Steel rail corrugation measurement method and device based on deep learning
By using multi-sensor fusion and deep learning algorithms, a benchmark measurement string is constructed for rail corrugation detection, which solves the problems of high cost, low efficiency and false judgment rate in existing technologies, and realizes high-precision and intelligent rail corrugation detection and prediction.
Patent Information
- Application Number
- CN202510980610.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-16
- Publication Date
- 2025-11-11
AI Technical Summary
Existing rail corrugation detection equipment is expensive, inefficient, and struggles to identify complex corrugation features, resulting in a high misjudgment rate. It also fails to achieve intelligent analysis and prediction, hindering efficient railway operation and maintenance.
By employing multi-sensor fusion and deep learning algorithms, a reference measurement string is constructed through a laser displacement sensor array. The signal is purified by combining wavelet transform and moving average filtering algorithms. One-dimensional convolutional neural networks and LSTM neural networks are used for feature extraction and prediction to achieve high-precision and intelligent rail condition detection.
It achieves high-precision and high-speed rail corrugation detection, improves the accuracy of complex corrugation feature recognition, can predict corrugation trends and provide intelligent maintenance suggestions, and reduces equipment costs and misjudgment rate.
Smart Images
Figure CN120922197A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of rail corrugation measurement technology, specifically to a rail corrugation measurement method and apparatus based on deep learning. Background Technology
[0002] While current rail corrugation detection methods can achieve high-precision measurements, they face three major technical bottlenecks: First, the cost of equipment procurement and maintenance is high; second, traditional detection methods are inefficient, with measurement speeds typically not exceeding 3 km / h, making it difficult to meet the demands of efficient operation and maintenance in modern railways; third, existing technologies are mostly based on fixed thresholds or empirical models, limiting their ability to identify complex corrugation characteristics (such as multi-wavelength coupling and non-uniform wear), and are easily affected by track welds and vibration noise, leading to a high misjudgment rate. Furthermore, traditional equipment lacks intelligent analysis capabilities, making it impossible to predict corrugation degradation trends and accurately classify them, severely hindering the development of preventative maintenance strategies. Summary of the Invention
[0003] The purpose of this invention is to provide a method and device for measuring rail corrugation based on deep learning. By using multi-sensor fusion and intelligent algorithms, it breaks through foreign technical barriers and achieves high-precision, high-efficiency, and intelligent rail condition detection and evaluation.
[0004] To achieve the above objectives, in a first aspect, the present invention provides a deep learning-based method for measuring rail corrugation, comprising: Step 1: Set up multiple sensors at equal intervals along the length of the rail, and collect multi-band corrugation signals of the rail through the sensors to obtain the rail corrugation wavelength distribution; Step 2: Based on the wavelength distribution of rail corrugation, optimize the spatial sampling density of the sensor; again, synchronously collect multi-band corrugation signals of the rail through the sensor, and output the multi-band corrugation data matrix of the rail; Step 3: Use wavelet transform combined with moving average filtering algorithm to process the multi-band warping data matrix, separate the real warping signal and interference signal from the multi-band warping signal, obtain the purified multi-band warping signal, and output the multi-band warping signal matrix. Step 4: Based on deep learning algorithms, extract the feature maps of each band from the multi-band corrugation signal matrix, and output the corrugation level and rail maintenance suggestions.
[0005] According to the deep learning-based rail corrugation measurement method provided by the present invention, the sensor is a laser displacement sensor; step 1 specifically includes: A measuring trolley that automatically travels along the rail is set up. The measuring trolley is equipped with multiple equally spaced arrayed laser displacement sensors, which form a reference measuring chord of fixed length. An adaptive leveling structure ensures the measurement stability of the laser displacement sensors. During the measurement process, each laser displacement sensor collects the vertical distance between its light spot and the rail surface in real time. By using the spatial constraint relationship of the fixed spacing between adjacent laser displacement sensors and the differential operation of multi-point measurement data, the rail surface contour sequence within the coverage area of the reference measuring chord is accurately reconstructed, and thus the rail corrugation wavelength distribution is obtained.
