On-board rail damage acoustic emission detection method based on dictionary fusion enhancement

Abstract

The present invention discloses a vehicle-mounted rail damage acoustic emission detection method based on dictionary fusion enhancement. First, a mirror-extended adaptive local mean decomposition algorithm is proposed to reconstruct a multi-channel data set to eliminate random noise in the signal. Through mirror expansion and adaptive adjustment of the sliding step size, the local details and global characteristics of the signal are balanced, and the endpoint effect and modal aliasing are avoided. At the same time, an innovative correlation constraint enhanced dictionary fusion algorithm based on Cramér's V coefficient is developed to train a unified dictionary, fuse redundant effective information in multi-channel signals, and further eliminate WRRN. Finally, the sample entropy envelope is extracted from the reconstructed fusion signal, and an adaptive threshold is constructed to accurately detect damage and prompt staff to maintain the rails in time. This method can effectively enhance the damage characteristics, detect damage signals submerged by noise, and provide guidance for rail damage analysis and evaluation.

Description

Technical Field

[0001] The present invention belongs to the field of structural health monitoring (SHM) of rails, and relates to a vehicle-mounted rail damage acoustic emission detection method, and specifically to a vehicle-mounted rail damage acoustic emission detection method based on dictionary fusion enhancement. Background Art

[0002] With the advancement of urbanization, the importance of railway transportation has become increasingly prominent. As an important part of the railway network, surface peeling and cracks are the main types of damage that lead to rail breakage and even train derailment. Therefore, timely and accurate identification of new damage is crucial to maintaining railway transportation safety.

[0003] Acoustic emission (AE) testing is a prominent nondestructive testing technology known for its high sensitivity and non-invasiveness, and can be used for passive rail SHM. However, it still faces many challenges in practical applications. The main obstacle is that the amplitude of the damage signal is usually weak and can be easily overwhelmed by the complex wheel rail rolling noise (WRRN) caused by wheel-rail contact and mechanical friction, resulting in missed detection and false detection. These WRRN contain both random noise distributed at low frequencies and low amplitudes, and self-excited noise with high kurtosis characteristics. These noises overlap with the frequency band distribution of the damage signal, making it difficult to remove them by traditional denoising methods. Local mean decomposition (LMD), as an emerging signal denoising method, reveals the intrinsic characteristics of non-stationary signals through adaptive decomposition and effectively eliminates random noise. However, the LMD algorithm has band aliasing and endpoint effects, and is heavily dependent on preset parameters, which weakens the denoising effect of LMD on random noise.

[0004] At the same time, the existing signal denoising and rail detection methods rely on the data of a single sensor, and the damage detection accuracy is limited by the accuracy of the sensor signal, and the monitoring range is limited. Using mobile vehicle-mounted multi-sensors to monitor the health status of a large range of rail structures can break through the limitations of the sensor detection range, and the data between multiple channels can form redundant complementarity. Dictionary fusion is a potential data fusion method that sparsely expresses damage information by training a unified fusion dictionary to achieve damage detection. Finding the intrinsic connection between multi-channel data and fully integrating the submerged effective information is the key to further eliminate WRRN and enhance damage information.

[0005] In addition, affected by the detection environment, the AE signals collected during each SHM process have different amplitudes. In addition to WRRN, the signals also contain some interference that cannot be eliminated, making it impossible to effectively detect damage signals by relying on a fixed threshold, which can easily lead to false detection. Summary of the invention

[0006] In order to achieve accurate rail SHM under complex background noise and provide guidance for rail structural health monitoring under noise background, the present invention provides a vehicle-mounted rail damage acoustic emission detection method based on dictionary fusion enhancement. This method can effectively enhance damage characteristics, detect damage signals submerged by noise, and provide guidance for rail damage analysis and evaluation.

