Wheel polygons are uneven wear along the circumference of the wheel, and the wheel radius has periodic changes along the circumference of the wheel. Rail corrugation is uneven wear distributed longitudinally along the surface of the rail head. Both wheel polygons and rail corrugation are common in subway, heavy-duty, high-speed and light rail systems, and have brought many negative effects, such as increasing wheel-rail interaction forces, accelerating the failure of vehicle and track components, and intensifying abnormal vibration and noise, which seriously affect the running quality of railway vehicles. Therefore, it is very necessary to develop a fast and accurate wheel-rail roughness monitoring system, which can accurately guide people to carry out wheel-rail operation and maintenance operations in the most needed places and at the most appropriate time, so that railway operators can manage the health of wheels and rails more scientifically.

[0004] Currently, wheel-rail roughness is usually diagnosed by manual visual inspection, special precision instruments or comprehensive inspection trains. It should be noted that with the rapid development of sensor and communication technology in the past two decades, installing low-cost sensors such as accelerometers in daily operating vehicles to monitor wheel-rail status has gradually become a mainstream development trend. This on-board monitoring system often does not require manual experience, is fully data-driven and automatically operated, and can avoid the high testing costs caused by the use of complex equipment. This type of monitoring system can also continuously monitor the initiation and evolution of wheel-rail damage in the entire railway network, so as to formulate more scientific operation and maintenance strategies, which is of great benefit to railway operators in achieving sustainable development of intelligent operation and maintenance of wheel-rail systems.

[0005] As we all know, wheel polygon and rail corrugation usually cause high-frequency dynamic response of the vehicle, and the ABA signal without attenuation of the primary and secondary suspension systems of the vehicle is very suitable for objectively reflecting the wheel-rail roughness characteristics corresponding to these high-frequency characteristics. Therefore, among all on-board monitoring systems, the research on wheel polygon and rail corrugation detection based on ABA has attracted more and more attention and has been widely reported.

[0006] Wheel polygon detection based on ABA: Sun et al. used angular domain synchronous averaging technology to process ABA signals and developed a wheel polygon detection system based on this. They pointed out that based on ABA data, the order and roughness characteristics of wheel polygons can be accurately evaluated. Xie et al. extracted the quasi-periodic components corresponding to wheel polygons in ABA signals by constructing steady-state test and similarity test methods. The results of long-term tracking tests on a high-speed train showed that this ABA-based method can quantitatively identify the main characteristics of single-order and multi-order wheel polygons of trains at different mileages. For low-order wheel polygons, Zhang et al. reported a method for signal decoupling of different faults using the principle of signal superposition. This solves the problem that the low-frequency characteristics of ABA signals are easily interfered with and improves the accuracy of online diagnosis of wheel polygons based on ABA. For high-order wheel polygons, Wei et al. combined time-frequency analysis and inertial reference method to study the high-frequency vibration characteristics of ABA and completed the corresponding polygon recognition task. For wheel polygons in the early stages of initiation, Xu et al. proposed an anti-interference wheel polygon detection model based on ABA. Simulation and field tests show that their signal processing method for ABA is effective. In addition to the above-mentioned traditional ABA signal processing, Ye et al. also built a deep learning model for the condition monitoring and fault diagnosis of high-speed train wheel out-of-roundness (including wheel polygons). Dong et al. established a hybrid deep learning algorithm and conducted research on the ABA-based heavy-duty locomotive wheel polygon detection method.

[0007] Rail corrugation detection based on ABA: Kojima et al. reported the first study on the detection of rail corrugation based on vehicle-mounted measurement data including ABA. Bocciolone et al., Molodova et al., and Karis et al. studied the correlation between ABA and rail corrugation from theoretical and experimental perspectives, respectively. Their research further developed and revealed the mapping relationship between ABA and rail corrugation. Based on the above results and experience, the use of vehicle-mounted ABA data to detect rail roughness is currently very popular worldwide. For example, Li et al. introduced the relevant results of the Dutch railway network based on ABA for automatic detection of rail corrugation. Salvador et al. conducted experiments in the Valencia (Spain) metropolitan railway network to determine the optimal sampling and filtering frequencies and the installation positions of accelerometers for ABA-based rail corrugation monitoring. Tanaka et al. studied the effective accuracy and application conditions of ABA-based rail corrugation condition monitoring and management on Japanese railways. Carrigan et al. proposed a frequency domain random inverse method based on a vehicle-track dynamic simulation model and subway measurement data between Wolverhampton and Birmingham in the West Midlands, UK, to derive the wavelength and roughness information of rail corrugation from the ABA data of field tests.