[0006] According to the deep learning-based rail corrugation measurement method provided by the present invention, the rail corrugation wavelength distribution ranges from 10mm to 3000mm.
[0007] According to the deep learning-based rail corrugation measurement method provided by the present invention, step 2 specifically includes: The wavelength distribution of the rail corrugation is divided into 5 bands: [10mm, 30mm), [30mm, 100mm), [100mm, 300mm), [300mm, 1000mm), [1000mm, 3000mm]; the sampling interval Δx between two adjacent laser displacement sensors satisfies: 20mm≤Δx≤1000mm, and ; In the formula, λ min This represents the lower limit value for the corresponding band.
[0008] According to the deep learning-based rail corrugation measurement method provided by this invention, the multi-band corrugation data matrix is as follows:
[0009] In the formula, N is the number of data points collected by the laser displacement sensor; each column in the multi-band wave-sloshing data matrix Y represents all the sampled data of the corresponding waveband, including signal amplitude, timestamp and spatial coordinates; each row in the multi-band wave-sloshing data matrix Y represents the synchronous measurement values of five wavebands at the same spatial location.
[0010] According to the deep learning-based rail corrugation measurement method provided by the present invention, step 3 specifically includes: The k-th band beating signal y k Wavelet decomposition is performed to obtain detail coefficients, high-frequency approximation coefficients, and low-frequency approximation coefficients; y k is the k-th column of the multi-band wave milling data matrix, k=1,2,...,5; Based on the dynamic threshold denoising algorithm, the denoising coefficient is obtained from the detail coefficient; Peak detection is performed on the high-frequency approximation coefficients to mark the interference locations; Based on the moving average filtering algorithm, the smoothing coefficient is obtained from the low-frequency approximation coefficient; Based on the denoising coefficient and smoothing coefficient, an inverse wavelet transform is performed to obtain the purified k-th band sprue signal. This outputs a five-band wave-scraping signal matrix. .
[0011] According to the deep learning-based rail corrugation measurement method provided by the present invention, step 4 specifically includes: The purified k-th band warping signal is processed using a parallel one-dimensional convolutional neural network. The feature map h of the k-th band is obtained. k ; The feature maps of all bands are weighted and fused using an attention mechanism to obtain the fused feature vector z; The fused feature vector z is input into the Gaussian mixture model, and the output is the distribution function of rail corrugation; Based on preset standards, the fused feature vector z is mapped to a ripple level; The five-band corrugation signal matrix is processed by an LSTM neural network to predict the future corrugation amplitude change of the rail. The system compares the corrugation level and the future change in rail corrugation amplitude with a preset threshold, and outputs rail maintenance instructions to provide rail maintenance recommendations.
[0012] According to the deep learning-based rail corrugation measurement method provided by the present invention, in step 1, the spacing of the laser displacement sensors is 200mm~300mm, and the length of the reference measurement chord is 1.2m~1.6m.
[0013] According to the deep learning-based rail corrugation measurement method provided by the present invention, in step 1, the measurement accuracy of the laser displacement sensor is ±0.01mm; each laser displacement sensor is calibrated by a laser tracker to ensure that the collinearity error of the laser displacement sensor spot center in the rail length direction is less than 0.05mm / m, and the perpendicularity deviation between the beam and the rail surface is within ±0.1°.