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

[0008] A vehicle-mounted rail damage acoustic emission detection method based on dictionary fusion enhancement includes the following steps:

[0009] Step 1: Perform a mirror flip operation on the multi-channel AE signal dataset X, and build a mirror extension-multi-channel AE dataset X based on the original data e , improve the continuity of the signal segment endpoints to avoid endpoint effects and modal aliasing, and then use the short-term energy of the signal segment to adjust the adaptive smoothing step size S L , to extract the local envelope of the signal and solve the FM signal S m,ij (t), and finally retain the product function (PF) with the largest mutual correlation coefficient to reconstruct the multi-channel AE dataset Eliminate random noise;

[0010] Step 2: For the reconstructed multi-channel AE dataset, the K-SVD algorithm is used to train a sub-dictionary Φ for the reconstructed signal of each channel q , then the atoms in each sub-dictionary are weighted according to the Cramér V coefficient and merged into a unified dictionary Φ f In the process of solving the sparse vector, the shared sparse coefficient B is solved by adding correlation constraints, and the ADMM algorithm is used to solve it to eliminate the residual WRRN, and finally reconstruct the fusion signal X re =Φ f Β;

[0011] Step 3: Based on the reconstructed fusion signal X re , for each column signal The sample entropy SE is extracted, and then the adaptive threshold Th is calculated based on the SE value of each column of the fused signal. Finally, the signal segments in the fused signal with amplitudes higher than the adaptive threshold are found to be damaged signals, and other signals below the adaptive threshold are non-damaged signals.

[0012] Compared with the prior art, the present invention has the following advantages:

[0013] 1) This paper proposes an enhanced ME-ALMD algorithm, which can adaptively adjust the sliding step size to preserve the local details and global characteristics of the signal. It can effectively eliminate the random noise in WRRN without relying on prior information and is not affected by endpoint effects and modal aliasing.

[0014] 2) The innovative EDF-RCC algorithm can fuse the redundant information in multi-channel AE signals to enhance the damage characteristics. The algorithm further eliminates WRRN by weighted sub-dictionary fusion and adding sparse constraints.

[0015] 3) The adaptive threshold method based on sample entropy envelope can accurately identify damage signals, and the effectiveness and reliability of this method have been verified on a customized experimental platform under various conditions, showing significant advantages over traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a conceptual diagram of the on-board rail damage acoustic emission detection method based on dictionary fusion enhancement.

[0017] Figure 2 Equipment diagram of the wheel-rail rolling test platform, (a) front view of the platform, (b) test wheel and on-board AE sensor, (c) prefabricated damage location, (d) strip crack, (e) square defect, (f) three VS900-RICAE sensors, and (g) VallenAE acquisition system.

[0018] Figure 3 It is the time domain waveform of the multi-channel signal under the initial condition of 30 degrees.

[0019] Figure 4 ME-ALMD decomposition results of 1-channel signal fragment under 30-degree initial conditions, (a) original signal, (b) PF1, (c) PF 2, (d) PF 3, (e) PF 4, (f) PF 5, (g) PF 6, (h) PF 7.

[0020] Figure 5 The time domain waveform of the multi-channel data set reconstructed based on the screened optimal product function under the initial condition of 30 degrees, (a) the denoising result of channel 1, (b) the denoising result of channel 2, and (c) the denoising result of channel 3.

[0021] Figure 6 Time domain waveforms of the reconstructed fusion signal obtained under four initial conditions: (a) 0 degrees, (b) 30 degrees, (c) 60 degrees, and (d) 90 degrees.

[0022] Figure 7It is the sample entropy and lower envelope curve of the fusion signal at an angle of 30 degrees.

[0023] Figure 8 The fusion signal and its adaptive threshold under each different initial experimental condition: (a) 0 degrees, (b) 30 degrees, (c) 60 degrees, and (d) 90 degrees. DETAILED DESCRIPTION

[0024] 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.