[0008] In summary, wheel polygon and rail corrugation detection based on ABA is an effective method for wheel-rail roughness diagnosis and a beneficial supplement to existing monitoring methods. However, there is an important and often overlooked problem with this type of method: wheel polygon and rail corrugation usually exist simultaneously on the same transportation line, and the signals excited by the two in ABA have a significant coupling effect. If the test vehicle has wheel polygons (rail corrugation), this poses a huge challenge to the accurate detection of rail corrugation (wheel polygon) based on ABA.

[0009] Pallas et al. reviewed several methods for separating wheel and rail roughness. They believed that the synchronous averaging (SA) method can obtain the best separation results by utilizing the periodic and random contribution characteristics of wheel and rail roughness to ABA response respectively. Subsequently, Carrigan et al. and Pieringer et al. confirmed the feasibility of synchronous averaging method in monitoring rail corrugation in wheel polygonal condition. Summary of the invention

[0010] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a wheel-rail roughness separation method and its application in the quantitative detection of rail corrugation. The present invention proposes to use the principle of compressed sensing to randomly downsample the ABA signal and accurately reconstruct it without changing the original data attributes. By amplifying the random characteristics of the random components in the ABA signal, the expression of the periodic components in the ABA can be enhanced, thereby separating different signal components corresponding to the wheel-rail roughness.

[0011] To achieve the above object, the technical solution adopted by the present invention is: a wheel-rail roughness separation method, comprising the following steps:

[0012] Step 1, resample the original time domain axle box acceleration signal with equal time increment, i.e., the ABA signal y(t) and interpolate it into an angular domain signal y(θ) with equal angle intervals;

[0013] Step 2: According to the total wheel rotation period M and the synchronous averaging window L, the angular domain signal y(θ) of the ABA signal is divided into N segments y1(θ), y2(θ), ..., y N (θ), N = floor(M / L), floor refers to the floor operator;

[0014] Step 3: Downsample the segmented fragments using a measurement matrix based on an independent and identically distributed Gaussian matrix to obtain a sequence

[0015] Step 4: Add the data points corresponding to each randomly downsampled signal and calculate the average value to obtain the sequence based on the wheel-rail roughness separation model, namely, the sparse angular domain synchronous average SADSA

[0016] Step 5: Reconstruct the sequence based on the orthogonal matching pursuit algorithm OMP in the basic theory of compressed sensing CS Get the periodic component extracted from the ABA signal

[0017] As a further improvement of the present invention, in step 1, the angular domain signal y(θ) of the ABA signal resampled to the angular domain is specifically as follows:

[0018] y(θ)=S(θ)+A(θ)+R(θ)

[0019] Among them, S(θ), A(θ) and R(θ) represent synchronous coherent signals, i.e. periodic signals, asynchronous signals, i.e. non-periodic signals and random noise respectively.

[0020] As a further improvement of the present invention, in step 2, the divided angular domain signal y(θ i ) are as follows:

[0021]

[0022] Wherein, C is the attenuation coefficient of the asynchronous signal A(θ), C<<1.

[0023] As a further improvement of the present invention, step 3 is specifically as follows:

[0024] Measurement data through the measurement matrix Φ Thus, the original or complete data y can be reconstructed; the basic theory of compressed sensing CS is as follows:

[0025]

[0026] in, is the original angular domain ABA signal, is the downsampled measurement signal, is the measurement matrix; They are the sparse transform basis and the sensing matrix respectively; Represents the sparse coefficients of the ABA signal in the frequency domain;

[0027] According to the CS principle, the transformed signal ω obtained by applying the transform basis sparse Ψ to y has sparseness, and the original signal y is accurately reconstructed by solving the following optimization equation:

[0028]

[0029] Where ||·||1 represents the l1 norm;