[0014] Secondly, the present invention provides a rail corrugation measurement device based on deep learning, comprising: The acquisition unit is used to set up multiple sensors at equal intervals along the length of the rail, and to acquire multi-band corrugation signals of the rail through the sensors to obtain the rail corrugation wavelength distribution. The secondary acquisition unit is used to optimize the spatial sampling density of the sensor based on the wavelength distribution of rail corrugation; it then synchronously acquires the multi-band corrugation signal of the rail again through the sensor and outputs the multi-band corrugation data matrix of the rail. The purification unit is used to process the multi-band warping data matrix using wavelet transform combined with moving average filtering algorithm, separate the real warping signal from the interference signal in the multi-band warping signal, obtain the purified multi-band warping signal, and output the multi-band warping signal matrix. The output unit is used to extract the feature maps of each band from the multi-band corrugation signal matrix based on deep learning algorithms, and output the corrugation level and rail maintenance suggestions.
[0015] This invention has at least the following technical effects: 1. Multi-scale 1D-CNN Intelligent Feature Extraction Architecture. This invention innovatively constructs a multi-scale one-dimensional convolutional neural network (1D-CNN) model for rail corrugation, achieving full-band feature capture through parallel convolutional branches. Differentiated convolutional kernels are designed for the wide wavelength range of 10-3000mm for corrugation signals, with kernel lengths strictly matched to the target wavelength. Hierarchical feature fusion technology is employed, weighting and fusing the high-dimensional feature maps output from each band through an attention mechanism, significantly improving the recognition accuracy of complex corrugations (such as harmonic superposition and local depressions).
[0016] 2. Dynamic Compensation Measurement Technology Based on the One-String, N-Point String Measurement Method. A rigid reference measurement string is constructed using a high-precision laser displacement sensor array. During the measurement process, each sensor in the array collects the vertical distance between its light spot and the rail surface in real time. Through the spatial constraint relationship of the fixed spacing between adjacent sensors and the differential calculation of multi-point measurement data, the rail surface profile sequence within the coverage area of the reference measurement string is accurately reconstructed, thus obtaining the rail corrugation wavelength distribution. This measurement method based on a fixed-length reference measurement string not only provides a stable spatial reference frame for corrugation detection across the 10-3000mm full-band, but also effectively suppresses the influence of low-frequency interference such as track unevenness on short-wavelength measurements, ensuring the accuracy of wavelength feature extraction and laying a reliable data foundation for subsequent multi-scale analysis and intelligent diagnosis. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0018] In the attached diagram: Figure 1 This is a flowchart of the rail corrugation measurement method based on deep learning according to the present invention; Figure 2 This is a structural block diagram of the rail corrugation measurement method based on deep learning according to the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0020] The following detailed description of some embodiments of the present invention will be provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0021] Please see Figure 1 and Figure 2 This invention provides a deep learning-based method for measuring rail corrugation, comprising the following steps: Step 1: Set multiple sensors at equal intervals along the length of the rail, and collect multi-band corrugation signals of the rail through the sensors to obtain the rail corrugation wavelength distribution; Specifically, a measuring trolley that automatically travels along the rail is set up. The measuring trolley is equipped with a high-precision laser displacement sensor array, which forms a reference measuring chord of fixed length. An adaptive leveling structure ensures the measurement stability of the laser displacement sensor and supports the acquisition of corrugation data across the entire wavelength range of 10mm-3000mm.
[0022] This invention constructs a fixed-length reference measurement chord using a precisely designed sensor array, which is the key foundation for achieving high-precision detection of rail corrugation. Specifically, 6-8 high-precision laser displacement sensors (measurement accuracy ±0.01mm) are rigidly mounted in a strictly equidistant (typically 200-300mm) linear array on an ultra-high rigidity reference beam made of Invar steel or carbon fiber composite material, forming a physical measurement reference with a fixed length of 1.2-1.6 meters. Each sensor is calibrated with a laser tracker to ensure that the collinearity error of its spot center along the length of the beam is less than 0.05mm / m, and the perpendicularity deviation of the beam to the rail surface is controlled within ±0.1°. During the measurement process, each sensor in the array collects the vertical distance between its light spot and the rail surface in real time. Through the spatial constraint relationship of the fixed spacing between adjacent sensors and the differential calculation of multi-point measurement data, the rail surface contour sequence within the coverage area of the reference measurement chord is accurately reconstructed, thereby obtaining the rail corrugation wavelength distribution. This measurement method based on a fixed-length reference chord not only provides a stable spatial reference frame for the detection of corrugation across the 10-3000mm wavelength range, but also effectively suppresses the influence of low-frequency interference such as track unevenness on shortwave measurements, ensuring the accuracy of wavelength feature extraction and laying a reliable data foundation for subsequent multi-scale analysis and intelligent diagnosis.
