Steel rail damage Mel spectrum feature detection method based on dynamic weight fusion driving

Abstract

The invention discloses a steel rail damage Mel spectrum feature detection method based on dynamic weight fusion driving, which comprises the following steps of: firstly, calculating a signal consistency variance of an acquired acoustic emission signal to determine a numerical value fusion weight; then, calculating a position reliability function of the sensor according to the actual installation position of the sensor, and calculating a position weight by combining a dynamic compensation mechanism; then, combining the obtained numerical value weight with the position weight to obtain a self-adaptive weight for signal fusion so as to carry out signal fusion; and finally, Mel spectrum roll-off point features are extracted from the fusion signal, and whether the steel rail is damaged or not is determined in combination with a statistical threshold. According to the method, the fusion weight is dynamically adjusted according to the signal consistency variance and the geometric position of the sensor, sensor data under different conditions can be more flexibly adapted, the Mel spectrum roll-off point is introduced as a feature extraction index, noise interference is effectively removed, damage information is reserved, and the method is suitable for the field of steel rail damage detection needing high precision.

Description

Technical Field

[0001] The present invention relates to a method for processing and detecting acoustic emission signals of rail damage, and in particular to a method for detecting Mel spectrum features of rail damage based on dynamic weight fusion drive. Background Art

[0002] As an important part of modern transportation, high-speed railways have significant advantages such as high speed, comfort and environmental protection. The condition assessment of rails is crucial to the operation quality and safety of the entire high-speed railway network. Due to the frequent operation cycles and huge load requirements, rails are particularly prone to damage, which not only affects the operation efficiency but also brings major safety hazards. Therefore, rail damage detection is crucial to ensure safety, improve transportation efficiency and reduce maintenance costs.

[0003] Acoustic emission technology is a nondestructive testing method based on monitoring and analyzing elastic waves generated when stress concentration or damage propagation occurs inside the material. Due to its advantages such as real-time monitoring, high sensitivity, and nondestructive testing, it has been widely used in rail damage detection in recent years. However, most current research relies on single sensor signal analysis. Such signals have a limited range, which can lead to problems such as low fault tolerance and poor stability, affecting the accuracy and reliability of detection. Therefore, it is urgent to explore multi-sensor data fusion methods to enhance signal confidence in order to support and improve the application of acoustic emission technology in track damage detection.

[0004] Multi-sensor data fusion is the process of integrating, processing and analyzing data from multiple sensors to generate more accurate, comprehensive and reliable information. According to the degree of data processing in the fusion process, it can be divided into three levels: data-level fusion, feature-level fusion and decision-level fusion. Data-level fusion is the direct integration and processing of raw data from different sensors; feature-level fusion is the integration of data features of each sensor; decision-level fusion refers to the feature extraction and preliminary diagnosis of the data collected by each sensor, and then the preliminary diagnosis results of each sensor are fused to make decisions. Although data-level fusion may reduce the anti-interference ability of the system, it can retain more details and reduce the dependence on data quality, which is very suitable for rail damage detection tasks. Among the data-level fusion algorithms, Bayesian estimation, correlation function fusion algorithm and weighted coefficient fusion algorithm are more common. However, Bayesian estimation relies on predefined probability models and prior knowledge, which are often difficult to obtain accurately in practical applications; the correlation function fusion algorithm mainly realizes data fusion by calculating correlation, but ignores the influence of sensor geometric position on signal acquisition ability. The weighted coefficient fusion algorithm considers a single perspective, which may lead to poor performance. Therefore, it is of extraordinary significance to propose a multi-sensor data fusion algorithm that does not require prior knowledge and probability models and can fully reflect the importance of sensors in different situations. In addition, since the nonlinear characteristics of acoustic emission signals and the large amount of noise interference contained in the collected signals seriously affect the subsequent judgment of the damage state of the rail, it is also of great significance to propose a feature that can characterize damage-related information under the background of noise. Summary of the invention