[0030] Orthogonal matching pursuit (OMP) is used to iteratively search for the support set of the sparse vector and reconstruct the original signal y using least squares estimation; the input of OMP is the downsampled measurement signal Sensing matrix Θ and sparse coefficient K; is the output of OMP, that is, according to the measurement signal Reconstructed signal:

[0031]

[0032] As a further improvement of the present invention, in step 4, the sequence of sparse angular domain synchronous average SADSA as follows:

[0033]

[0034] As a further improvement of the present invention, in step 5, the periodic component extracted from the ABA signal is as follows:

[0035]

[0036] The present invention also provides an application of a wheel-rail roughness separation method in the quantitative detection of rail corrugation, wherein the periodic component extracted by the wheel-rail roughness separation method as described above is Quantitative detection of wheel-rail corrugation information.

[0037] As a further improvement of the present invention, according to the periodic component The inertial reference method or RCNet is used to quantitatively detect the corrugation information of the rail.

[0038] The beneficial effects of the present invention are:

[0039] This paper proposes a method for separating wheel and rail roughness from ABA signals, and introduces an application case of quantitative detection of rail corrugation in the presence of wheel polygons. First, the data characteristics of the ABA signal itself are studied, and a wheel-rail roughness separation model, sparse angle domain synchronous averaging (SADSA), is constructed based on the theory of random downsampling in synchronous averaging and compressed sensing. Subsequently, field tests are carried out to prove the effectiveness of the proposed SADSA under different operating conditions and its superiority in quantitative detection of rail corrugation roughness under wheel polygon interference. The results show that SADSA can effectively attenuate asynchronous signals and random noise in ABA, thereby accurately separating synchronous signals caused by periodic wheel roughness excitation. Comparative analysis and ablation tests show that compared with ABA processed by the most advanced algorithm as the input of the rail corrugation detection model, the ABA input based on SADSA can improve the fitting degree of rail corrugation detection from 62.6% to 90.6%, and the error level can be reduced from 29.2 (dBre1μm) to 24.8 (dBre1μm). 。 BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a framework diagram for quantitatively identifying rail corrugation roughness based on ABA when wheel polygons exist in an embodiment of the present invention;

[0041] Figure 2 Schematic diagram of angular domain synchronous averaging in an embodiment of the present invention;

[0042] Figure 3 Schematic diagram of compressed sensing in an embodiment of the present invention;

[0043] Figure 4 This is a flow chart of a sparse angular domain synchronous averaging method according to an embodiment of the present invention;

[0044] Figure 5 1 is a field test diagram in an embodiment of the present invention; wherein (a) is a rotation speed test, (b) is an axle box acceleration test, (c) is a wheel polygon test, and (d) is a rail corrugation test;

[0045] Figure 6 Schematic diagram of field measured data in an embodiment of the present invention; wherein (a) is the rotation speed and the derived vehicle speed signal, (b) is the ABA signal, (c) is the wheel polygon signal, and (d) is the rail corrugation signal;

[0046] Figure 7Schematic diagram of wheel-rail roughness separation results in an embodiment of the present invention; wherein (a) is the result based on ADSA, and (b) is the result based on SADSA;

[0047] Figure 8 It is a schematic diagram of wheel-rail roughness separation results under variable speed and uniform speed conditions in an embodiment of the present invention;

[0048] Fig. 9 Schematic diagram of rail corrugation roughness detection results under different ABA inputs in an embodiment of the present invention; wherein (ac) represents variable speed and (df) represents uniform speed. DETAILED DESCRIPTION

[0049] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0050] Example

[0051] A wheel-rail roughness separation method, comprising:

[0052] 1. Rail corrugation detection framework based on ABA under wheel polygon interference:

[0053] Aiming at wheel-rail roughness separation, a case study of rail corrugation quantitative detection based on ABA under wheel polygon interference is proposed. Figure 1 As shown in the figure, it mainly includes two modules. The first is about the signal preprocessing of ABA, which mainly explains the sparse angle domain synchronous averaging (SADSA) model based on synchronous averaging and compressed sensing. Then, a rail corrugation detection model is introduced for quantitative regression prediction of rail corrugation roughness.