[0023] Step 2: Based on the wavelength distribution of rail corrugation, optimize the spatial sampling density of the sensor; again, synchronously collect multi-band corrugation signals of the rail through the sensor, and output the multi-band corrugation data matrix of the rail; Specifically, the wavelength distribution of rail corrugation is divided into five bands: [10mm, 30mm), [30mm, 100mm), [100mm, 300mm), [300mm, 1000mm), and [1000mm, 3000mm]. By employing anti-interference hardware design, multi-scale corrugation signals at 10-30mm, 30-100mm, 100-300mm, and 300-1000mm scales can be acquired simultaneously, ensuring data integrity.
[0024] Based on the track surface contour sequence y acquired by the N laser displacement sensors obtained in step 1 i Let i = 1, 2, ..., N, where the sensor spatial coordinates x i Satisfying non-uniform distribution: the sampling interval between two adjacent sensors is Δx = x i+1 -x i ∈[20,1000]mm. According to the Nyquist-Shannon sampling theorem, to avoid spectral aliasing, the sampling interval Δx must also satisfy:
[0025] in: This represents the lower limit of the corresponding band, for example, when in the 10-30mm band. ≤5mm.
[0026] In one specific embodiment, when the wave-polishing wavelength When the frequency bands are in the following five bands, the corresponding sampling intervals should be selected as follows:
[0027] The final output five-band wave-mold data matrix is as follows:
[0028] In the formula, N represents the number of data points collected by the laser displacement sensor; each column of the multi-band wave-erosion data matrix Y represents all sampled data for the corresponding waveband, including signal amplitude, timestamp, and spatial coordinates; each row of the multi-band wave-erosion data matrix Y represents the synchronous measurement values of five wavebands at the same spatial location. The number of wavebands is 5 (shortwave, medium-shortwave, medium-longwave, longwave, and ultra-longwave), corresponding to the five channels for subsequent multi-scale analysis.
[0029] Step 3: Process the multi-band warping data matrix using wavelet transform combined with moving average filtering algorithm to separate the real warping signal from the multi-band warping signal, obtain the purified multi-band warping signal, and output the multi-band warping signal matrix; this can improve the signal-to-noise ratio and provide clean data for subsequent analysis.
[0030] Specifically, the interference signals include track joints, welds, etc.
[0031] In some embodiments, step 3 specifically includes: Step 31: Transfer the k-th band beveling signal y k Perform wavelet decomposition to obtain detail coefficients. High-frequency approximation coefficients (j=1 or 2) and low-frequency approximation coefficients (j≥3); y k The k-th column of the multi-band wave erosion data matrix, k=1,2,...,5, can separate different wavelength components from noise.
[0032] Specifically, a 5-level discrete wavelet transform (DWT) is performed using the Db4 wavelet basis:
[0033] Let the wavelet basis functions be scale j and translation n; represents the detail coefficients (high-frequency components) at scale j. These are approximate coefficients (for low-frequency components). The physical wavelength range corresponding to each layer: .
[0034] Step 32: Based on the dynamic threshold denoising algorithm, the denoising coefficient is obtained according to the detail coefficient; this can preserve effective ripple features and suppress noise.
[0035] Among them, the threshold is optimized based on Stein's unbiased risk estimation (SURE):
[0036] In the formula, Soft threshold function ; The noise variance of the j-th layer is estimated using Median Absolute Deviation:
[0037] The divergence of the threshold function; It is an L2 norm; This is the median operator.