[0025] The present invention provides a vehicle-mounted rail damage acoustic emission detection method based on dictionary fusion enhancement. Firstly, for WRRN in multi-channel AE signals, a mirror extension-based adaptive local mean decomposition algorithm (Mirror Extension-based Adaptive LMD, ME-ALMD) is proposed based on the LMD algorithm to reconstruct the multi-channel data set to eliminate random noise in the signal. Through mirror extension and adaptive adjustment of the sliding step size, the local details and global features of the signal are balanced, and the endpoint effect and modal aliasing are avoided. Compared with the traditional method, the algorithm has a smaller root mean square error at the end of the decomposition result, while retaining the local details and global features of the original signal. The random noise in the WRRN is eliminated by utilizing the optimal PF screened by cross-correlation. At the same time, an innovative enhanced dictionary fusion algorithm with relevance constraints based on Cramér's V coefficient (Enhanced Dictionary Fusion Algorithm with RelevanceConstraints based on Cramér's VCoefficient, EDF-RCC) is also developed to train a unified dictionary, fuse the redundant effective information in the multi-channel signal, and further eliminate the WRRN. The fused signal has more obvious damage characteristics. Finally, the sample entropy envelope is extracted from the reconstructed fused signal and an adaptive threshold is constructed to accurately detect damage and prompt the staff to maintain the rails in time. Figure 1 As shown, the specific steps are as follows:

[0026] Step 1: Perform a mirror flip operation on the multi-channel AE signal data set, and build a mirror extension-multi-channel AE data set based on the original data to improve the continuity of the signal segment endpoints to avoid endpoint effects and modal aliasing. Then use the short-term energy of the signal segment to adjust the adaptive smoothing step size to extract the local envelope of the signal and solve the frequency modulation signal. Finally, retain the product function with the largest cross-correlation coefficient to reconstruct the multi-channel AE data set and eliminate random noise. The specific steps are as follows:

[0027] Step 1: For a multi-channel AE dataset Where Q is the number of channels, q is the channel number, M is the length of each column of signals, and N is the number of signal columns for each channel. is an integer domain, and the signal fragment of the jth column in the qth channel is Reverse along the time axis to generate mirror data symmetrical to the original signal in is the mth element in the jth column signal in the qth channel.

[0028] Step 1 and 2: Combine the generated image data with the original data to construct an image extension-AE signal containing the original signal and its image:

[0029]

[0030] in, represents the mirror extension-AE signal of the jth column in the qth channel, and then Constructing mirror extension-multi-channel AE dataset Through this operation, the continuity of the signal segments at the endpoints is significantly enhanced, thereby effectively reducing the endpoint effect and modal aliasing and improving the overall signal quality.

[0031] Step 13: Calculate the short-time energy of each signal segment, and construct an adaptive step size S that can be dynamically adjusted according to the size of the short-time energy. L , the formula is as follows:

[0032]

[0033] in, is the initial step length, η is a constant, For rounding up operation, η≥2 is the minimum iteration step, ensuring that the minimum iteration step is used to decompose the signal. The smoothing step is dynamically adjusted according to the signal energy. The energy of the damaged part of the signal is concentrated and is related to S L The value of is inversely proportional to the value of S, so that more detail information can be mined with a smaller step size and the detail resolution can be improved. In the noise part, due to the dispersion of signal energy, S L If the value of is large, a large step size is used to ignore the WRRN in the area susceptible to noise, reduce interference and improve the computational efficiency of the algorithm.

[0034] Step 14: Extract X e The local mean α of each column signal in i and the local envelope γ i , the formula is as follows:

[0035] α i=(σ i +σ i+1 ) / 2

[0036] γ i =|(σ i -σ i+1 ) / 2|

[0037] Among them, σ i is the i-th extreme point of each column signal.

[0038] Step 15: Based on the adaptive step size, the local mean function (LMF) and envelope estimation function (EEF) are calculated by averaging the sliding operation. The formulas are as follows:

[0039]

[0040] Among them, j represents the column number, Ls is the initial sliding step size, and f a,ij (t) represents the i-th element of the j-th column of the local mean function at time t, f γ,ij (t) represents the i-th element of the j-th column of the envelope estimation function at time t.