[0005] In order to overcome the problems that the existing fusion algorithm relies on prior information, cannot fully reflect the importance of different sensors, and the acoustic emission signal is difficult to characterize the damage state, the present invention provides a rail damage Mel spectrum feature detection method based on dynamic weight fusion drive. This method dynamically adjusts the fusion weight according to the signal consistency variance and the sensor geometric position, can more flexibly adapt to sensor data in different situations, and introduces the Mel spectrum roll-off point as a feature extraction indicator to effectively remove noise interference and retain damage information. It is suitable for the field of rail damage detection that requires high precision.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A rail damage mel spectrum feature detection method based on dynamic weight fusion drive includes the following steps:

[0008] Step 1: Load the acoustic emission signal y obtained from the i-th sensor in the vehicle-mounted rail damage detection experiment i , calculate the signal consistency variance to obtain the numerical weight w ni, and so on, calculate the numerical weight w corresponding to I sensors n ;

[0009] Step 2: Load the i-th sensor angle information θ obtained from the rotary encoder and laser i , calculate the position reliability f(θ i ) and combined with the dynamic compensation mechanism to obtain the sensor position weight w pi , and so on, calculate the position weight w corresponding to I sensors p , combine it with the numerical weight and normalize it to obtain the adaptive weight W, and then obtain the fusion signal fy;

[0010] Step 3: Extract the Mel spectrum roll-off point feature MRP of the fusion signal fy, calculate the upper and lower detection thresholds LT and UT to determine the damage state of the rail.

[0011] Compared with the prior art, the present invention has the following advantages:

[0012] 1. The present invention dynamically adjusts the fusion weight according to the signal consistency variance and the sensor geometric position, does not rely on a predefined probability model or fixed weight, can more flexibly adapt to sensor data in different situations, and enhances the robustness and reliability of the system.

[0013] 2. The present invention introduces the Mel spectrum roll-off point as a feature extraction indicator, which can filter out a large amount of noise and other interference information in the signal while retaining key damage information. This indicator plays a key role in detection and improves the accuracy of detection.

[0014] 3. The present invention solves the problems that the current rail damage detection method relies on prior knowledge and probability models, the fusion weights are not fully considered, and the detection accuracy is affected by noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 The present invention is a flow chart of the rail damage Mel spectrum feature detection method based on dynamic weight fusion drive.

[0016] Figure 2 It is the amplitude diagram of multi-sensor signals.

[0017] Figure 3 This is a comparison chart of location reliability and location weight.

[0018] Figure 4 It is the amplitude diagram of the fused signal. DETAILED DESCRIPTION

[0019] 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.

[0020] The present invention provides a method for detecting rail damage Mel spectrum features based on dynamic weight fusion drive. First, for the collected acoustic emission signal, the signal consistency variance is calculated to determine the numerical fusion weight; then, the position reliability function of the sensor is calculated according to the actual installation position, and the position weight is calculated in combination with the dynamic compensation mechanism; then, the numerical weight and position weight obtained by the above steps are combined to obtain an adaptive weight for signal fusion to perform signal fusion; finally, the Mel spectrum roll-off point feature is extracted from the fused signal, and combined with the statistical threshold to determine whether the rail is damaged.

[0021] like Figure 1 As shown, the specific steps are as follows:

[0022] Step 1: Load the acoustic emission signal y obtained from the i-th sensor in the vehicle-mounted rail damage detection experiment i , calculate the signal consistency variance to obtain the numerical weight w ni , and so on, calculate the numerical weight w corresponding to I sensors n , the specific steps are as follows:

[0023] Step 1: Load the acoustic emission signal obtained from the i-th sensor in the vehicle-mounted rail damage detection experiment in represents the signal collected by the i-th sensor at the m-th sampling point, and each group of signals contains M sampling points;

[0024] Step 1 and 2: Calculate the signal consistency variance SCV obtained by the i-th sensor i :

[0025]

[0026] Step 13: Calculate the numerical weight w of the i-th sensor ni :

[0027]

[0028] Step 14: Calculate the numerical weight w corresponding to I sensors n :

[0029]

[0030] Among them, SCV I The signal consistency variance obtained by the Ith sensor.