[0054] 1.1. Sparse angular domain synchronous averaging:

[0055] 1.1.1 Basic theory of angular domain synchronous averaging:

[0056] The ABA signal measured on site usually contains various amplitude, frequency and phase modulations, and is affected by random noise. This embodiment uses the idea of ​​synchronous averaging to separate the wheel-rail roughness component in ABA, that is, to extract the relevant features of different types (periodic and non-periodic) wheel-rail roughness. The ABA response of a railway vehicle under periodic excitation (such as a wheel polygon) is periodically distributed in the angular domain along the circumference of the wheel. When the wheel speed is constant, it also exhibits periodic characteristics in the time domain. However, vehicles are often accompanied by acceleration and braking operations during operation, making this assumption difficult to achieve. During synchronous averaging, even small speed fluctuations (about 0.1%) will result in a large amount of information loss. In order to solve the problem of speed change, the measured ABA signal can be converted from equal time intervals to equal angle intervals, and angle domain synchronous averaging (ADSA) can be performed, that is, the ABA sequence is resampled and the ABA signal is interpolated at a fixed angle interval θ. A typical ADSA such as Figure 2 As shown. The reference signal during synchronous averaging is phase-locked with the wheel rotation angle position and is used to process the time domain ABA signal. The reference signal can be obtained from a dedicated tachometer installed on the vehicle. The ABA signal y(θ) resampled to the angular domain consists of three parts:

[0057] y(θ)=S(θ)+A(θ)+R(θ) (1)

[0058] Among them, S(θ), A(θ) and R(θ) represent synchronous coherent (periodic) signals, asynchronous (non-periodic) signals and random noise, respectively. Assuming that the ABA signal y(θ) contains a total of M wheel rotation periods, and a synchronous averaging window is L wheel rotation periods during ADSA processing, the signal y(θ) can be divided into N segments. In order to balance computational efficiency and spatial resolution, L is set to 10. N = floor(M / L) is the synchronous average sample size during ADSA calculation, and floor refers to the floor operator. The corresponding data points of each segment are then added to calculate the average value. Due to the phase shift of the asynchronous signal A(θ) and the irrelevance of the random noise R(θ), the two signal components can be greatly attenuated.

[0059]

[0060] azasxADSA's output signal x(θ i ) can be obtained by i ) is calculated on average:

[0061]

[0062] After ADSA calculation, the asynchronous signal A(θ) is attenuated to C times of the original sequence, and the output noise is times. Note that the information of the synchronous coherent signal S(θ) has been accurately extracted. Combining the above formulas, the discrete signal after ADSA processing can be expressed as:

[0063]

[0064] It can be seen that ADSA can use the wheel angle domain position as a reference when averaging y(θ), amplify the synchronous signal S(θ), and weaken the asynchronous signal A(θ) and random noise R(θ). In this way, the quasi-periodic component corresponding to the fixed wheel roughness and the random separation of the time-varying rail corrugation roughness in ABA can be separated.

[0065] 1.1.2 Basic theory of compressed sensing:

[0066] like Figure 3 As shown in Figure 2, the essence of compressed sensing (CS) is to use the redundant characteristics of information to measure a small amount (or incomplete) of data through the measurement matrix Φ. Thus, the original (or complete) data y can be accurately reconstructed

[32] . Formula (5) shows the basic theory of CS:

[0067]

[0068] in, is the original angular domain ABA signal, is the downsampled measurement signal, is the measurement matrix. CS has two important prerequisites, namely sparsity and incoherence. The former means that the original signal y must be sparse in a certain expression domain, that is, the number of non-zero elements in the signal is much smaller than the total number of signals. Although the ABA signal is not sparse in the time domain and the angular domain, it can meet the approximate sparse characteristics required by CS after being processed with a sparse basis Ψ (such as discrete cosine transform (DCT) and fast Fourier transform (FFT)) in the frequency domain. Incoherence requires that the measurement matrix Φ must satisfy the restricted isometry property (RIP). Its equivalent condition is that the measurement matrix Φ is incoherent with the sparse basis Ψ. Random Gaussian matrices with independent and identical distributions can be used as general CS measurement matrices Φ. Formula (5) They are respectively a sparse transform basis (DCT is used in this embodiment) and a sensing matrix. represents the sparse coefficients of the ABA signal in the frequency domain. According to the CS principle, the transformed signal ω obtained by applying the transform basis Ψ to y is usually sparse enough and the original signal y can be reconstructed as accurately as possible by solving the following optimization equation

