Dictionary fusion steel rail damage positioning method based on dynamic time warping

Through dynamic time regularization and dictionary fusion technology, combined with sparse expression and Mahalanobis distance algorithm, the problems of large-scale monitoring and noise interference in rail acoustic emission detection are solved, and high-precision rail damage positioning is achieved.

CN119125323AActive Publication Date: 2024-12-13HARBIN INST OF TECH

Patent Information

Application Number
CN202411122496.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-12-13
Estimated Expiration
2044-08-15

AI Technical Summary

Technical Problem

The existing rail acoustic emission detection technology is difficult to achieve large-scale monitoring, and the detection accuracy is limited by the sensor performance, which is easily masked by the rolling noise of the wheel and rail, resulting in false detection.

Method used

Using a dictionary fusion method based on dynamic time regularity, combining sparse expression and Mahalanobis distance algorithm, the on-board AE sensor detection wheel rolling distance estimation algorithm and a multi-dictionary signal fusion algorithm are developed, and a dual-feature adaptive threshold is designed to accurately locate rail damage.

Benefits of technology

It breaks through the limitations of fixed sensors, improves detection accuracy, reduces wheel and rail rolling noise interference, and realizes high-precision rail damage detection and positioning.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119125323A_ABST
    Figure CN119125323A_ABST
Patent Text Reader

Abstract

The invention discloses a dictionary fusion steel rail damage positioning method based on dynamic time warping, and the method comprises the steps: segmenting a multi-channel AE signal into peak detection frames, and calculating the time difference between adjacent frames, so as to estimate the rolling distance of a wheel. Meanwhile, the multi-channel AE signal can train a sub-dictionary and a reference dictionary in previous work. The multi-dictionary SR algorithm based on the improved DTW is innovated, a unified fusion dictionary is synthesized through sub-dictionary weighting, damage feature information is enhanced, and most WRRN is reduced. In the process, the innovative ADTW-SDM algorithm reflects the global features of the dictionary atoms, and the similarity between the atoms is accurately evaluated through the second-order difference value of the Mahalanobis distance. A double-feature adaptive threshold is designed for detecting and positioning damage on the steel rail. The method provides guidance for follow-up accurate detection and positioning of large-range steel rail damage based on the advanced AE technology.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of structural health monitoring (SHM) of rails and relates to a rail damage locating method, and specifically to a dictionary fusion method based on dynamic time warping for a mobile rail damage locating method under acoustic emission (AE) technology. Background Art

[0002] Nowadays, modern rail transportation is developing towards high speed and heavy load. Frequent wear of rails can easily lead to defects and even cause serious rail breakage accidents. Therefore, it is imperative to implement efficient and accurate SHM for railway infrastructure. SHM of rails requires not only reliable damage diagnosis, but also accurate damage location for timely maintenance and repair.

[0003] As an emerging non-destructive testing technology, acoustic emission testing has been widely used in structural health monitoring and defect location due to its non-invasiveness, sensitivity and freedom from site geometry. However, most of the current acoustic emission detection technologies for rail damage rely on deep learning models, which are difficult to implement on a large scale due to their lack of interpretability and strong dependence on system parameters. In addition, these methods are also limited by fixed installation positions, requiring a large number of sensors to cover a wide range of rail structural health monitoring.

[0004] This situation shows that it is necessary to develop an efficient rail acoustic emission detection technology that can be widely used in practical applications. At the same time, among the various existing damage location methods, the Time Difference of Arrival (TDOA) algorithm stands out for its real-time, accuracy and multi-sensor collaboration. However, TDOA can only determine the position of the damage signal relative to the tread of the detection wheel when it appears, but cannot determine its position on the rail. Therefore, calculating the rolling distance of the detection wheel is a key part of rail damage location.

[0005] In addition, the damage AE signal originates from the deformation or expansion of the rail at the potential defect location under the action of load stress. The contact between the loaded wheel and the rail will generate high-amplitude wheel rail rolling noise (WRRN), which will mask the basic damage information. Due to the overlapping frequency band distribution, it is very challenging to eliminate WRRN by filtering alone. Therefore, detecting the submerged damage signal from the noise background is another key task. Summary of the invention