[0041] Step 16: Combine f a,ij (t) and f γ,ij (t) Construct frequency modulation signal S m,ij (t), the formula is as follows:

[0042]

[0043] Among them, PF h is the product function generated by the hth iteration, x(t) represents the signal element at time t, and the above iterative process is repeated until a pure FM signal is generated, and the product function PF of the hth round is obtained. h .

[0044] Step 17: In order to eliminate the random noise in the signal, choosing the stop condition of decomposition is also crucial for effective denoising. Given that the random noise in WRRN follows a Gaussian distribution and its correlation is lower than that of the damage signal, the retained PF is determined based on the cross-correlation coefficient between different channels and the multi-channel data set is reconstructed. in, represents the reconstructed signal set corresponding to the qth channel, and the formula is as follows:

[0045]

[0046] in, represents the PF generated by the h1th iteration in the q1 channel, Rc represents the cross-correlation function, and IME(·) is the inverse operation of the mirror extension. The PF with the largest sum of cross-correlation coefficients is retained, thereby maximizing the retention of effective information.

[0047] Through the above steps, the integrity of local details can be maintained while improving signal continuity, and random noise and valid information in the signal can be effectively distinguished during the processing process. The random noise can be eliminated from the reconstructed multi-channel data set by screening the optimal PF.

[0048] Step 2: In order to fuse and reconstruct the redundant information in the multi-channel data set and further eliminate the residual WRRN that has frequency band overlap with the damage signal, the present invention proposes an EDF-RCC algorithm for training a unified fusion dictionary and reconstructing the denoised fusion signal for subsequent high-precision damage detection. The EDF-RCC algorithm innovatively weights each sub-dictionary atom according to the Cramér V coefficient and fuses them into a unified dictionary, thereby fusing multi-channel data and enhancing the ability of atoms to characterize damage characteristics. In the process of sparse vector solution, the shared sparse coefficients are solved through correlation constraints to minimize the residual invalid components, thereby completely eliminating WRRN. The specific steps are as follows:

[0049] Step 21: Since the reconstructed multi-channel data set is generated by the cross-correlation coefficient, it means that each channel contains a lot of redundant information. Therefore, the fusion reconstructed multi-channel data set can integrate this effective information and enhance the damage characteristics. Specifically, the sparse expression form of the reconstructed multi-channel data set is constructed as follows:

[0050]

[0051] Among them, Φ q is the sub-dictionary corresponding to the qth channel, is the sparse coefficient matrix Β corresponding to the qth channel q In the dictionary training stage, the contribution of each atom to the same damage representation can be measured by the degree of association between atoms at corresponding positions in each sub-dictionary. The sub-dictionary corresponding to each channel is trained using the K-order singular value decomposition (K-SVD) algorithm.

[0052] Step 22: Since the contribution of each atom to the same damage representation can be measured by the degree of association between atoms at corresponding positions in each sub-dictionary during the dictionary training stage, the present invention introduces Cramér's V coefficient to measure the strength of association between two random variables. It can better explain the internal connection of unstable vectors from the perspective of global characteristics than the traditional inner product method. The formula of Cramér's V coefficient is as follows:

[0053]

[0054] Among them, χ 2 represents the chi-square statistic, θ1 and θ2 represent the two sequences in the cross table, g and u represent the number of columns and rows in the cross table, respectively. In the process of dictionary fusion reconstruction of damage signals, the contribution of each atom based on Cramér's V coefficient can be formulated as:

[0055]

[0056] in, for The k-th column signal in k q represents the k-th column of the sub-dictionary atom, Represents the sparse coefficient of the kth row corresponding to the qth sub-dictionary.

[0057] Step 2 and 3: Weight each sub-dictionary atom by its atomic contribution to construct the fusion dictionary atom φ of the kth column f,k , the formula is as follows:

[0058]

[0059] The denominator represents the normalized sum of contributions.

[0060] Step 24: Fusion dictionary Φ f The optimization objective can be formulated as follows:

[0061]

[0062] Construct a unified fusion dictionary Φ by weighting sub-dictionary atoms f The fused atoms contain the intrinsic characteristics of all channel atoms, thereby achieving the purpose of enhancing damage information.

