A Wheel Abrasion Detection Method Based on Improved Mapping Fusion Features and Thresholds

Abstract

The present invention discloses a wheel abrasion detection method based on improved mapping fusion features and thresholds, and the method is as follows: First, load P groups of wheel acoustic emission signals, extract seven features from each group of signals, determine the extreme values of each group of features, and calculate the average value of the extreme values of the P groups of features; Second, use the obtained average value as the detection threshold to preliminarily detect the acoustic emission signals, calculate the detection rate and precision rate of the abrasion signals of each feature, perform feature screening on the multi-dimensional features to obtain a low-dimensional feature parameter set, and map and fuse the low-dimensional feature parameter set to obtain a one-dimensional fusion feature; Third, calculate the standard deviation, mean value and minimum value of the mapped fusion feature, calculate the improved adaptive detection threshold within each time window, and determine whether each feature belongs to the abrasion signal according to the improved adaptive threshold discrimination algorithm, so as to complete the detection of the wheel abrasion acoustic emission signal. The present invention can greatly improve the detection rate and detection accuracy of wheel abrasion signals, and meet the requirements for ensuring the safe operation of trains.

Description

Technical Field

[0001] The present invention relates to a method for processing and detecting wheel abrasion signals, and more particularly to a method for detecting wheel abrasion based on improved mapping fusion features and thresholds. Background Art

[0002] Railways are important national transportation facilities and the main artery of the country's economic development, which are closely related to the production and life of the people. In the 1860s, the first railway in China, the Wusong Railway, emerged. So far, the Chinese railway has a development history of more than 160 years. Since the 21st century, with the rapid development of materials science and manufacturing technology, the Chinese railway industry has achieved leapfrog development. It is expected that by the end of 2021, the total operating mileage of the Chinese railway will exceed 1.46 million kilometers, ranking first in the world. The "eight vertical and eight horizontal" layout of the Chinese railway covers all parts of the country, covering about 99% of the cities with a population of over 200,000, connecting the north-south economy and facilitating the development of the east-west regions, providing sufficient transport capacity for the rapid development of the Chinese economy. However, the wheels of trains will experience varying degrees of abrasion due to long-term operation, resulting in emergency repairs or potential accidents, causing delays, unnecessary costs, and even casualties. Therefore, in order to ensure the safe, stable, and high-speed operation of railways, real-time monitoring of wheels, research on new wheel abrasion detection methods, and improvement of detection accuracy have profound practical significance.

[0003] In the existing wheel abrasion detection methods, methods such as magnetic flux leakage detection, ultrasonic detection, vibration detection, laser and high-speed camera detection are widely used. However, for the characteristics of fast running speed and large wheel-rail friction noise of train wheels, the above detection methods have many deficiencies. The magnetic flux leakage detection technology is limited by the existing mechanism and can only detect surface abrasions of wheels but cannot detect internal conditions, and the detection result accuracy is low; the ultrasonic detection technology relies on the ultrasonic waves emitted by the instrument and is likely to cause secondary abrasions to the wheels to be detected, affecting the detection accuracy; the vibration detection technology is easily interfered by external factors, and there are many interferences in the obtained vibration signals, seriously affecting subsequent detection and analysis; the laser and high-speed camera detection technology is mainly applied to the laboratory environment because its equipment is expensive and the installation is complex, and it is not suitable for actual train wheel abrasion detection. Compared with the above detection technologies, acoustic emission is a non-destructive detection technology, and the detected energy comes from the wheel to be detected itself rather than being provided by the detection instrument, and it can more accurately provide information on the actual condition of the wheel to be detected. Therefore, the acoustic emission detection technology has also been applied to the field of wheel abrasion detection in recent years. However, due to the large amount of noise generated by the intense friction between the wheels and the rails during train operation, the detection result accuracy is affected.