[0006] In view of the problem that traditional rail damage detection relies on a fixed single sensor, cannot achieve large-scale rail SHM, and the detection accuracy is limited by the sensor performance, and is easily submerged by WRRN and causes false detection, the present invention provides a dictionary fusion rail damage positioning method based on dynamic time warping. This method combines the sparse representation (SR) method to propose a new multi-dictionary fusion method based on improved dynamic time warping (DTW), which is used for large-scale rail SHM and damage positioning based on AE technology. In this method, a detection wheel rolling distance estimation algorithm based on three on-board AE sensors is developed to monitor the moving distance of the wheel. In addition, a multi-dictionary SR algorithm based on improved DTW is also innovated to fuse multi-channel AE signals, reduce WRRN and enhance the characteristics of damage signals. In addition, an adaptive DTW algorithm (Adaptive DTWbase on the Second-Order Difference of the Mahalanobis Distance, ADTW_SDM) was developed based on the second-order difference of the Mahalanobis distance algorithm to accurately evaluate the similarity between the dictionary and the reference atoms and weight the sub-dictionary of the fusion signal. Finally, a dual-feature adaptive threshold was designed to detect damage signals and locate the location of damage on the rail, providing guidance for advanced and accurate rail structural health monitoring.

[0007] The objective of the present invention is achieved through the following technical solutions:

[0008] A dictionary fusion rail damage location method based on dynamic time warping includes the following steps:

[0009] Step 1: Divide the multi-channel AE data set collected by multiple vehicle-mounted AE sensors into detection frames and perform normalization processing on them. Then, the frames with energy greater than the threshold are recorded as peak detection frames. The maximum delay between adjacent peak detection frames is calculated, and the relative position of the sensor to the detection wheel at the corresponding time is determined. The rolling distance of the detection wheel is estimated by measuring two adjacent relative positions.

[0010] Step 2: Establish a sub-dictionary for each channel, and use the K-order singular value decomposition algorithm (K-SVD) and the alternating direction multiplier method (ADMM) to solve the sparse coefficients of the sub-dictionary and the corresponding channel respectively, and use the improved DTW algorithm to calculate the weights for fusing multi-channel sub-dictionaries and joint sparse coefficients. Finally, reconstruct the fused signal by fusing the dictionary and the joint sparse coefficients.

[0011] Step 3: Introduce the DTW algorithm, use the second-order difference of the Mahalanobis distance to replace the Euclidean distance, add a penalty function when calculating the DTW path, calculate the similarity between the multi-channel sub-dictionary atoms and the base dictionary atoms, and provide the best dictionary fusion weight;

[0012] Step 4: Combine the short-time zero-crossing rate to identify the damage information by detecting the frequency of the signal crossing the zero axis, and construct a dual-feature adaptive threshold to further make up for the defects of the RPG feature.

[0013] The present invention novelly combines vehicle-mounted sensor data fusion with signal processing algorithms to locate rail damage by mobile detection. Compared with the prior art, it has the following advantages:

[0014] 1) The detection wheel rolling distance estimation algorithm uses the on-board AE sensor to accurately track the rolling distance under speed changes, breaking through the limitations of fixed sensors and minimizing system complexity.

[0015] 2) The multi-dictionary SR algorithm based on the improved DTW can effectively fuse multi-channel data, enhance damage features and reduce WRRN interference.

[0016] 3) The ADTW-SDM algorithm improves data fusion accuracy by revealing the global properties of modulation damage information and accurately weighting the fusion sub-dictionaries.

[0017] 4) The dual-feature adaptive threshold achieves high-precision rail damage detection and location, and the reliability and effectiveness of the method are verified through experimental tests under various conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a conceptual diagram of the dictionary fusion rail damage location method based on dynamic time warping.

[0019] Figure 2 Figure 2: Wheel-rail rolling test platform with prefabricated damage. (a) Front view of the platform, (b) Rail with prefabricated damage, (c) Strip damage at 350 mm, (d) Square damage at 684 mm, (e) Hydraulic device, (f) Detection wheel and onboard AE sensor, (g) Three VS900-RIC AE sensors, (h) Vallen AE acquisition system.

[0020] Figure 3 It is a detection wheel rolling distance estimation algorithm based on three vehicle-mounted sensors.

[0021] Figure 4 Time domain waveforms of multi-channel signals under 30-degree initial conditions: (a) channel 1, (b) channel 2, and (c) channel 3.

[0022] Figure 5 Peak detection frames of the time domain signal under 30-degree initial conditions, (a) channel 1, (b) channel 2, and (c) channel 3.

[0023] Figure 6 Peak detection box and rolling distance curves of the detection wheel under four initial conditions: (a) 0 degrees, (b) 30 degrees, (c) 60 degrees, and (d) 0 degrees.

[0024] Figure 7 The fusion signals obtained using the proposed method under different initial conditions: (a) 0 degrees, (b) 30 degrees, (c) 60 degrees, and (d) 0 degrees.

[0025] Figure 8 The mapping relationship between two dictionary atoms, (a) traditional DTW algorithm, (b) ADTW-SDM algorithm.