[0063] Step 25: Considering the similarity of the impairment signals of each channel, they have the same sparsity, so they can be sparsely expressed using a common sparse coefficient matrix. Based on the ADMM algorithm, the solution process of the joint sparse coefficient is as follows:

[0064]

[0065] Among them, the l2 norm enhances the simplicity of shared sparse coefficients, B is the joint sparse coefficient matrix, η is the regularization parameter, and R a(·) is the autocorrelation function, and ρ is the penalty coefficient. By adding the correlation constraint term, atoms with higher correlation are preferentially selected during the update process, thereby protecting the damage characteristics. At the same time, in the coefficient expression process, the WRRN part with low correlation will be discarded, thereby removing the interference term.

[0066] Step 26: Based on the fusion dictionary and the shared sparse coefficient matrix, the reconstructed fusion signal X can be obtained re , the formula is as follows:

[0067] X re =Φ f Β

[0068] The K-SVD algorithm is currently the most widely used algorithm due to its advantages of fast convergence speed, simplicity and high efficiency. Alternating Direction Multiplier Method (ADMM) is a popular multi-task distributed convex optimization algorithm in recent years, which can efficiently solve the coefficient expression problem of multi-channel signals. Therefore, the present invention combines K-SVD and ADMM to alternately solve the fusion dictionary atoms and sparse coefficients, and reconstructs the fusion signal X using the obtained fusion dictionary and sparse coefficients. re .

[0069] Step 3: During the structural health monitoring of rails, changes in the detection environment and wheel-rail contact status will change the amplitude of the collected AE data. In order to adapt to the changes in signal amplitude caused by different detection environments and wheel-rail contact states, and further eliminate the influence of interference waveforms on the detection results, an adaptive threshold is constructed based on the reconstructed fusion signal to extract the lower envelope of the sample entropy, and the signal segments with amplitudes higher than the adaptive threshold in the fusion signal are found as damaged signals, and other signals below the adaptive threshold are non-damaged signals. The specific steps are as follows:

[0070] Step 31: Since sample entropy can effectively track the probability of vector similarity fluctuating with increasing dimension to describe the complexity of the signal, for the more similar damaged signal, its sample entropy will decrease accordingly. Therefore, for the fused signal of column j in express , constructing a vector for each value

[0071] Step 32: Calculate the distance d(i1,i2) between each value. The formula is as follows:

[0072]

[0073] Where u = 1, 2, ..., M, i1 and i2 represent the element numbers, and the statistics are less than the similarity threshold T r= 0.1, denoted as w d [d(i1,i2)], then the formula for calculating sample entropy (SE) is as follows:

[0074]

[0075] Step 33: Use the lower envelope of SE of each fusion signal segment to construct an adaptive threshold Th, the formula is as follows:

[0076] Th=ωE lower (SE)+δ

[0077] Where ω = 0.5, δ is the penalty coefficient, and δ = 0.45, E lower (·) indicates the operation of extracting the lower envelope. The damage signal is confirmed by finding the peak of the fused signal with an amplitude greater than the adaptive threshold. When the fused signal amplitude is greater than the threshold, it indicates that there is damage in the signal segment, the structural health of the rail is damaged, and it should be maintained and replaced in time. Otherwise, the signal segment will be regarded as a non-damaged part, thereby achieving accurate detection of rail damage.