[0004] Most traditional acoustic emission scratch detection algorithms only use a single feature to describe part of the scratch information, and cannot accurately represent all the information of the wheel scratch, thus affecting the detection rate of the scratch signal. At the same time, traditional detection algorithms judge whether a wheel scratch occurs based on a fixed adaptive threshold, that is, the same threshold is used to judge the scratch for each group of acoustic emission signals. The detection results brought by this threshold detection method often contain a large number of noise signals, seriously reducing the detection accuracy and unable to accurately ensure the safe and stable operation of the train. Summary of the Invention

[0005] To solve the problems of slow speed and low accuracy of the existing scratch signal detection methods, the present invention provides a wheel scratch detection method based on improved mapping fusion features and thresholds. This method is based on the processing of wheel rolling acoustic emission signals, extracts multi-dimensional feature information, characterizes the wheel scratch information from multiple angles, and at the same time uses time information to improve the traditional adaptive threshold, and proposes a multi-feature mapping fusion and improved adaptive threshold detection algorithm. It makes full use of the time-frequency domain feature information and time information of the samples, screens the feature information according to the detection rate and detection accuracy, further reduces the operation time and effectively reduces the interference of noise signals, accurately detects the wheel scratch information, and realizes the detection of the wheel scratch acoustic emission signal while improving the detection accuracy. The present invention has a fast operation rate and high detection accuracy, and has important social significance and economic value in the field of train wheel damage detection.

[0006] The object of the present invention is achieved by the following technical solutions:

[0007] A wheel scratch detection method based on improved mapping fusion features and thresholds includes the following steps:

[0008] Step 1: Load and obtain P groups of wheel acoustic emission signals containing scratches and noises Extract seven features from each group of signals : root mean square value, entropy, kurtosis spectrum, skewness spectrum, mel-frequency cepstral coefficients, gamma-frequency cepstral coefficients, and third-order wavelet energy entropy, and determine the extreme values of each group of features and calculate the average value of the extreme values of the P groups of features;

[0009] Step 2: Use the average value of the extreme values of the P groups of features obtained in Step 1 as the detection threshold to preliminarily detect the acoustic emission signals, calculate the detection rate and precision rate of the scratch signals of each feature, screen the multi-dimensional features obtained, and obtain a low-dimensional feature parameter set with higher detection rate and precision rate and containing more scratch information S 2 represents the dimension of the screened low-dimensional feature parameter set. According to the multi-feature mapping fusion algorithm, map and fuse the low-dimensional feature parameter set to obtain a one-dimensional mapping fusion feature

[0010] Step 3: Fuse the one-dimensional mapping features obtained in Step 2 Divide them into several time windows, and each time window contains mapping fusion features H of length α v (β). According to the scratch signal detection rate F 1 and precision rate F 2 Comprehensive coefficient J selects the most appropriate length of α, calculates the standard deviation, mean and minimum value of H v (β), combines the correlation coefficient to calculate the improved adaptive detection threshold ATH within each time window, and based on the improved adaptive threshold discrimination algorithm, discriminates whether each feature belongs to the scratch signal, and completes the detection of the acoustic emission signal of the wheel scratch.

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

[0012] 1. The present invention uses a total of seven features, namely root mean square value, entropy, kurtosis spectrum, skewness spectrum, mel-frequency cepstral coefficients, gamma-frequency cepstral coefficients and third-order wavelet energy entropy, to characterize the scratch information containing the acoustic emission signal of the train wheel. Compared with the prior art, the scratch information is characterized from more and more perfect angles. Based on the designed scratch signal detection rate and accuracy, the above features are screened to obtain a more effective feature vector with higher detection rate and accuracy. A multi-feature information mapping fusion algorithm is proposed to map and fuse the more effective feature vectors obtained by screening into one-dimensional features, which can show the wheel scratch information from multiple angles, reduce the feature dimension, reduce the data operation cost, and greatly save the running time.

[0013] 2. The present invention improves the traditional adaptive threshold detection algorithm. Combining the characteristics of the mapping fusion features and time information, a time window of a certain length is selected to slide on the mapping fusion feature vector, and the improved adaptive threshold is calculated by combining the standard deviation, mean and minimum value information of the mapping fusion features and the environmental parameters determined by the actual train operation conditions. According to the improved adaptive threshold discrimination algorithm, the acoustic emission signal of the wheel to be detected is subjected to scratch detection. Compared with the existing traditional adaptive threshold detection technology, the improved adaptive threshold detection method proposed by the present invention has higher detection accuracy.