[0026] Fig. 9 The best paths obtained by the traditional DTW algorithm and the ADTW-SDM algorithm.

[0027] Fig.10 Normalized RPG and normalized STZCR, (a) normalized PRG, (b) normalized STZCR.

[0028] Fig.11 Damage detection and location results under different initial conditions: (a) 0 degrees, (b) 30 degrees, (c) 60 degrees, and (d) 0 degrees. DETAILED DESCRIPTION

[0029] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0030] The present invention provides a dictionary fusion rail damage location method based on dynamic time warping, in which the wheel rolling distance is estimated by segmenting the multi-channel AE signal into peak detection frames and calculating the time difference between adjacent frames. At the same time, the multi-channel AE signal trains the sub-dictionary and the benchmark dictionary in previous work. The multi-dictionary SR algorithm based on improved DTW has been innovated to synthesize a unified fusion dictionary through sub-dictionary weighting, enhance the damage feature information, and reduce most of the WRRN. In this process, the innovative ADTW-SDM algorithm embodies the global characteristics of dictionary atoms, and accurately evaluates the similarity between atoms through the second-order difference of the Mahalanobis distance. A dual-feature adaptive threshold is designed for detecting and locating damage on rails. The reliability and stability of the method proposed in the present invention under various conditions were further verified on the wheel-rail experimental platform, proving its potential in rail SHM, and providing guidance for the subsequent precise detection and location of large-scale rail damage based on advanced AE technology. Figure 1 As shown, the specific steps are as follows:

[0031] Step 1: Divide the multi-channel AE data set collected by multiple vehicle-mounted AE sensors into detection frames and perform normalization processing on them. Then, the frames with energy greater than the threshold are recorded as peak detection frames. By calculating the maximum delay between adjacent peak detection frames and determining the relative position of the sensor on the detection wheel at the corresponding time, the rolling distance of the detection wheel is estimated by measuring two adjacent relative positions. The specific steps are as follows:

[0032] Step 1: In order to accurately track the rolling distance of the detection wheel under the speed change state, the present invention proposes a wheel rolling distance estimation algorithm using adjacent peak detection frames. After the train runs for a distance, the sensors closest to the wheel-rail contact surface are marked as sensor q1 and sensor q2 from short to long distance, and their normal angles with the wheel-rail contact surface are θ1 and θ2 respectively. The time required from the wheel-rail contact point to the two sensor positions is expressed as Γ1 and Γ2 respectively. The wheel radius is defined as r, and the rolling distance during this period is recorded as R.

[0033] To be more universal, for a set of multi-channel AE datasets collected by vehicle-mounted AE sensors Where q is the serial number of the sensor and data channel, Q is the number of sensors, M is the length of each column of signals, and N is the number of signal columns for each channel. is the domain of integers, is the field of real numbers, Each column in is regarded as a frame signal, where Represents X qThis frame segmentation preprocessing helps improve the efficiency of subsequent operations. To avoid the influence of energy differences between each sensor, the signal of each channel is normalized as follows:

[0034]

[0035] in, represents the j-th normalized element in the i-th column of the q-th channel, represents the j-th element of the i-th column in the q-th channel signal set, where min(·) and max(·) are the minimum and maximum values, respectively.

[0036] Step 1 and 2: Extract the difference ΔE between the root mean square of each frame signal and the average energy value of the entire signal to perform peak detection. The specific formula is as follows:

[0037]

[0038]

[0039] Where ave(·) is the averaging operation, and W=128 is the length of the detection window. A detection frame can be expressed as Indicates that the i-th column in the detection frame corresponding to the q-th channel contains more than two non-zero The frames are called peak detection frames, where P is the number of peak detection frames.

[0040] Step 13: Since the relative distance between the onboard sensor and the rail will change with the rotation of the wheel, the amplitude and delay between the non-zero detection frames corresponding to each channel are different. To determine the order of the distance between the sensor and the track, the arrival time of each channel data in all peak detection frames can be traversed and its sequence number can be sorted in ascending order to form a list Λ q =[q1,q2,...,q Q ], the first two channels q1 and q2 are the channels closest to the track. Then, the arrival time difference between two adjacent sensors can be calculated by the maximum delay obtained by the correlation of the corresponding peak detection frames, as follows:

[0041]

[0042] in, represents the maximum delay of the t-th peak detection frame between the q1-th and q2-th channels, τ′ is the number of delay points, represents the jth point in the tth peak detection frame in the q1 channel. Then the maximum delay can be expressed as:

[0043]

[0044] Among them, θ s is the angle between two adjacent sensors, v represents the propagation speed of the signal on the tread of the detection wheel, then the distance that the detection wheel rolls between two adjacent peak detection frames can be expressed as:

[0045]

[0046] Among them, θ1′(q i+1 ) represents the new angle between sensor 1 and the rail surface normal in the i+1th peak detection frame, θ1(q i ) represents the angle between sensor 1 and the rail surface in the i-th peak detection frame. By estimating the rolling distance between all adjacent peak detection frames, the complete wheel trajectory can be tracked. In addition, combined with the proposed method, the identification and location of track damage can also be achieved.