[0031] Step 2: Load the i-th sensor angle information θ obtained from the rotary encoder and laser i , calculate the position reliability f(θ i ) and combined with the dynamic compensation mechanism to obtain the sensor position weight w pi , and so on, calculate the position weight w corresponding to I sensors p , combine it with the numerical weight and normalize it to obtain the adaptive weight W, and then obtain the fusion signal fy. The specific steps are as follows:

[0032] Step 21: Load angle information θ i , calculate the position reliability of the i-th sensor:

[0033]

[0034] Step 22: Combine the dynamic compensation mechanism to obtain the position weight w of the i-th sensor pi :

[0035]

[0036] Where α is the protection constant;

[0037] Step 23: Calculate the position weight w corresponding to I sensors p :

[0038] w p =[w p1 ,w p2 ,…,w pI ];

[0039] Among them, w pI is the position weight of the I-th sensor;

[0040] Step 24: Combine the numerical weight and the position weight and normalize them to obtain the sensor's adaptive fusion weight W:

[0041]

[0042] W=[W1,W2,…,W I ];

[0043] Where ⊙ is the Hadamard product, W I is the adaptive fusion weight of the I-th sensor;

[0044] Step 25: Fusion of multi-sensor signals based on adaptive weights to obtain fused signal fy:

[0045] fy=W1×y1+W2×y2+…+W I ×yI .

[0046] Step 3: Extract the Mel spectrum roll-off point feature MRP of the fusion signal fy, calculate the upper and lower detection thresholds LT and UT to determine the damage state of the rail. The specific steps are as follows:

[0047] Step 31: Convert the fused signal fy into a Mel spectrum and calculate the roll-off point of the spectrum to obtain the Mel spectrum feature MRP of the signal:

[0048]

[0049] Where L is the number of Mel filter banks, S[K] is the energy of the Kth band in the Mel spectrum, β is the threshold ratio of the roll-off point, and l ro is the index of the frequency band corresponding to the roll-off point, u[l ro ] is the first ro The center frequency of the Mel filter bank;

[0050] Step 32: Calculate the upper and lower detection thresholds LT and UT of the damage based on the concept of outlier statistics:

[0051]

[0052] Among them, LT and UT are the lower and upper detection thresholds, respectively, μ and σ are the mean and standard deviation of MRP, respectively, and k1 and k2 are detection threshold constants;

[0053] Step 33: Determine the damage status of the rail based on the Mel spectrum roll-off point feature MRP and the upper and lower detection thresholds LT and UT:

[0054]

[0055] in, Represents the damage state of the rail corresponding to the eigenvalue MRP. Undamaged means the rail is in a non-damaged state, and damaged means the rail is in a damaged state.

[0056] Example:

[0057] This embodiment illustrates the specific implementation of the present invention in combination with the acoustic emission signals collected in the vehicle-mounted rail damage detection experiment:

[0058] Execution step 1: In the vehicle-mounted rail damage detection experiment, there are 3 sensors, that is, I = 3. The signal amplitude diagram obtained is as follows: Figure 2 As shown. Set the number of sampling points in each group of signals to 4096, that is, M = 4096. Calculate the signal consistency variance SCV of the first group of signals of the three sensors, respectively [3.0474×10 -5,3.9064×10 -5 ,4.6899×10 -5 ], and then calculate the numerical weights w of the three sensors n is [3281.49, 2559.90, 2132.24]. And so on, calculate the numerical weight corresponding to each group of signals.

[0059] Execute step 2: Obtain the angle information of the three sensors corresponding to each group of signals according to the rotary encoder and the laser. When collecting the first group of signals, the angles θ of the three sensors are: [163.8281°, 283.8281°, 43.8281°], and calculate the position reliability f(θ) of each sensor as: [0.0198, 0.6195, 0.8607]; take the protection constant α = 0.1, and combine the dynamic compensation mechanism to obtain the position weight w of each sensor when collecting the first group of signals. p is: [0.1678, 0.7076, 0.9246]. Similarly, the sensor position reliability and position weight corresponding to each group of signals are calculated. The comparison curve is as follows Figure 3 As shown; combined with the numerical weight w n With position weight w p We get w = [5.5066 × 10 3 ,1.8113×10 4 ,1.9715×10 4 ], and then normalize to get the adaptive weight W=[0.1271, 0.4180, 0.4550] corresponding to the first group of signals collected by the three sensors, and calculate the adaptive weight corresponding to each group of signals by analogy; fuse the first group of signals according to the adaptive weight W to get the fused signal fy, and calculate the fused signal corresponding to each group of signals by analogy. The specific fusion results are as follows: Figure 4 shown.