[0078] Example:

[0079] The specific implementation of the present invention is described below in conjunction with rail tensile fracture test data:

[0080] This embodiment uses a wheel-rail rolling test platform to collect track damage AE signals. Figure 2 As shown, the platform is based on rails and consists of several key components. These include a 2-meter rail (made of U71Mn) cut from an in-service rail, a detection wheel (diameter: 300 mm), a wheel frame, three onboard AE sensors, and three magnetic clamps. The attached hydraulic device can drive the hydraulic rod to provide a radial stress of up to 3KN111 to simulate the running load of a real train. Figure 2 The three vehicle-mounted AE sensors (VS 900-RIC) in (b) and (f) are arranged on the detection wheel at an angle of 120 degrees and fixed by a magnetic fixture. Figure 2 (c), (d) and (e). The rail has two artificial prefabricated damages, such as Figure 2As shown in (g), the AE data acquisition system (model: VallenAMSY-6) receives real-time signals from all three on-board sensors in continuous waveform mode and records them in the data matrix. When the inspection wheel moves, the multi-channel AE data will continuously capture the WRRN. When the inspection wheel passes the defect location, the damage signal generated is easily overwhelmed by the high-amplitude WRRN. Different AE signals can be obtained by setting different wheel start states. The rail AE data obtained by this experimental platform under a variety of initial setting conditions are closer to the actual inspection environment, which helps to verify the effectiveness and reliability of the method proposed in the present invention.

[0081] Execution step 1: In each wheel-rail rolling experiment, the wheel passes through two damage points in turn, which means that each set of multi-channel AE signals contains two damage signal segments. At the same time, the pre-start and braking parts of the signal are removed from the original signal to ensure authenticity and facilitate calculation. Each set of signals is divided into 5000 signal segments, each signal segment is 2048 sampling points long, and the total duration is 2.048 seconds. The multi-channel signal set is The complete time domain signal obtained is as follows Figure 3 shown.

[0082] Figure 3 The signal in the figure comes from an experiment where sensor No. 1 is at an angle of 30 degrees to the wheel-rail contact surface. The red dashed lines indicate the two moments of damage, which are located at 1.28s and 1.83s respectively. From the time domain waveform, in the initial stage, due to the low rolling speed, the signal amplitude fluctuates slightly. As the speed increases (after 0.4 seconds), the noise caused by mechanical vibration and friction gradually increases. The damage signal has been completely submerged by the noise, and the signal of each channel contains a large number of high-amplitude WRRN. Therefore, it is necessary to pre-process the multi-channel signal for denoising. Figure 4 The PFs waveform of a signal segment under the initial condition of 30 degrees obtained based on the ME-ALMD algorithm is shown. Compared with the original time domain waveform, it is found that the decomposition result has no obvious distortion or unnatural fluctuations at the beginning and end of the signal, which is obvious in the end segments of PF6 and PF7. The multi-channel data set is then reconstructed based on the best PF screened like Figure 5 As shown in the figure. The red dotted line is the random noise waveform, and the red solid line is the high-amplitude waveform, indicating that there are suspected damage signals or self-excited noise in the WRRN, and the random noise in the WRRN is mainly weakened. Figure 6 As shown in Figure 1, the best PF is filtered out according to the cross-correlation characteristics and used to reconstruct the multi-channel AE signal. The red dotted line is a random noise waveform, and the red solid line is a high-amplitude waveform, indicating that there is a suspected damaged signal or self-excited noise in the WRRN.

[0083] and Figure 5Compared with (a), (b) and (c), the random noise waveform after denoising becomes smoother. At the same time, from the perspective of the time domain integrated waveform, the reconstructed signal continues the ups and downs of the original signal. This result shows that the algorithm proposed in the present invention adaptively adopts a larger sliding step size in the noise part, which improves the denoising efficiency while maintaining the global characteristics of the signal. At the same time, in the high amplitude and high peak part, the peak waveform is well preserved, which shows that ME-ALMD automatically adopts a smaller step size for decomposition operation, retaining the dynamic details of the damaged signal. This feature greatly helps further damage detection work.

[0084] Execute step 2: Use the EDF-RCC algorithm to eliminate the remaining WRRN in the signal. Set the number of dictionary atoms for each channel to 6000, and obtain the initial dictionary D∈R based on the KSVD algorithm. 2048×6000 , the number of iterations is 100, then the reconstructed fusion signal obtained under the initial conditions of 0 degrees, 30 degrees, 60 degrees, and 90 degrees is as follows Figure 6 shown.