[0047] Step 2: In order to solve the problem that WRRN in multi-channel data can easily drown out effective damage information and cause false detection, the present invention proposes a multi-dictionary SR algorithm based on improved DTW, which is used to fuse effective damage information in multi-channel data, enhance damage characteristics and reduce WRRN. The algorithm establishes a sub-dictionary for each channel, and uses K-SVD and ADMM to solve the sparse coefficients of the sub-dictionary and the corresponding channel respectively, and calculates the weights through the improved DTW algorithm to fuse the multi-channel sub-dictionary and the joint sparse coefficients. Finally, the fused signal is reconstructed by fusing the dictionary and the joint sparse coefficients, thereby enhancing the damage characteristics and eliminating WRRN. The specific steps are as follows:

[0048] Step 21: Given a multi-channel AE dataset X, according to the SR theory, we can find a set of sparse bases, map them to a high-dimensional space, and construct a multi-channel AE joint sparse expression:

[0049]

[0050] Among them, D q represents the sub-dictionary corresponding to the qth channel, which is trained by the K-SVD algorithm, w q is the corresponding sub-dictionary D q The weight of represents the sparse vector of the i-th row corresponding to the q-th channel, and λ represents a regularization parameter greater than zero.

[0051] Step 22: Further calculate the residual of the SR process:

[0052]

[0053] Where K represents the number of dictionary atoms, Represents the sub-dictionary D corresponding to the qth channelq exist The reconstructed residual on Indicates D q The k-th column of atoms in . A new auxiliary variable z is introduced and combined with the residual to solve the sparse coefficient, the formula is as follows:

[0054]

[0055] Where D represents the fusion dictionary of weighted joint sparse expression, s u,i represents the joint sparse coefficient of the i-th column. μ ≥ 0 is the Lagrange multiplier, ξ k is the residual of the kth column after the joint sparse expression. After solving the sub-dictionary and sparse matrix corresponding to each channel, we can further perform multi-dictionary fusion and joint sparse expression by adjusting the weights.

[0056] In the process of joint sparse expression, WRRN follows Gaussian distribution and has no autocorrelation, so it cannot be sparsely expressed. On the contrary, the effective information of damage in multi-channel AE data is retained and integrated into the fusion dictionary.

[0057] Step 3: In order to accurately calculate the fusion weights between sub-dictionaries, the present invention proposes the ADTW-SDM algorithm for accurately evaluating the correlation between atoms and damage information, thereby efficiently fusing multi-channel sub-dictionaries. In this algorithm, the DTW algorithm is introduced, and the second-order difference of the Mahalanobis distance is used to replace the Euclidean distance to improve the accuracy of distance calculation. In addition, a penalty function is added to the DTW path calculation to avoid incorrect time axis stretching, thereby accurately calculating the similarity between the multi-channel sub-dictionary atoms and the reference dictionary atoms, and providing the best dictionary fusion weight. The specific steps are as follows:

[0058] Step 31: Use the K-SVD algorithm to train a baseline impairment dictionary Φ = [φ k |k=1,2,...,K], where φ k is the kth column atom in Φ, and K is the total number of atoms in the base dictionary. It is used to evaluate the contribution of atoms in each sub-dictionary to the damage information, so that the weight can be accurately estimated. Then a time warping matrix is ​​created To store the distance between two dictionary atoms, For L d The element at row j2 and column j1 in .

[0059] Step 32: DTW distance reflects the similarity between atoms and reference atoms. Atoms similar to reference atoms are given larger weights to enhance damage features, while atoms containing WRRN are weighted less. Replacing Euclidean distance with Mahalanobis distance can reduce noise interference. At the same time, the second-order difference of the Mahalanobis Distance (SDM) between adjacent points can more intuitively reveal the shape trend of the mapped atoms and incorporate their global features into the measurement. The Mahalanobis volume distance formula between two points is as follows:

[0060]

[0061] Among them, Λ represents the covariance between two vectors. Subsequently, the SDM between adjacent points is calculated to replace the Euclidean distance in the original DTW, and the formula is as follows:

[0062]

[0063] Where k1 and k2 are the column numbers of the dictionary atoms, and j1 and j2 are the numbers of the elements in the atoms. In addition, incorrect time axis stretching will produce more uneven mappings, resulting in longer mapping paths calculated by DTW. Therefore, the path penalty function C is introduced p , let the algorithm give priority to points within a limited range to minimize incorrect stretching and improve computational efficiency, then the distance formula between two dictionary atoms is as follows:

[0064]

[0065]

[0066] in, is the j1th element in the k1th column of Φ, D q The j2th element in the k2th column of .