[0060] Execute step 3: Extract the Mel spectrum roll-off point feature MRP of the fusion signal and calculate its mean value as μ = 9.6198 × 10 5 , the standard deviation is σ=1.7768×10 5 . Taking k1=1.4, k2=2.0, the upper and lower detection thresholds are LT=7.1323×10 5 ,UT=1.3173×10 6 The damage state of the rail is judged based on the Mel spectrum roll-off point feature MRP and the upper and lower detection thresholds LT and UT. The final rail damage detection result is obtained with an accuracy rate of 83.91%.

Claims

1. A rail damage mel spectrum feature detection method based on dynamic weight fusion drive, characterized in that The method comprises the following steps: Step 1: Load the acoustic emission signal y obtained from the i-th sensor in the vehicle-mounted rail damage detection experiment i , calculate the signal consistency variance to obtain the numerical weight w ni , and so on, calculate the numerical weight w corresponding to I sensors n ; Step 2: Load the i-th sensor angle information θ obtained from the rotary encoder and laser i , calculate the position reliability f(θ i ) and combined with the dynamic compensation mechanism to obtain the sensor position weight w pi , and so on, calculate the position weight w corresponding to I sensors p , combine it with the numerical weight and normalize it to obtain the adaptive weight W, and then obtain the fusion signal fy; Step 3: Extract the Mel spectrum roll-off point feature MRP of the fusion signal fy, calculate the upper and lower detection thresholds LT and UT to determine the damage state of the rail.

2. The method for detecting rail damage Mel spectrum features based on dynamic weight fusion drive according to claim 1 is characterized in that The specific steps of step one are as follows: Step 1: Load the acoustic emission signal obtained from the i-th sensor in the vehicle-mounted rail damage detection experiment in represents the signal collected by the i-th sensor at the m-th sampling point, and each group of signals contains M sampling points; Step 1 and 2: Calculate the signal consistency variance SCV obtained by the i-th sensor i : Step 13: Calculate the numerical weight w of the i-th sensor ni : Step 14: Calculate the numerical weight w corresponding to I sensors n : Among them, SCV I The signal consistency variance obtained by the I-th sensor.

3. The rail damage Mel spectrum feature detection method based on dynamic weight fusion drive according to claim 1 is characterized in that The specific steps of step 2 are as follows: Step 21: Load angle information θ i , calculate the position reliability of the i-th sensor: Step 22: Combine the dynamic compensation mechanism to obtain the position weight w of the i-th sensor pi : Where α is the protection constant; Step 23: Calculate the position weight w corresponding to I sensors p : In p =[in p1 ,In p2 ,…,In pI ]; Among them, w pI is the position weight of the I-th sensor; Step 24: Combine the numerical weight and the position weight and normalize them to obtain the sensor's adaptive fusion weight W: In=[In1,In2,…,In I ]; Where ⊙ is the Hadamard product, W I is the adaptive fusion weight of the I-th sensor; Step 25: Fusion of multi-sensor signals based on adaptive weights to obtain fused signal fy: fy=W1×y1+W2×y2+…+W I ×y I 。 4. The method for detecting rail damage Mel spectrum features based on dynamic weight fusion drive according to claim 1 is characterized in that The specific steps of step three are as follows: Step 31: Convert the fused signal fy into a Mel spectrum and calculate the roll-off point of the spectrum to obtain the Mel spectrum feature MRP of the signal: Where L is the number of Mel filter banks, S[K] is the energy of the Kth band in the Mel spectrum, β is the threshold ratio of the roll-off point, and l ro is the index of the frequency band corresponding to the roll-off point, u[l ro ] is the first ro The center frequency of the Mel filter bank; Step 32: Calculate the upper and lower detection thresholds LT and UT of the damage based on the concept of outlier statistics: Among them, LT and UT are the lower and upper detection thresholds, respectively, μ and σ are the mean and standard deviation of MRP, respectively, and k1 and k2 are detection threshold constants; Step 33: Determine the damage status of the rail based on the Mel spectrum roll-off point feature MRP and the upper and lower detection thresholds LT and UT: in, Represents the damage state of the rail corresponding to the eigenvalue MRP. Undamaged means the rail is in a non-damaged state, and damaged means the rail is in a damaged state.

Images