[0014] 3. The existing wheel abrasion signal detection methods often only apply to laboratory operation data or a small amount of measured data. Usually, these data only contain some simple known noise signals and do not fully cover the wheel-rail rolling noise during train operation, which is quite different from the actual train operation conditions. At the same time, the existing wheel abrasion detection methods are only based on a certain feature and cannot accurately represent the abrasion information of the wheel. Moreover, the traditional adaptive threshold is often calculated based on a small amount of information, which is prone to detection errors and seriously affects the detection accuracy. The multi-feature mapping fusion and improved adaptive threshold algorithm proposed by the present invention can map and fuse features of multiple dimensions into one-dimensional features, which can show the wheel abrasion information from multiple angles. At the same time, the improved adaptive threshold combined with the time window information can adjust the detection threshold in a timely manner according to the actual feature performance, effectively excluding the interference of noise signals, improving the detection accuracy of wheel abrasion signals, and meeting the safety guarantee requirements in actual train operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 FIG. is a flowchart of the wheel abrasion detection method based on improved mapping fusion features and thresholds of the present invention.

[0016] Figure 2 FIG. is a picture of the experimental site.

[0017] Figure 3 FIG. is a picture of wheel abrasion.

[0018] Figure 4 FIG. is a time-amplitude diagram of the transmitted signal.

[0019] Figure 5 FIG. is a time-amplitude diagram of the mapped fusion features.

[0020] Figure 6 FIG. is a diagram showing the corresponding relationship between the time window length and the comprehensive index

[0021] Figure 7 FIG. is an adaptive threshold abrasion signal detection diagram

[0022] Figure 8 FIG. is an example diagram of signal detection. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] The technical solutions of the present invention will be further described below in conjunction with the drawings, but are not limited thereto. Any modification or equivalent replacement of the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention shall be covered by the protection scope of the present invention.

[0024] The present invention provides a method for detecting wheel abrasion based on improved mapping fusion features and thresholds. First, a feature set is designed to be extracted from the acquired acoustic emission signals, including root mean square value, entropy, kurtosis spectrum, skewness spectrum, mel-frequency cepstral coefficients, gamma-frequency cepstral coefficients, and third-order wavelet energy entropy, a total of seven features. Then, three screening factors, namely the detection rate, precision rate, and comprehensive coefficient of the abrasion signal, are proposed to obtain several features with higher detection rate, precision rate, and comprehensive coefficient. Then, according to the proposed multi-feature mapping fusion algorithm, the screened features are mapped and fused to obtain one-dimensional mapping fusion features. Calculate the standard deviation, average value, and minimum value of the mapping fusion features, set the correlation coefficient according to the actual working conditions, and use the above parameters and time information to improve the traditional adaptive threshold detection algorithm. Finally, find the feature segments exceeding the threshold, determine the time points corresponding to these segments, and extract the signals from the initially acquired acoustic emission signals according to the time point information, which are the abrasion signals, and complete the detection of the acoustic emission signals of wheel abrasion. As Figure 1 shown, the specific steps are as follows:

[0025] Step 1: Load the acquired P groups of wheel acoustic emission signals containing abrasion and noise From each group of signals extract seven features: root mean square value, entropy, kurtosis spectrum, skewness spectrum, mel-frequency cepstral coefficients, gamma-frequency cepstral coefficients, and third-order wavelet energy entropy, and determine the extreme values of each group of features and calculate the average value of the extreme values of the P groups of features. The specific steps are as follows:

[0026] 1) Load the acquired P groups of wheel acoustic emission signals containing abrasion and noise where N i represents the length of each group of acoustic emission signals, that is, the number of sampling points, i = 1, 2,..., P;

[0027] 2) Respectively extract the root mean square value, entropy, kurtosis spectrum, skewness spectrum, mel-frequency cepstral coefficients, gamma-frequency cepstral coefficients, and wavelet energy entropy from each group of acoustic emission signals a total of seven features, which are respectively denoted as The length of each feature parameter set is N' i (N' i < N i ), because the long sampling time of the acoustic emission signal leads to too large a length of each signal segment, affecting the detection rate, so the acoustic emission signal is grouped using the sampling interval Q, and each feature is extracted from the grouped signals to reduce the amount of calculation and improve the detection rate;

[0028] 3) Determine the extreme values of the root mean square value, entropy, kurtosis spectrum, skewness spectrum, mel-frequency cepstral coefficients, gamma-frequency cepstral coefficients, and wavelet energy entropy of each group of acoustic emission signals, and calculate the average value of the extreme values of the P groups of features, which are respectively denoted as A R , AE , A K , A S , A M , A G and A X 。