[0067] Step 33: Calculate in sequence and and select the smallest value as Then the entire matrix L d Backtrack until the minimum path is found as the improved DTW distance, the formula is as follows:

[0068]

[0069] Furthermore, the fusion weight formula between each sub-dictionary atom is as follows:

[0070]

[0071] The weight is used to perform dictionary fusion and reconstruct the fusion signal. The formula is as follows:

[0072]

[0073]

[0074] in is the i-th column in the reconstructed fusion signal.

[0075] Step 4: In order to accurately detect the damage signal from the fused signal, the present invention designs a dual-feature adaptive threshold to achieve accurate damage detection and positioning. Resonance Peaks Gain (RPG) is a feature widely used in signal processing and acoustic analysis, but it is easily disturbed by signal amplitude fluctuations. Combined with the Short-Time Zero Crossing Rate (STZCR), the damage information is identified by detecting the frequency of the signal crossing the zero axis, and an adaptive threshold (AT) is constructed to further compensate for the defects of the RPG feature. The specific steps are as follows:

[0076] Step 41: Calculate the RPG feature and SRZCR feature for each column of fusion signal respectively. The formula is as follows:

[0077]

[0078]

[0079] Among them, R PG,i is the fusion signal of column i RPG features, S TZCR,i is the fusion signal of column i The STZCR characteristics of , sgn(·) is the sign function, and DFT(·) is the discrete Fourier transform.

[0080] Step 42: Use the regularized RPG feature and the regularized STZCR feature to construct a dual-feature adaptive threshold. The formula is as follows:

[0081]

[0082] Among them, f N (·)∈[-1,1] is the regularization function, T w =150 is the threshold window length, and η=0.5 are both constants. By finding the moment when the amplitude in the fusion signal exceeds the adaptive threshold AT, it is recorded as a damage signal. At the same time, combined with the rolling distance of the detection wheel in the corresponding time period obtained in step 1, the distance between the damage and the detection starting point can be obtained, thereby achieving accurate rail damage positioning.

[0083] The specific implementation of the present invention is described below in conjunction with rail tensile fracture test data:

[0084] In order to obtain AE signals closer to the actual SHM environment, a special wheel-rail rolling test platform with prefabricated defects was designed, such as Figure 2 The front view of the platform is shown in Figure 2 As shown in (a), the foundation of the experimental platform uses a two-meter-long rail (made of U71Mn) cut from an in-service rail. Figure 2 As shown in (b), (c) and (d), there are two artificial prefabricated defects in the rail surface etching. One is located between 349 mm and 357 mm (50×8×2 mm 3 ) and the other is located between 684 mm and 719 mm (25×35×2mm 3 ) square damage. Figure 2 As shown in (e), the additional hydraulic device connected to the frame drives the hydraulic rod and provides radial stress to simulate the running load of the actual train. At the same time, the detection wheel (diameter: 300 mm) fixed on the frame rolls on the rail to generate WRRN and AE signals, as shown in Figure 2 (f) Figure 2 As shown in (g), three onboard AE sensors (VS900-RIC) are fixed on the surface of the detection wheel through a 120-degree magnetic clamp to collect multi-channel AE signals. Figure 2 As shown in (h), the AE signal acquisition system (model: Vallen AMSY-6) is used to receive multi-channel AE signals in real time and store them as a data matrix in a continuous mode for subsequent analysis. Different AE data can be obtained by setting different initial conditions for the angle between the sensor and the track surface normal. The platform can verify the reliability and effectiveness of the proposed method by changing the angle between the initial sensor and the track surface normal.

[0085] Specifically, the sensor closest to the track surface is marked as No. 1, and the other sensors are marked as No. 2 and No. 3 in counterclockwise order, such as Figure 3As shown. First, adjust the wheel position so that sensor No. 1 is located directly above the track surface (the angle with the normal line of the track surface is 0 degrees). Connect it to the AE acquisition system through a shielded signal cable. Set the sampling rate to 5MHz, continuous mode. The system will save the acquisition signals of each channel into a data matrix in chronological order, with 2048 samples as a column. Then push the detection wheel to roll at a constant speed and stop after passing two defect positions. The multi-channel data of the entire process is saved as a set of experimental data for subsequent analysis. Then adjust the detection wheels respectively so that the angles between sensor No. 1 and the normal line of the rail surface are 30 degrees, 60 degrees and 90 degrees respectively. Repeat the above steps. The experiment at each angle is repeated twice, and a total of eight sets of data are obtained. Then the method proposed in the present invention is used to fuse the multi-channel signals and locate the damage position.