[0085] from Figure 6 It can be found that WRRN has been eliminated in the reconstructed fusion signal, and only a few high-amplitude interference signals still exist. Compared with the time domain waveform before fusion, although the signal energy is reduced, the high-amplitude damage signal is more obvious, which is helpful for accurate detection of damage. Figure 6 In (c) and (d), the fused signal retains the persistence of the damage information in the original signal.

[0086] Execute step 3: extract sample entropy for each column of reconstructed fusion signal and calculate its lower envelope to construct an adaptive threshold. For example, the sample entropy and lower envelope of the fusion signal at an angle of 30 degrees are as follows: Figure 7 As shown. Figure 6 and Figure 7 It can be found that the sampling entropy of the fused signal is different from its time domain waveform. After denoising, the waveform of the fused signal presents a smooth straight line in the noise part, and peaks only appear when there is high amplitude interference or damage. Although the sample entropy varies with the complexity of the signal itself, it drops rapidly when interference or damage signals appear. Therefore, the adaptive threshold constructed based on sample entropy can keep the threshold at high amplitude when passing noise to avoid false detection. When the damage signal appears, the threshold drops rapidly, thereby achieving accurate damage detection.

[0087] In order to further prove that the method proposed in this invention has the ability to accurately detect damage, Figure 8 The adaptive threshold and fused signal for each different initial experimental condition are shown, each containing two corrupted signals. Figure 8As can be seen in , the adaptive threshold identifies all corrupted signals and does not falsely detect interference. Figure 8 In (a) and (b), the fluctuation of the adaptive threshold is large. This is because the time domain signal contains more noise and interference, resulting in a large change in signal complexity. Figure 8 In (c) and (d), the adaptive threshold tends to be flat due to the less noise in the AE signal. Figure 8 The amplitude of the first damage signal in (c) is very close to the threshold. This is because the amplitude of this part of the AE wave is weak as a whole, resulting in a small fluctuation in the adaptive threshold. The results show that the method proposed in the present invention can accurately detect damage under background noise. Compared with before fusion, the average kurtosis of the signal increased by 25.67, and the structural similarity with the damage signal increased by 27.11%. The above experimental results show that the method proposed in the present invention can accurately detect damage signals submerged by background noise, and can effectively improve the characteristics of damage information, eliminate WRRN in multi-channel data, and provide guidance for the practical application of rail structure health monitoring based on acoustic emission.