[0029] Step 2: Using the average value of the P-group characteristic extreme values obtained in Step 1 as the detection threshold, preliminarily detect the acoustic emission signal, calculate the detection rate and precision rate of the scratch signal for each characteristic, perform feature screening on the obtained multi-dimensional features, and obtain a low-dimensional feature parameter set with higher detection rate and precision rate and containing more scratch information S 2 represents the dimension of the low-dimensional feature parameter set obtained by screening. According to the multi-feature mapping fusion algorithm, map and fuse the low-dimensional feature parameter set to obtain a one-dimensional mapped and fused feature The specific steps are as follows:

[0030] 1) Set the average value A of the P-group characteristic extreme values obtained in Step 1 R , A E , A K , A S , A M , A G and A X as the detection threshold for each characteristic. The signal exceeding the threshold is the scratch signal, which is compared with the known scratch signal, and the different signal is the interference signal;

[0031] 2) According to the actual operation requirements of the train, the detection rate and precision rate of the scratch signal are important parameter bases for judging whether the train can operate safely. Therefore, use the scratch signal detection rate F 1 and precision rate F 2 to screen the obtained multi-dimensional features, where:

[0032]

[0033]

[0034] In the formula, L a represents the number of scratch signals actually existing in the acoustic emission signal, L d and L i respectively represent the number of scratch signals detected by the feature and the number of interference signals;

[0035] 3) Considering that a single index can only reflect the characteristics of a certain aspect of the test result, in order to comprehensively measure the test result, the present invention designs a weighted comprehensive coefficient to fuse the above two parameters, specifically expressed as:

[0036] J = ε 1 F 1 + ε2 F 2

[0037]

[0038] In the formula, ε i represents the weight of the importance of each performance evaluation index, and the value is considered according to the actual operation conditions; according to the multi-feature mapping fusion algorithm, the feature set is mapped and fused to obtain a one-dimensional mapped fusion feature

[0039] 4) According to the scratch signal detection rate F 1 and precision rate F 2 and comprehensive coefficient J, the feature set E i is screened to obtain a low-dimensional feature parameter set with higher detection rate and precision rate and containing more scratch information Design and construct the key transformation matrix of the mapping fusion algorithm Among them:

[0040]

[0041] In the formula, "T" and "-1" represent the transpose operation and the inverse operation respectively, is a unified vector;

[0042] 5) Using the transformation matrix The present invention can creatively map the low-dimensional feature parameter set to a one-dimensional mapped fusion feature Among them:

[0043]

[0044] Step three: Divide the one-dimensional mapped fusion feature obtained in step two into several time windows, and each time window contains a mapped fusion feature H of length α v (β). According to the scratch signal detection rate F 1 and precision rate F 2 and comprehensive coefficient J, select the most appropriate length of α, calculate the standard deviation, mean and minimum value of H v (β), calculate the improved adaptive detection threshold ATH within each time window in combination with the correlation coefficient, and judge whether each feature belongs to the scratch signal according to the improved adaptive threshold discrimination algorithm, so as to complete the detection of the wheel scratch acoustic emission signal. The specific steps are as follows:

[0045] 1) Map the one-dimensional mapped fusion feature into θ time windows, and each time window contains a mapped fusion feature H of length α v (β) to obtain more time information of wheel scratches. Among them:

[0046] H v H(β) = {H v (m), m = 1 + (β - 1)δ,..., α + (β - 1)δ};

[0047] Where δ is the sliding interval of the time window, usually taken as δ = α / 3, β = 1, 2,..., θ, and g(·) is the rounding function;

[0048] θ = g(N′ i / α);

[0049] 2) Calculate the standard deviation S of the mapped fusion feature H v (β) within a time window α, where: v Among them:

[0050]

[0051] Among them, represents the operation of finding the average value of the mapped fusion feature H v (β) within a time window α;

[0052] 3) The mapped fusion feature takes on both positive and negative values. Therefore, based on the standard deviation, mean, and minimum value of H v (β), the calculation of the improved adaptive detection threshold ATH is divided into two cases:

[0053]

[0054] Among them, H represents the entire mapped fusion feature min(·) represents the operation of taking the minimum value, and λ, ω 1 and ω 2 respectively represent the environmental parameter variables determined by the working conditions;

[0055] 4) According to the improved adaptive detection threshold ATH calculation formula, select the most appropriate time window length α based on the scratch signal detection rate F 1 , precision rate F 2 and comprehensive coefficient J, and obtain the optimal detection threshold ATH;

[0056] 5) According to the improved adaptive detection threshold ATH obtained in 4), perform wheel scratch signal discrimination. The discrimination method is as follows:

[0057]

[0058] If the calculation result R is 1, then this feature belongs to the scratch signal; otherwise, it belongs to the noise signal. Discriminate whether each feature belongs to the scratch signal to complete the detection of wheel scratch acoustic emission signals.