[0069]

[0070] Where ||·||1 represents the l1 norm. The above formula can be solved directly using greedy iteration or convex optimization algorithm. The most classic method is orthogonal matching pursuit (OMP), which is used to iteratively search the support set of sparse vectors and use least squares estimation to reconstruct the original signal y. The input of OMP is the downsampled measurement signal Sensing matrix Θ and sparse coefficient K. Since we usually focus on wheel polygons within order 40, K is set to 40. is the output of OMP, that is, according to the measurement signal Reconstructed signal.

[0071]

[0072] 1.1.3 Synchronous averaging under the compressed sensing framework:

[0073] ADSA is used to attenuate asynchronous signals and random noise in ABA to extract periodic signals in ABA caused by wheel roughness. CS is introduced to take advantage of its ability to randomly downsample and accurately reconstruct signals. Before ADSA processing, the original angular domain ABA signal y is downsampled to y using a random Gaussian measurement matrix Φ. Then the downsampled measurement signal Enter the ADSA calculation to obtain the average signal Finally, based on SA is reconstructed using the OMP algorithm The random downsampling process can further attenuate the asynchronous signal and random noise in the ABA. On the other hand, the periodic component with important physical significance in the ABA signal is retained and accurately reconstructed due to its strong sparsity in the frequency domain. In other words, the reconstructed signal It has the same periodic component as the original signal y, but the asynchronous and random noise signals have been greatly weakened. Therefore, based on the reconstructed signal As input ADSA gets It can better avoid the interference of asynchronous signals and random noise, so as to obtain more objective and real signal characteristics in the frequency domain and a smoother waveform in the angular domain.

[0074]

[0075] like Figure 4 As shown, the basic steps of the synchronous averaging model based on the compressed sensing framework, sparse angle domain synchronous averaging (SADSA), can be described as follows.

[0076] (1) The original time domain (equal time increment) ABA signal y(t) is resampled and interpolated into an angular domain signal y(θ) with equal angle intervals.

[0077] (2) According to the total wheel rotation period M and the synchronous averaging window L, the ABA signal y(θ) is divided into N segments y1(θ), y2(θ), ..., y N (θ), N = floor (M / L).

[0078] (3) Use the measurement matrix based on independent and identically distributed Gaussian matrix to downsample the segmented fragments and obtain the sequence

[0079] (4) Add the data points corresponding to each randomly downsampled signal and calculate the average value to obtain the SADSA-based sequence

[0080] (5) Reconstructing the sequence based on the OMP algorithm in CS Get the periodic component extracted from ABA

[0081] 1.2. Rail corrugation and roughness detection model:

[0082] The proposed SADSA can extract the periodic components in ABA caused by wheel roughness, and then the inertial reference method or RCNet can directly evaluate the severity of wheel roughness by quantitatively detecting the corrugation information of the rail. For example, wheel polygons and wheel flats were successfully identified using periodic signals based on ADSA. This embodiment aims to use the proposed SADSA to study rail corrugation detection in the presence of wheel polygons, that is, to use SADSA to separate the different ABA components excited by wheel and rail roughness, and then input the ABA signal corresponding to the rail roughness into the corresponding detection model to automatically output rail corrugation information.

[0083] 2. Engineering application:

[0084] 2.1. Field test:

[0085] Field tests were conducted on an actual operating subway line in a city in my country. Figure 5 As shown in Figure 1, the test contents include wheel speed, ABA, wheel roughness and rail roughness. The collected data are used to verify the effectiveness of the proposed wheel-rail roughness separation algorithm (SADSA). It should be noted that the wheels of the test train are generally required to be in the initial turning state to ensure that the measured ABA is mainly affected by the rail roughness. However, this condition is not always met. On some subway lines, there are both wheel polygons and rail corrugations. Figure 6 Figure 2 shows the measured signal in this case. Figure 6 (a) in the figure is the speed pulse signal measured by the photoelectric sensor and the vehicle speed information derived therefrom. Figure 6 (b) in the figure is the ABA signal collected by installing it on the axle box. Figure 6 (c) in FIG. 5 shows the wheel roughness data obtained using the BST device. Figure 6 (d) is the rail corrugation signal measured by CAT equipment. It can be seen that the test wheel has obvious 15-order polygonal wear, and there is obvious rail corrugation on the curve with a radius of 360m.