[0086] Execute step 1: Based on the above experimental process, the data of each channel obtained contains two damaged signals. For the convenience of calculation, the parts before starting and stopping have been deleted from the original signal. Each channel in the data set contains 6500 columns of signal matrix, with a length of 2048 sampling points and a total duration of 2.6624 seconds. Taking the time domain signal at an angle of 30 degrees as an example, the time domain waveforms of each channel are as follows Figure 4 shown.

[0087] It is obvious that in the time domain waveforms of the three channels, the damage signal has been completely submerged in the WRRN. In addition, since sensor No. 1 is closer to the rail, the acquired AE signal has a higher amplitude and energy. In order to determine the wheel rolling distance, it is recommended to use adjacent peak detection frames to estimate the wheel rolling distance. After normalizing each channel signal separately, the detection frame is divided into columns. Invalid frames are eliminated, and the peak detection frame is indicated by a red asterisk, as shown in Figure 2. Figure 5 shown.

[0088] It can be seen from the figure that the peak detection frames are widely distributed in the three channel signals. In the startup phase, due to the low speed, fewer peak detection frames are generated, while as the speed increases, the peak detection frames are evenly distributed with a higher density. A large number of peak detection frames improves the accuracy of wheel rolling speed estimation. Adjacent peak detection frames are used to estimate the rolling distance of the detection wheel under four initial conditions, such as Figure 6 As shown. The blue line and red asterisk represent the normalized channel 1 signal and its peak detection frame, respectively. The red curve represents the rolling distance curve. The initial stage of about 0s to 0.2s is the acceleration stage, followed by the uniform motion stage, and then enters the deceleration and stop stage after 2.5s. This result proves that the adjacent peak detection framework proposed in the present invention can effectively monitor the variable speed motion of the detection wheel and its rolling distance.

[0089] Execute step 2: While recording the rolling distance of the detection wheel, detecting the damage information in the multi-channel data set is crucial for the SHM of the rail. In order to clarify the effectiveness of the method proposed in the present invention, the method proposed in the present invention is used for fusion denoising in the dictionary fusion process. Under the initial condition of 30 degrees, the fusion signal obtained based on the algorithm proposed in the present invention is as follows Figure 7 shown.

[0090] Figure 7 The red arrow indicates the amplitude of the damage signal, and the black arrow indicates the amplitude of the interference signal. The maximum damage signal amplitude of the algorithm proposed in this invention reaches 5.8, which is 134% higher than the original signal. The maximum interference amplitude is 0.73, which is 56.29% lower than the original amplitude. The algorithm proposed in this invention has the most obvious damage signal and the smallest interference.

[0091] Execute step three: The ADTW-SDM proposed in the present invention takes into account the global properties of atoms, replaces the Euclidean distance with the second-order difference of the Mahalanobis distance, reduces the warping error, and improves the accuracy of atomic distance and similarity evaluation. In order to verify its performance, the distances between the three channel sub-dictionary atoms and the reference dictionary atoms under the initial condition of 30 degrees are calculated using ADTW-SDM. For the sake of convenience, 90 points are intercepted, and the corresponding relationship between the square-damaged sub-dictionary atoms and the reference dictionary atoms is shown in the figure. Figure 8 The black dashed line indicates the mapping relationship between the corresponding points of two atoms.

[0092] Compared with the traditional DTW algorithm, the mapping relationship between two atoms in ADTW-SDM is more uniform and broader. The ADTW-SDM algorithm proposed in this invention can capture a wider range of atomic global features and avoid the calculation errors of interatomic distances caused by incorrect time axis stretching. In order to intuitively display the warping and mapping relationship between atoms, Fig. 9 Describes the path between the sub-dictionary atoms and the base dictionary atoms under the 30-degree initial condition. The abscissa on the curve represents the ordinal number of the element in the sub-dictionary atom, and the ordinate is the ordinal number of the mapped element on the base dictionary atom. It is obvious that the path of the ADTW-SDM algorithm is shorter than the traditional DTW algorithm that maps and stretches uniform distribution.

[0093] Execute step 4: Based on the fused signal, extract the normalized STZCR and RPG features to construct the adaptive threshold. The normalized STZCR and RPG of the fused signal under 30 degrees are as follows: Fig.10 shown.