[0059] The specific implementation manner of the present invention will be described below in conjunction with the measured wheel acoustic emission signal data.

[0060] The experimental verification data of the present invention is derived from the on-site acquisition of wheel actual rolling acoustic emission signals. The experimental site conditions are as Figure 2 shown. A scratch with a length of 5 cm, a width of 2 cm, and a depth of 2 mm is artificially made on the experimental wheel, and the shape is as Figure 3 shown. The wheel is manually pushed to roll back and forth on a 7-meter-long railway track at a speed of 3.5 km / h. Then, the signal receiving sensor is fixed on one side of the railway track through a mechanical fixture to receive the electrical signals generated by the wheel rolling. The amplifier increases the amplitude of the electrical signals, and then the acoustic emission acquisition device converts the electrical signals into acoustic emission signals. A total of 54 groups of effective wheel rolling acoustic emission signals are obtained in the experiment.

[0061] Execute Step 1: Load the 54 groups of wheel rolling acoustic emission signals containing scratches and noise obtained from actual measurements. The signal sampling rate is 2 MHz, and the length of each group of signals is different. Randomly select one group for display, and its time amplitude is as Figure 4 shown, with a total of 15,041,024 sampling points. It is known that the scratches of the selected signal occur at 0.816 seconds and 6.4 seconds.

[0062] Next, extract the root mean square, entropy, kurtosis spectrum, skewness spectrum, Mel-frequency cepstral coefficients, gamma-frequency cepstral coefficients, and third-order wavelet energy entropy of each group of acoustic emission signals, a total of seven features, and set Q = 256. Among them, the Mel-frequency cepstral coefficients select the 3rd-dimensional feature, the gamma-frequency cepstral coefficients select the 3rd and 5th-dimensional features, and the third-order wavelet energy entropy selects the first-dimensional feature. In this way, each group of acoustic emission signals has 8-dimensional features.

[0063] Then determine the extreme values of each group of features, and calculate the average values of the 54 groups of feature extreme values, which are: 2.1929, 0.0016, 499.2659, 14.2426, -322.1328, -7.983, -4.6371, and 0.7426 respectively.

[0064] Execute Step 2: Use the average values of the 54 groups of feature extreme values obtained in Step 1 as the detection threshold to preliminarily detect the wheel scratch signals. The signals exceeding the threshold are scratch signals, and compare them with the known scratch signals. The different signals are interference signals. According to the number of detected scratch signals and interference signals, calculate the scratch signal detection rate F 1 and precision rate F 2 and comprehensive coefficient J. At the same time, according to the actual operation safety guarantee requirements of the train, the detection rate and precision rate account for the same proportion. Therefore, the feature weight coefficients ε 1 and ε 2They are 0.5 and 0.5 respectively. The detection results of each dimension feature are shown in Table 1.

[0065] Table 1 Detection rate, precision rate and comprehensive coefficient of scratch signals for each dimension feature

[0066]

[0067] According to the results in Table 1, features with excellent performance in each index are selected. The gamma spectrum cepstrum coefficient 3, gamma spectrum cepstrum coefficient 5 and third-order wavelet energy entropy 1 are thus selected to construct a three-dimensional feature parameter set containing more scratch information. The three-dimensional feature set is input into the multi-feature mapping fusion algorithm proposed by the present invention to obtain a one-dimensional mapping fusion feature. The mapping fusion feature of the acoustic emission signal selected in Step 1 is as Figure 5 shown. Repeating the above process, 54 groups of mapping fusion features of acoustic emission signals can be obtained.

[0068] Execute Step 3: Calculate that the mean value and minimum value of the mapping fusion feature of the acoustic emission signal selected in Step 1 are: 0.2426 and 0.9877 respectively. At the same time, set the environmental parameters λ, ω 1 and ω 2 to be 1.2, 0.35 and 0.4 respectively.