[0086] 2.2. Wheel-rail roughness separation effect of SADSA on measured data:

[0087] Figure 7 The black, green, and orange lines in the figure represent the measured raw ABA data and the ABA components related to rail and wheel roughness after signal separation, namely “ABA-rail” and “ABA-wheel”. Figure 7 (a) and (b) show the overall results of wheel-rail roughness separation based on ADSA and the proposed SADSA, respectively. However, the “ABA-wheel” extracted by ADSA is slightly larger than the result based on SADSA. The reason may be that ADSA regards the periodic component excited by the rail roughness as a component related to the wheel roughness, which causes ADSA to mistakenly misjudge part of the energy belonging to the rail roughness as wheel roughness. In addition, a random Figure 7 The ABA signals of the medium speed change section and the constant speed section are analyzed in detail. Figure 8 (a) and (d) in the figure show the ABA component excited by the wheel polygon at variable speed and constant speed, namely "ABA-wheel". It can be seen that, whether it is variable speed or constant speed, the extraction result based on SADSA more intuitively represents the wheel polygon characteristics of order 15 than the extraction result based on ADSA, and the waveform is smoother, which can better describe the harmonic polygon wear characteristics of the wheel. When the order of the wheel polygon is determined by the local minimum (maximum) method, the result based on ADSA will lead to misjudgment because several erroneous peaks (valleys) will appear in the waveform. SADSA can solve this problem due to the use of sparse random downsampling method. It can filter the random interference of the signal and play the role of shifting peaks and filling valleys, so that the extracted periodic signal can more accurately reflect the wheel roughness. Figure 8(b), (c) and (d), (f) show the ABA signals corresponding to rail roughness separated under different speed conditions and separation methods. It can be seen that the energy of the "ABA-rail" signal based on SADSA is slightly higher than that of the ADSA method, because SADSA makes appropriate corrections to the "ABA-wheel" signal. Although the differences between the ABA-rail signals in different methods seem small, when the ABA-rail signals are used for rail corrugation detection tasks, completely different results may be obtained.

[0088] 2.3. Rail corrugation roughness detection under wheel polygon interference:

[0089] The proposed SADSA is superior to the rail corrugation and roughness detection task under wheel polygon interference. Figure 8 The signal described in is the research object. The original ABA signal (without wheel-rail roughness separation), ABA input based on ADSA processing and ABA input based on SADSA processing are used as the input of the rail corrugation detection model IRCNet to perform quantitative detection of rail corrugation. The corresponding results are shown in Fig. 9 As shown. It can be seen that the contribution of the vibration component caused by the wheel polygon to the ABA signal cannot be ignored, and even drowns out the ABA component excited by the rail corrugation. It can be seen that the rail corrugation prediction value after wheel-rail roughness separation is significantly better than the result based on the original ABA input. This proves that wheel-rail roughness separation processing is very important in ABA-based rail corrugation detection, especially when the wheel has obvious out-of-round characteristics. Otherwise, the detection results may be far from the actual situation, resulting in improper wheel-rail operation and maintenance strategies, and ultimately even causing serious safety hazards and economic losses. Fig. 9 Further research shows that the rail corrugation roughness detection results based on SADSA are closer to the measured values ​​than those based on ADSA, regardless of whether the speed is constant or variable. This is because the proposed SADSA can more scientifically separate the interference of wheel polygons from ABA than ADSA, thereby providing a purer input for the rail corrugation detection model IRCNet, which is very important for accurately detecting the main features of rail corrugation (such as wavelength and roughness). Table 1 quantitatively describes Fig. 9 The detection results in Figure 2. It can be seen that the ABA input based on SADSA has significant advantages in the fitting degree and error level of rail corrugation roughness detection. Compared with the current most advanced algorithm ADSA, the fitting degree based on SADSA can be improved from 62.6% to 90.6% at constant speed, and the error level can be reduced from 29.2 (dB re1μm) to 24.8 (dB re1μm).