[0038] Step 33: Perform peak detection on the high-frequency approximation coefficients and mark the interference locations; avoid misjudging interference locations such as weld seams as corrugations.
[0039] Specifically, interference suppression at track welds: Weld seam signals are characterized by transient impacts, and their feature extraction employs the following methods:
[0040] This is an indicator function (1 if the condition is met, 0 otherwise).
[0041] Impact location detection:
[0042] In the formula N peak This represents the number of impact pulses.
[0043] Step 34: Based on the moving average filtering algorithm, obtain the smoothing coefficient according to the low-frequency approximation coefficient; eliminate macroscopic irregularities in the track.
[0044] Applying a moving average to the low-frequency band (j≥3) suppresses vibration noise:
[0045] in: For window length ( (The wavelength corresponding to the current scale).
[0046] Step 35: Based on the denoising coefficient and smoothing coefficient, perform inverse wavelet transform to obtain the purified k-th band sprue signal. This outputs a five-band wave-scraping signal matrix. .
[0047] Specifically, the denoised signal is reconstructed using inverse wavelet transform (IDWT):
[0048] Reconstruction error is controlled within:
[0049] The scaling function is at scale 5; This is the reconstructed k-th band signal.
[0050] In some embodiments, compensation processing is also required for ultra-long wavelengths to improve the accuracy of the 1000-3000mm band.
[0051] For the 1000-3000mm band, an improved moving average is used:
[0052] In the formula: L=round(λ / 2Δx) is the adaptive window length, and round is the rounding function; α = 0.01 is the Gaussian weighting coefficient; The original ultra-long wave signal after IMU (Inertial Measurement Unit) compensation.
[0053] The final output is a purified five-band signal matrix: .
[0054] Step 4: Based on deep learning algorithms, extract feature maps of each band from the multi-band corrugation signal matrix, and output corrugation level and rail maintenance recommendations. This enables intelligent rail condition diagnosis and maintenance decision-making.
[0055] In some embodiments, step 4 specifically includes: Step 41: Process the purified k-th band wave erosion signal using a parallel one-dimensional convolutional neural network. The feature map h of the k-th band is obtained. k It can extract wavelength-related local / global features.
[0056] Specifically, each band undergoes independent convolution operations:
[0057] In the formula, ∈R L×C , representing the l-th layer convolution kernel of the k-th band (L=λ) k / Δx); For bias terms; This represents a one-dimensional convolution operation; C is the number of feature map channels.
[0058] For the shortwave branch, L1 = 10mm / 5mm = 2 (core size); Ultra-long wavelength branch L5 = 3000mm / 300mm = 10 (nuclear size).
[0059] Step 42: Weighted fusion of all band feature maps using an attention mechanism to obtain a fused feature vector z; this can enhance the contribution of key bands.
[0060] Specifically, weighted fusion is achieved through an attention mechanism:
[0061] The global features are represented as:
[0062] For multilayer perceptron computation, denoted as the attention weight for the k-th band.
[0063] Step 43: Input the fused feature vector z into the Gaussian mixture model and output the distribution function p(λ) of rail corrugation; the area of concentrated corrugation energy can be visualized.
[0064] Specifically, a Gaussian mixture model is used:
[0065] Parameter estimation:
[0066] Let be the mixing coefficient of the m-th Gaussian component. This represents the Gaussian distribution function.
[0067] Step 44: Based on preset standards, map the fused feature vector z to the corrugation level; this can guide the maintenance decisions of the rail (such as grinding / replacement).
[0068] Step 45: Process the five-band corrugation signal matrix using an LSTM (Long Short-Term Memory) neural network to predict the future change in rail corrugation amplitude; this can predict the rate of corrugation development.