[0069] Next, select the time window length between 0.1 ms and 10 ms, calculate the corresponding adaptive threshold, and detect the acoustic emission signal according to the adaptive threshold to determine the number of scratch signals and the number of interference signals. Further calculate the scratch signal detection rate F 1 and precision rate F 2 and comprehensive coefficient J. According to the corresponding relationship between the time window length and the comprehensive coefficient, determine the size of the nearest time window length. The corresponding relationship between the time window length and the comprehensive coefficient is as Figure 6 shown.

[0070] According to Figure 4 it can be seen that when the time window length is 7.2 ms, the corresponding comprehensive coefficient value is the largest. Therefore, set the time window length to 7.2 ms, and calculate the adaptive threshold according to the mean value, standard deviation and minimum value of the mapping fusion feature, as well as the set environmental parameter variables. The improved adaptive threshold of the acoustic emission signal selected in Step 1 is as Figure 7 shown. Finally, based on the improved adaptive threshold decision algorithm, the wheel acoustic emission signal is discriminated, and the discrimination result is as Figure 8 shown. The positions of the signals with the discrimination result of 1 are marked in the figure, and this section of the signal is the wheel scratch signal.

[0071] Repeat the above steps to complete the scratch detection of 54 groups of wheel acoustic emission signals. Statistically analyze the number of scratch signals and interference signals detected based on the multi-feature mapping fusion and improved adaptive threshold detection algorithm proposed in the present invention, and calculate the scratch signal detection rate F 1 and the precision rate F 2 and the comprehensive coefficient J. At the same time, according to the traditional adaptive threshold calculation method, calculate the average value of the mapped fusion feature extreme values of 54 groups of wheel acoustic emission signals to be 0.4856, and detect the acoustic emission signals based on this threshold. The detection results of the improved adaptive threshold detection algorithm proposed in the present invention and the traditional adaptive threshold method are shown in Table 2

[0072] Table 2 Comparison between the improved adaptive threshold and the traditional method

[0073]

[0074] It can be found from Table 2 that the improved adaptive threshold algorithm proposed in the present invention significantly outperforms the traditional adaptive threshold detection algorithm in terms of the three indicators of scratch signal detection rate, precision rate, and comprehensive coefficient. Therefore, through the above experimental verification, the following conclusion can be obtained: The multi-feature mapping fusion and improved adaptive threshold detection algorithm proposed in the present invention can significantly improve the scratch signal detection rate and detection accuracy of wheels, meeting the requirements for ensuring the safe operation of trains