Claims

1. A vehicle-mounted rail damage acoustic emission detection method based on dictionary fusion enhancement, characterized in that The method comprises the following steps: Step 1: Perform a mirror flip operation on the multi-channel AE signal dataset X, and build a mirror extension-multi-channel AE dataset X based on the original data e , improve the continuity of the signal segment endpoints to avoid endpoint effects and modal aliasing, and then use the short-term energy of the signal segment to adjust the adaptive smoothing step size S L , to extract the local envelope of the signal and solve the FM signal Finally, the PF with the largest mutual correlation coefficient is retained to reconstruct the multi-channel AE dataset To eliminate random noise, the specific steps are as follows: Step 1: For a multi-channel AE dataset Where Q is the number of channels, q is the channel number, M is the length of each column of signals, and N is the number of signal columns for each channel. is an integer domain, and the j-th column signal fragment in the q-th channel is Reverse along the time axis to generate mirror data symmetrical to the original signal in is the Mth element in the jth column signal in the qth channel; Step 1 and 2: Combine the generated mirror data with the original data to construct the mirror extension-AE signal in the jth column of the qth channel, which contains the original signal and its mirror image: in, represents the mirror extension-AE signal of the jth column in the qth channel, and then Constructing mirror extension-multi-channel AE dataset Step 13: Calculate the short-time energy of each signal segment, and construct an adaptive step size S that can be dynamically adjusted according to the size of the short-time energy. L : in, is the initial step length, η is a constant, It is a round-up operation, and η≥2 is the minimum iteration step size; Step 14: Extract X e The local mean α of each column signal in i′ and the local envelope γ i′ : a i′ =(s i′ +s i′+1 ) / 2 c i′ =|(σ i′ -s i′+1 ) / 2| Among them, σ i′ is the i′th extreme point of each column signal; Step 15: Based on the adaptive step size, the local mean function and the envelope estimation function are calculated respectively through the average sliding operation: Among them, j represents the column number, L s is the initial sliding step length, f α,i′j (t) represents the i′th element of the jth column of the local mean function at time t, f γ,i′j (t) represents the i′th element of the jth column of the envelope estimation function at time t; Step 16: Combine f α,i′j (t) and f γ,i′j (t) Construct frequency modulation signal S f,i′j (t): Among them, PF h is the product function generated by the hth iteration, x(t) represents the signal element at time t, and the iterative process of steps 15 and 16 is repeated until a pure FM signal is generated; Step 17: Determine the retained PF based on the cross-correlation coefficients between different channels and reconstruct the multi-channel data set in, represents the reconstructed signal set corresponding to the qth channel, and the formula is as follows: in, represents the PF generated by the h1th iteration in the q1 channel, R c represents the cross-correlation function, IME(·) is the inverse operation of the mirror extension; Step 2: For the reconstructed multi-channel AE dataset, the K-SVD algorithm is used to train a sub-dictionary Φ for the reconstructed signal of each channel q , then the atoms in each sub-dictionary are weighted according to Cramér's V coefficient and merged into a unified dictionary Φ f In the process of solving the sparse vector, the shared sparse coefficient B is solved by adding correlation constraints, and the ADMM algorithm is used to solve it to eliminate the residual WRRN, and finally reconstruct the fusion signal X re =Φ f Β, the specific steps are as follows: Step 21: Construct the following sparse representation for the reconstructed multi-channel dataset: Among them, Φ q is the sub-dictionary corresponding to the qth channel, is the jth column in the sparse coefficient matrix corresponding to the qth channel. In the dictionary training stage, the contribution of each atom to the same damage representation is measured by the degree of association between atoms at corresponding positions in each sub-dictionary. The sub-dictionary corresponding to each channel is trained using the K-SVD algorithm. Step 22: Introduce Cramér's V coefficient to measure the strength of the association between two random variables. The formula of Cramér's V coefficient is as follows: Among them, χ 2 represents the chi-square statistic, θ1 and θ2 represent the two sequences in the cross table, g and u represent the number of columns and rows in the cross table, respectively. In the process of dictionary fusion reconstruction of damage signals, the contribution of each atom based on Cramér's V coefficient is formulated as: in, for The k-th column signal in , represents the k-th column of the sub-dictionary atom, Indicates the sparse coefficient of the kth row corresponding to the qth sub-dictionary; Step 2 and 3: Weight each sub-dictionary atom by its atomic contribution to construct the fusion dictionary atom φ of the kth column f,k , the formula is as follows: Among them, the denominator term represents the normalized sum of contribution; Step 24: The optimization objective of the fusion dictionary is formulated as follows: Step 25: Using the common sparse coefficient matrix for sparse expression, based on the ADMM algorithm, the solution process of the joint sparse coefficients is as follows: Among them, the l2 norm enhances the simplicity of shared sparse coefficients, η′ is the regularization parameter, and R a (·) is the autocorrelation function, ρ is the penalty coefficient; Step 26: Based on the fusion dictionary and the shared sparse coefficient matrix, the reconstructed fusion signal is obtained: X re =Φ f B; Step 3: Based on the reconstructed fusion signal X re , for each column signal Extract the sample entropy SE, and then calculate the adaptive threshold Th based on the SE value of each column of the fused signal. Finally, find the signal segments in the fused signal with an amplitude higher than the adaptive threshold as damaged signals, and other signals lower than the adaptive threshold as non-damaged signals. The adaptive threshold Th is constructed by the following formula: Th=ωE lower (SE)+δ Where ω = 0.5, δ is the penalty coefficient, E lower (·) indicates the operation of extracting the lower envelope.

Images