[0069] Step 46: Compare the corrugation level and the future change in rail corrugation amplitude with preset thresholds, and output rail maintenance instructions to provide rail maintenance recommendations. Maintenance instructions may include emergency repairs, preventative grinding, continued monitoring, and then output specific maintenance plans.
[0070] The degradation trend prediction includes: LSTM neural network time series modeling and hidden state update.
[0071] The predicted output is:
[0072] In the formula Δ A t+30 This represents the change in amplitude after 30 days.
[0073] Output for the forget gate. It is the sigmoid activation function.
[0074] Lifespan prediction models can be based on Paris's Law:
[0075] The remaining lifespan is:
[0076] ΔK represents the crack propagation rate, and ΔK represents the stress intensity factor amplitude.
[0077] The final output includes: Wavelength-amplitude distribution diagram: {(λ i A i )}, i=1, 2, 3……N.
[0078] Wave grinding level:
[0079] Maintenance recommendations:
[0080] The weighting coefficient of the k-th band (∑w) k =1).
[0081] Based on the same inventive concept, another embodiment of the present invention provides a rail corrugation measurement device based on deep learning. This device corresponds to the method of the foregoing embodiment and includes: The acquisition unit is used to set up multiple sensors at equal intervals along the length of the rail, and to acquire multi-band corrugation signals of the rail through the sensors to obtain the rail corrugation wavelength distribution. The secondary acquisition unit is used to optimize the spatial sampling density of the sensor based on the wavelength distribution of rail corrugation; it then synchronously acquires the multi-band corrugation signal of the rail again through the sensor and outputs the multi-band corrugation data matrix of the rail. The purification unit is used to process the multi-band warping data matrix using wavelet transform combined with moving average filtering algorithm, separate the real warping signal from the interference signal in the multi-band warping signal, obtain the purified multi-band warping signal, and output the multi-band warping signal matrix. The output unit is used to extract the feature maps of each band from the multi-band corrugation signal matrix based on deep learning algorithms, and output the corrugation level and rail maintenance suggestions.
[0082] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. It should be understood that the invention is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.
Claims
1. A deep learning-based method for measuring rail corrugation, characterized in that, include: Step 1: Set multiple sensors at equal intervals along the length of the rail, and collect multi-band corrugation signals of the rail through the sensors to obtain the rail corrugation wavelength distribution; Step 2: Based on the wavelength distribution of rail corrugation, optimize the spatial sampling density of the sensor; again, synchronously acquire the multi-band corrugation signal of the rail through the sensor, and output the multi-band corrugation data matrix of the rail; Step 3: Process the multi-band warping data matrix using wavelet transform combined with moving average filtering algorithm, separate the real warping signal and interference signal from the multi-band warping signal, obtain the purified multi-band warping signal, and output the multi-band warping signal matrix. Step 4: Based on deep learning algorithms, extract the feature maps of each band from the multi-band corrugation signal matrix, and output the corrugation level and rail maintenance suggestions.
2. The rail corrugation measurement method based on deep learning according to claim 1, characterized in that, The sensor is a laser displacement sensor; step 1 specifically includes: A measuring trolley that automatically travels along the rail is set up. The measuring trolley is equipped with multiple equally spaced arrayed laser displacement sensors, which form a reference measuring chord of fixed length. An adaptive leveling structure ensures the measurement stability of the laser displacement sensors. During the measurement process, each laser displacement sensor collects the vertical distance between its light spot and the rail surface in real time. By using the spatial constraint relationship of the fixed spacing between adjacent laser displacement sensors and the differential operation of multi-point measurement data, the rail surface contour sequence within the coverage area of the reference measuring chord is accurately reconstructed, and thus the rail corrugation wavelength distribution is obtained.
3. The deep learning-based rail corrugation measurement method according to claim 2, characterized in that, The wavelength distribution of the rail corrugation ranges from 10mm to 3000mm.