Claims

1. A method for detecting wheel abrasion based on improved mapping fusion features and thresholds, characterized in that the method comprises the following steps: Step 1: Load the obtained P groups of wheel acoustic emission signals X containing abrasions and noises Ni , and extract seven features from each group of signals X Ni : Root mean square value, entropy, kurtosis spectrum, skewness spectrum, Mel-frequency cepstral coefficients, gamma-frequency cepstral coefficients, and third-order wavelet energy entropy, and determine the extreme values of each group of features and calculate the average value of the extreme values of the P-group features, where N i represents the length of each group of acoustic emission signals, that is, the number of sampling points, i = 1, 2,..., P; Step 2: Using the average value of the P groups of characteristic extreme values obtained in Step 1 as the detection threshold, preliminarily detect the acoustic emission signal, calculate the detection rate and precision rate of the scratch signal for each characteristic, screen the obtained multi-dimensional characteristics, and obtain a low-dimensional characteristic parameter set with higher detection rate and precision rate and containing more scratch information represents the dimension of the low-dimensional characteristic parameter set obtained by screening, N i ′ is the length of each characteristic parameter set. According to the multi-characteristic mapping fusion algorithm, the low-dimensional characteristic parameter set E S2*Ni′ is mapped and fused to obtain a one-dimensional mapped fusion characteristic The specific steps are as follows: 1) Set the average value A of the P groups of characteristic extreme values obtained in Step 1 R , A E , A K , A S , A M , A G and A X are the detection thresholds for each characteristic. A signal exceeding the threshold is a scratch signal, which is compared with the known scratch signal. A different signal is an interference signal; 2) Using the abrasion signal detection rate F 1 and the precision rate F 2 screen the obtained multi-dimensional features, where: where L a represents the number of actual scratch signals present in the acoustic emission signal, L d and L i represent the number of scratch signals detected by the feature and the number of interference signals respectively; 3) Design a weighted comprehensive coefficient to fuse the detection rate and precision rate of abrasion signals, specifically expressed as: J = ε 1 F 1 + ε 2 F 2 where ε i represents the weight of the importance of each performance evaluation index. According to the multi-feature mapping fusion algorithm, the feature set is mapped and fused to obtain a one-dimensional mapped fusion feature 4) According to the abrasion signal detection rate F 1 , precision rate F 2 and comprehensive coefficient J, screen the feature set E i to obtain a low-dimensional feature parameter set with higher detection rate and precision rate and containing more abrasion information Design and construct the key transformation matrix of the mapping fusion algorithm Where: where "T" and "-1" represent the transpose operation and the inverse operation respectively, is the unified vector; 5) Using a transformation matrix to map the low-dimensional feature parameter set to a one-dimensional mapped fusion feature Where: Step 3: Fuse the one-dimensional mapping features obtained in Step 2 Divide them into several time windows, and each time window contains mapping fusion features H v (β) of length α. According to the detection rate F 1 of abrasion signals and the precision rate F 2 of the comprehensive coefficient J, select the most suitable length of α, calculate the standard deviation, mean and minimum value of H v (β), calculate the improved adaptive detection threshold ATH within each time window in combination with the correlation coefficient, and according to the improved adaptive threshold discrimination algorithm, determine whether each feature belongs to the abrasion signal, and complete the detection of the acoustic emission signal of wheel abrasion. The specific steps are as follows: 1) Map one-dimensional mapped fusion features Divide it into θ time windows, and each time window contains mapped fusion features H of length α v (β) to obtain more time information of wheel abrasion, where: H v H(β) = {H v (m), m = 1 + (β - 1)δ,..., α + (β - 1)δ}; where δ is the sliding interval of the time window, usually taking δ = α / 3, β = 1, 2,..., θ, and g(·) is the rounding function; θ = g(N i ′ / α); 2) Calculate the mapped fusion feature H within a time window α v The standard deviation S of (β) v , where: In the formula, represents the operation of obtaining the mapped fusion feature H within a time window α v (β) average value operation; 3) According to H v (β), the standard deviation, mean value, and minimum value, the improvement of the calculation of the adaptive detection threshold ATH is divided into two cases: Wherein, H represents the integrated mapping fusion feature for the whole segment min(·) represents the minimum value operation, λ, ω 1 and ω 2 respectively represent the environmental parameter variables determined by the working conditions; 4) According to the improved adaptive detection threshold ATH calculation formula, select the most appropriate time window length α based on the scratch signal detection rate F 1 , precision rate F 2 and comprehensive coefficient J, and obtain the optimal detection threshold ATH; 5) According to the improved adaptive detection threshold ATH obtained in 4), perform wheel abrasion signal discrimination, and the discrimination method is as follows: If the calculation result R is 1, then this feature belongs to the abrasion signal; otherwise, it belongs to the noise signal; discriminate whether each feature belongs to the abrasion signal to complete the detection of the wheel abrasion acoustic emission signal.

2. The method for detecting wheel abrasion based on improved mapping fusion features and thresholds according to claim 1, characterized in that the specific steps of step one are as follows: 1) Load the obtained P groups of wheel acoustic emission signals X containing abrasions and noises Ni = {x(1), x(2),..., x(Ni)}, where Ni represents the length of each group of acoustic emission signals, that is, the number of sampling points, and i = 1, 2,..., P; 2) Extract the root mean square value, entropy, kurtosis spectrum, skewness spectrum, Mel-frequency cepstral coefficients, gamma-frequency cepstral coefficients, and wavelet energy entropy from each group of acoustic emission signals X Ni respectively, a total of seven features, which are denoted as The length of each set of feature parameters is N i ′, N i ′ < N i , group the acoustic emission signals using the sampling interval Q, and extract each feature from the grouped signals; 3) Determine the root mean square values, entropy, kurtosis spectra, skewness spectra, Mel-frequency cepstral coefficients, gamma-frequency cepstral coefficients, and extreme values of wavelet energy entropy of each group of acoustic emission signals, and calculate the average of the extreme values of the P-group features, denoted as A R , A E , A K , A S , A M , A G and A X .

Images