Feature extraction method for abnormal abrasion sound of carrier roller of non-coal mine belt conveyor

Abstract

The invention discloses a non-coal mine belt conveyor carrier roller wear abnormal sound feature extraction method, and particularly relates to the technical field of navigation positioning, and the feature extraction method comprises the following steps: S1, constructing an anti-interference sound collection system, and collecting a sound signal in a carrier roller operation process; s2, preprocessing the collected sound signals, wherein the preprocessing comprises adaptive filtering and denoising, signal segmentation and abnormal peak suppression; s3, adopting a multi-domain fusion feature extraction strategy to extract time domain features, frequency domain features and time-frequency domain features from the preprocessed sound signals; s4, screening out an optimal feature subset sensitive to the wear state of the carrier roller through feature importance evaluation and redundancy analysis; the non-coal mine belt conveyor carrier roller wear abnormal sound feature extraction method can effectively resist noise interference of a non-coal mine complex environment, accurately extract carrier roller wear abnormal sound features, and provide a reliable basis for intelligent diagnosis of a carrier roller wear state.

Description

Technical Field

[0001] This invention relates to the field of navigation and positioning technology, and more specifically, this application relates to a method for feature extraction of abnormal wear sound of conveyor belt rollers in non-coal mines. Background Technology

[0002] Belt conveyors are key equipment for material transportation, and idlers, as an important component of belt conveyors, directly affect the operating efficiency and safety of the conveyor. During long-term operation, idlers are prone to wear due to factors such as material impact, dust erosion, and friction. If this wear is not detected in time, it may lead to idler failure, belt misalignment, or even breakage, causing production interruptions or safety accidents.

[0003] Currently, idler roller wear detection mainly relies on manual inspection and vibration testing. Manual inspection is inefficient, highly subjective, and difficult to adapt to the complex and harsh environment of non-coal mines; vibration testing is easily affected by strong background vibrations in mines, resulting in poor feature extraction accuracy. Sound signals, as an important carrier of equipment operating status, contain rich fault information, and sound acquisition has advantages such as being non-contact and easy to install. However, the non-coal mine environment is filled with strong interference such as mechanical noise and material impact noise, which drowns out the sound characteristics of abnormal idler roller wear. Traditional sound feature extraction methods are unable to effectively extract these features, leading to low detection accuracy.

[0004] Therefore, a feature extraction method for abnormal wear sound of conveyor belt rollers in non-coal mines is proposed to address the above problems. Summary of the Invention

[0005] To achieve the above objectives, this application provides the following technical solution: a method for feature extraction of abnormal wear sound of conveyor belt idler rollers in non-coal mines, the feature extraction method comprising the following steps:

[0006] S1. Construct an anti-interference sound acquisition system by deploying array-type sound sensors in the idler group area of ​​belt conveyors in non-coal mines to collect sound signals during the operation of the idlers.

[0007] S2. Preprocess the acquired audio signal, including adaptive filtering for noise reduction, signal segmentation, and suppression of abnormal spikes;

[0008] S3. A multi-domain fusion feature extraction strategy is adopted to extract time-domain features, frequency-domain features, and time-frequency-domain features from the preprocessed audio signal;

[0009] S4. Through feature importance assessment and redundancy analysis, the optimal feature subset sensitive to the wear state of the idler roller is selected.

[0010] Furthermore, the array of acoustic sensors consists of 4-8 sensors, which are evenly distributed on the frame on both sides of the roller assembly. The acquisition frequency range is 20Hz-20kHz, and each sensor acquires data synchronously. The acquired data is transmitted to the data processing terminal via optical fiber.

[0011] Furthermore, the adaptive filtering and denoising in S2 employs an adaptive filter based on the minimum mean square error criterion, using the sound signals of adjacent idlers as reference noise to denoise the sound signal of the target idler. The filter output formula is: y(n) = w T The error signal formula is: e(n)=d(n)-y(n), and the weight update formula is: w(n+1)=w(n)+μe(n)x(n), where y(n) represents the output signal of the filter at time n, w(n) is the weight vector used to weight the input signal at time n, and w T (n) is the transpose of the weight vector, x(n) represents the input signal at time n, e(n) is the error signal at time n, d(n) is the expected signal at time n, i.e. the ideal output signal, w(n+1) is the weight vector updated at time n+1, and μ is the step size parameter used to control the magnitude of the weight update.