4. The rail corrugation measurement method based on deep learning according to claim 3, characterized in that, Step 2 specifically includes: The wavelength distribution of the rail corrugation is divided into 5 bands: [10mm, 30mm), [30mm, 100mm), [100mm, 300mm), [300mm, 1000mm), [1000mm, 3000mm]; the sampling interval Δx between two adjacent laser displacement sensors satisfies: 20mm≤Δx≤1000mm, and ; In the formula, λ min This represents the lower limit value for the corresponding band.
5. The deep learning-based rail corrugation measurement method according to claim 4, characterized in that, The multi-band wave milling data matrix is as follows: In the formula, N is the number of data points collected by the laser displacement sensor; each column in the multi-band wave-sloshing data matrix Y represents all the sampled data of the corresponding waveband, including signal amplitude, timestamp and spatial coordinates; each row in the multi-band wave-sloshing data matrix Y represents the synchronous measurement values of five wavebands at the same spatial location.
6. The deep learning-based rail corrugation measurement method according to claim 5, characterized in that, Step 3 specifically includes: The k-th band beating signal y k Wavelet decomposition is performed to obtain detail coefficients, high-frequency approximation coefficients, and low-frequency approximation coefficients; the y k is the k-th column of the multi-band wave milling data matrix, k=1,2,...,5; Based on the dynamic threshold denoising algorithm, the denoising coefficients are obtained according to the detail coefficients; Peak detection is performed on the high-frequency approximation coefficients to mark the interference locations; Based on the moving average filtering algorithm, the smoothing coefficient is obtained according to the low-frequency approximation coefficient; Based on the denoising coefficients and smoothing coefficients, an inverse wavelet transform is performed to obtain the purified k-th band sprue signal. This outputs a five-band wave-scraping signal matrix. .
7. The deep learning-based rail corrugation measurement method according to claim 6, characterized in that, Step 4 specifically includes: The purified k-th band warping signal is processed using a parallel one-dimensional convolutional neural network. The feature map h of the k-th band is obtained. k ; The feature maps of all bands are weighted and fused using an attention mechanism to obtain the fused feature vector z; The fused feature vector z is input into the Gaussian mixture model, and the output is the distribution function of rail corrugation; Based on preset standards, the fused feature vector z is mapped to a ripple level; The five-band corrugation signal matrix is processed by an LSTM neural network to predict the future change in the corrugation amplitude of the rail. The system compares the corrugation level and the future change in rail corrugation amplitude with a preset threshold, and outputs rail maintenance instructions to provide rail maintenance recommendations.
8. The deep learning-based rail corrugation measurement method according to claim 2, characterized in that, In step 1, the spacing between the laser displacement sensors is 200mm~300mm, and the length of the reference measuring string is 1.2m~1.6m.
9. The deep learning-based rail corrugation measurement method according to claim 2, characterized in that, In step 1, the measurement accuracy of the laser displacement sensor is ±0.01mm; each laser displacement sensor is calibrated by a laser tracker to ensure that the collinearity error of the laser spot center in the rail length direction is less than 0.05mm / m, and the perpendicularity deviation between the beam and the rail surface is within ±0.1°.
10. A rail corrugation measuring device based on deep learning, characterized in that, include: The acquisition unit is used to set multiple sensors at equal intervals along the length of the rail, and to acquire multi-band corrugation signals of the rail through the sensors to obtain the rail corrugation wavelength distribution. The secondary acquisition unit is used to optimize the spatial sampling density of the sensor based on the wavelength distribution of rail corrugation. The multi-band corrugation signal of the rail is collected again through the sensor, and the multi-band corrugation data matrix of the rail is output. The purification unit is used to process the multi-band warping data matrix using wavelet transform combined with moving average filtering algorithm, separate the real warping signal and the interference signal from the multi-band warping signal, obtain the purified multi-band warping signal, and output the multi-band warping signal matrix. The output unit is used to extract the feature maps of each band from the multi-band corrugation signal matrix based on deep learning algorithms, and output the corrugation level and rail maintenance suggestions.