[0094] exist Fig.10In the above, the normalized RPG is quite sensitive to the damaged signal, but it is easily interfered. In contrast, the normalized average STZCR has strong anti-interference ability, but it is easy to ignore the effective information. Therefore, the AT developed in the present invention combines the advantages of both and realizes the positioning function by accurately capturing the defect information. In order to verify the positioning accuracy of the method proposed in the present invention, the positioning deviation index is introduced to evaluate the positioning accuracy of the algorithm, and the formula is as follows:

[0095]

[0096] Among them, C1 and C2 are the front and rear ends of the damage location, respectively. e is the damage location located by the algorithm. Under the four initial conditions, the damage detection and location results based on the fusion signal are as follows: Fig.11 As shown in Figure 2, for strip damage, the average error and LD under each experimental condition are 56 mm and 4.00%, respectively. For square damage, the average error and LD under all conditions are 16.25 mm and 1.16%, respectively. The above results confirm that the proposed method can effectively detect and locate damage from a noisy background, providing guidance for the practical application of large-scale rail SHM based on acoustic emission technology.

Claims

1. A dictionary fusion rail damage location method based on dynamic time warping, characterized by The method comprises the following steps: Step 1: Divide the multi-channel AE data set collected by multiple vehicle-mounted AE sensors into detection frames and perform normalization processing on them. Then, the frames with energy greater than the threshold are recorded as peak detection frames. The maximum delay between adjacent peak detection frames is calculated, and the relative position of the sensor to the detection wheel at the corresponding time is determined. The rolling distance of the detection wheel is estimated by measuring two adjacent relative positions. Step 2: Establish a sub-dictionary for each channel, and use K-SVD and ADMM to solve the sparse coefficients of the sub-dictionary and the corresponding channel respectively, and use the improved DTW algorithm to calculate the weights for fusing multi-channel sub-dictionaries and joint sparse coefficients. Finally, reconstruct the fused signal by fusing the dictionary and the joint sparse coefficients. Step 3: Introduce the DTW algorithm, use the second-order difference of the Mahalanobis distance to replace the Euclidean distance, add a penalty function when calculating the DTW path, calculate the similarity between the multi-channel sub-dictionary atoms and the base dictionary atoms, and provide the best dictionary fusion weight; Step 4: Combine the short-time zero-crossing rate to identify the damage information by detecting the frequency of the signal crossing the zero axis, and construct a dual-feature adaptive threshold to further make up for the defects of the RPG feature.

2. The method for locating rail damage based on dictionary fusion of dynamic time warping according to claim 1 is characterized in that The specific steps of step one are as follows: Step 1: After the train runs for a certain distance, the sensors closest to the wheel-rail contact surface are marked as sensor q1 and sensor q2 according to the distance from short to long. The angles between them and the normal of the wheel-rail contact surface are θ1 and θ2 respectively. The time required from the wheel-rail contact point to the two sensor positions is Γ1 and Γ2 respectively. The wheel radius is defined as r, and the rolling distance during this period is recorded as R. For a set of multi-channel AE data sets collected by on-board AE sensors Where q is the serial number of the sensor and data channel, Q is the number of sensors, M is the length of each column of signals, and N is the number of signal columns for each channel. is the domain of integers, is the field of real numbers, Each column in is regarded as a frame signal, where Represents X q The i-th column in , normalizes the signal of each channel, the formula is as follows: in, represents the normalized element of the i-th column in the j-th channel, represents the jth element in the i-th column of the q-th channel signal set, min(·) and max(·) are the minimum and maximum values, respectively; Step 1 and 2: Extract the difference between the root mean square of each frame signal and the average value of the entire signal energy to perform peak detection. The specific formula is as follows: Where ave(·) is the averaging operation, W is the length of the detection window, and a detection frame is expressed as Indicates that the i-th column in the detection frame corresponding to the q-th channel contains more than two non-zero The frames are called peak detection frames, where P is the number of peak detection frames; Step 13: Traverse the arrival time of each channel data in all peak detection frames, and sort their sequence numbers in ascending order to form a list Λ q =[q1,q2,...,q Q ], the first two channels q1 and q2 are the channels closest to the track, and the arrival time difference between two adjacent sensors is calculated by the maximum delay obtained by the correlation of the corresponding peak detection frames, as follows: in, represents the maximum delay of the t-th peak detection frame between the q1-th and q2-th channels, represents the jth point in the tth peak detection frame in the q1 channel, τ′ is the number of delay points, and the maximum delay is expressed by the propagation arc length of the signal as: Among them, θ s is the angle between two adjacent sensors, v represents the propagation speed of the signal on the tread of the detection wheel, and the distance that the detection wheel rolls between two adjacent peak detection frames is expressed as: Among them, θ1′(q i+1 ) represents the new angle between sensor 1 and the rail surface normal in the i+1th peak detection frame, θ1(q i ) represents the angle between sensor 1 and the rail surface in the i-th peak detection frame. The complete wheel trajectory is tracked by estimating the rolling distance between all adjacent peak detection frames.