[0090] Table 1. Comparison of rail corrugation test results under different ABA inputs

[0091]

[0092] In ABA-based wheel-rail roughness monitoring, it is very important to separate the ABA components corresponding to different wheel-rail excitation sources, especially when wheel-rail roughness coexists in the same track system. Otherwise, incorrect or even misleading results may be obtained.

[0093] A wheel-rail roughness separation model SADSA based on synchronous averaging and random downsampling theory in compressed sensing is proposed. Field measured data show that SADSA can effectively attenuate asynchronous signals and random noise in ABA and accurately separate the periodic signal corresponding to the wheel polygon in ABA.

[0094] Compared with the current most advanced algorithm ADSA, the fitting degree of rail corrugation roughness based on SADSA can be improved from 62.6% to 90.6%, and the error level can be reduced from 29.2 (dBre1μm) to 24.8 (dBre1μm).

[0095] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A wheel-rail roughness separation method, characterized in that: The following steps are involved: Step 1, resample the original time domain axle box acceleration signal with equal time increment, i.e., the ABA signal y(t) and interpolate it into an angular domain signal y(θ) with equal angle intervals; Step 2: According to the total wheel rotation period M and the synchronous averaging window L, the angular domain signal y(θ) of the ABA signal is divided into N segments y1(θ), y2(θ), ..., y N (θ), N = floor(M / L), floor refers to the floor operator; Step 3: Downsample the segmented fragments using a measurement matrix based on an independent and identically distributed Gaussian matrix to obtain a sequence Step 4: Add the data points corresponding to each randomly downsampled signal and calculate the average value to obtain the sequence based on the wheel-rail roughness separation model, namely, the sparse angular domain synchronous average SADSA Step 5: Reconstruct the sequence based on the orthogonal matching pursuit algorithm OMP in the basic theory of compressed sensing CS Get the periodic component extracted from the ABA signal 2. The wheel-rail roughness separation method according to claim 1, characterized in that: In step 1, the angular domain signal y(θ) of the ABA signal resampled to the angular domain is specifically as follows: y(θ)=S(θ)+A(θ)+R(θ) Among them, S(θ), A(θ) and R(θ) represent synchronous coherent signals, i.e. periodic signals, asynchronous signals, i.e. non-periodic signals and random noise respectively.

3. The wheel-rail roughness separation method according to claim 2, characterized in that: In step 2, the segmented angular domain signal y(θ i ) are as follows: Wherein, C is the attenuation coefficient of the asynchronous signal A(θ), C<<1.

4. The wheel-rail roughness separation method according to claim 3, characterized in that: Step 3 is as follows: Measurement data through the measurement matrix Φ Thus, the original or complete data y can be reconstructed; the basic theory of compressed sensing CS is as follows: in, is the original angular domain ABA signal, is the downsampled measurement signal, is the measurement matrix; They are the sparse transform basis and the sensing matrix respectively; Represents the sparse coefficients of the ABA signal in the frequency domain; According to the CS principle, the transformed signal ω obtained by applying the transform basis sparse Ψ to y has sparseness, and the original signal y is accurately reconstructed by solving the following optimization equation: Where ||·||1 represents the l1 norm; Orthogonal matching pursuit (OMP) is used to iteratively search for the support set of the sparse vector and reconstruct the original signal y using least squares estimation; the input of OMP is the downsampled measurement signal Sensing matrix Θ and sparse coefficient K; is the output of OMP, that is, according to the measurement signal Reconstructed signal:

5. The wheel-rail roughness separation method according to claim 4, characterized in that: In step 4, the sequence of sparse angular domain synchronous average SADSA as follows:

6. The wheel-rail roughness separation method according to claim 5, characterized in that: In step 5, the periodic component extracted from the ABA signal is as follows:

7. Application of a wheel-rail roughness separation method in the quantitative detection of rail corrugation, characterized in that: The periodic component extracted by the wheel-rail roughness separation method according to any one of claims 1 to 6 Quantitative detection of wheel-rail corrugation information.

8. The wheel-rail roughness separation method according to claim 7 and its application in the quantitative detection of rail corrugation, characterized in that: According to the periodic component The inertial reference method or RCNet is used to quantitatively detect the corrugation information of the rail.

Images