3. The method for locating rail damage based on dictionary fusion based on dynamic time warping according to claim 1 is characterized in that The specific steps of step 2 are as follows: Step 21: Given a multi-channel AE dataset X, according to the SR theory, find a set of sparse bases, map them to a high-dimensional space, and construct a multi-channel AE joint sparse expression: Among them, D q represents the sub-dictionary corresponding to the qth channel, w q is the weight of the corresponding sub-dictionary, represents the sparse vector of the i-th row corresponding to the q-th channel, and λ represents a regularization parameter greater than zero; Step 22: Calculate the residual of the SR process: Where K represents the number of dictionary atoms, Represents the sub-dictionary reconstruction residual corresponding to the qth channel, represents the k-th column atom in the q-th sub-dictionary; Introduce the auxiliary variable z and combine it with the residual to solve the sparse coefficient. The formula is as follows: Where D represents the fusion dictionary of weighted joint sparse expression, s u,i represents the joint sparse coefficient of the i-th column, μ ≥ 0 is the Lagrange multiplier, ξ k is the residual of the kth column after the joint sparse expression; after solving the sub-dictionary and sparse matrix corresponding to each channel, multi-dictionary fusion and joint sparse expression are further performed by adjusting the weights.

4. The method for locating rail damage based on dictionary fusion of dynamic time warping according to claim 1 is characterized in that The specific steps of step three are as follows: Step 31: Use the K-SVD algorithm to train a baseline impairment dictionary Φ = [φ k |k=1,2,...,K], where φ k is the kth column atom in Φ, K is the total number of atoms in the base dictionary, and is used to evaluate the contribution of atoms in each sub-dictionary to the damage information, creating a time-warping matrix Stores the distance between two dictionary atoms, where For L d The element at row j2 and column j1 in ; Step 32: The Mahalanobis volume distance formula between two points is as follows: Among them, Λ represents the covariance between two vectors, and the SDM between adjacent points is calculated to replace the Euclidean distance in the original DTW. The formula is as follows: Among them, k1 and k2 are the column numbers of the dictionary atoms, j1 and j2 are the numbers of the elements in the atoms; Introducing path penalty function C p , then the distance formula between two dictionary atoms is as follows: in, is the j1th element in the k1th column of Φ, D q The j2th element in the k2th column of ; Step 33: Calculate in sequence and and select the smallest value as For the entire matrix L d Backtrack until the minimum path is found as the improved DTW distance, the formula is as follows: Furthermore, the fusion weight formula between each sub-dictionary atom is as follows: The weight is used to perform dictionary fusion and reconstruct the fusion signal. The formula is as follows: in, is the i-th column in the reconstructed fusion signal.

5. The method for locating rail damage based on dictionary fusion of dynamic time warping according to claim 1 is characterized in that The specific steps of step 4 are as follows: Step 41: Calculate the RPG feature and SRZCR feature for each column of fusion signal respectively. The formula is as follows: Among them, R PG,i is the fusion signal of column i RPG features, S TZCR,i is the fusion signal of column i The STZCR characteristics of , sgn(·) is the sign function, and DFT(·) is the discrete Fourier transform; Step 42: Use the regularized RPG feature and the regularized STZCR feature to construct a dual-feature adaptive threshold. The formula is as follows: Among them, f N (·)∈[-1,1] is the regularization function, T w is the threshold window length, and η are both constants; by finding the moment when the amplitude of the fused signal exceeds the adaptive threshold AT, it is recorded as a damage signal, and combined with the rolling distance of the detection wheel in the corresponding time period obtained in step 1, the distance between the damage and the detection starting point is obtained, thereby achieving accurate rail damage positioning.

Citation Information

Patent Citations

  • Robust Speech Feature Extraction Method Based on Sparse Decomposition and Reconstruction

    CN102290047A

  • Gesture recognition method based on acceleration sensor

    CN103984416A

  • Speech conversion method between any sources and target speech

    CN107221321A

  • Two-stage damage location recognition method based on guided-wave signal sparse decomposition and damage localization

    CN110542723A

  • Steel rail health monitoring method based on information enhancement and variable step size sparse expression

    CN115236208A

Cited By

  • Multi-sensor correlation joint sparse expression steel rail structure health monitoring method

    CN120177568A