[0012] The signal segmentation adopts the sliding window method, with a window length of 50-200ms and a sliding step size of 1 / 4-1 / 2 of the window length;

[0013] The abnormal spike suppression identifies and replaces abrupt spikes in the signal by setting a dynamic threshold. The dynamic threshold is 3-5 times the root mean square (RMS) value of the current window signal. The RMS value is calculated using the following formula: Where RMS represents the root mean square value of the signal in the current window, N is the number of signal samples in the current window, and x i It is the value of the i-th signal sample within the window, where i = 1, 2, ..., N.

[0014] Furthermore, the time-domain features in S3 include the signal's peak factor, kurtosis, impulse factor, margin factor, and waveform factor, wherein the peak factor is calculated using the following formula:

[0015] The formula for calculating kurtosis is: The formula for calculating the impulse factor is: The formula for calculating the margin factor is: The formula for calculating the waveform factor is: Where x max This represents the peak value of the signal.

[0016] Furthermore, the frequency domain features in S3 include the power spectrum peak value, centroid frequency, mean square frequency, and frequency variance obtained through Fast Fourier Transform (FFT). The FFT formula is as follows:

[0017] Where k = 0, 1, ..., N′-1, the formula for calculating the centroid frequency is: The formula for calculating the mean square frequency is:

[0018] The formula for calculating frequency variance is: Where X(k) is the value of the frequency domain signal after discrete Fourier transform at the k-th frequency point, N′ is the number of sampling points, and j is the imaginary unit, and j 2 =-1, f i For the i-th frequency point, P(f i ) represents the power spectral density value at the i-th frequency point, and M represents the number of frequency points.

[0019] Furthermore, in S3, the time-frequency domain features are extracted using wavelet packet decomposition and Mel frequency cepstral coefficients. The wavelet packet decomposition uses the db4 wavelet basis with 4-6 decomposition layers. The energy proportion, wavelet energy entropy, and wavelet singular entropy of each frequency band are calculated. The formula for calculating the frequency band energy is:

[0020] x i (t) represents the signal in the i-th frequency band, and the frequency band energy percentage is: E represents the number of decomposition layers. i This represents the energy percentage of the i-th frequency band, i.e., the energy of the i-th frequency band in all 22 frequency bands. d The proportion of total energy in each frequency band; the formula for calculating wavelet energy entropy is: Wavelet singular entropy is obtained by performing singular value decomposition on the signal to obtain the singular value λ. i The formula for calculating wavelet singular entropy is:

[0021] Where r is the number of non-zero singular values.

[0022] Furthermore, the Mel frequency cepstral coefficients are obtained by performing Mel filtering bank processing, logarithmic energy calculation, and discrete cosine transform on the signal. The calculation formula is as follows:

[0023] Where L is the number of Mel filter banks, m is the order of the cepstral coefficients, taking the first 12-16 orders, and X(k′) is the amplitude value of the signal at frequency point k′ after Mel filter bank processing.

[0024] Furthermore, in S4, the feature importance assessment employs a feature importance ranking method based on random forests, calculating the reduction in Gini index for each feature during the classification process to evaluate the feature's ability to distinguish the wear state of the idler rollers. The Gini index calculation formula is as follows:

[0025] Where p k Let K be the probability of class k, where K is the total number of classes.

[0026] Redundancy analysis was performed using the Pearson correlation coefficient, calculated as follows:

[0027] Where, x′ i For the i-th sample value of the first feature, Let y′ be the mean of the first feature. i For the i-th sample value of the second feature, The mean of the second characteristic;

[0028] Redundant features with an absolute correlation coefficient greater than 0.8 are removed. A recursive feature elimination algorithm is used to gradually remove the least important features until the optimal feature subset is obtained.

[0029] The technical effects and advantages of this application are as follows:

[0030] 1. Compared with existing technologies, the feature extraction method for abnormal wear sound of conveyor belt rollers in non-coal mines can effectively resist noise interference from the complex environment of non-coal mines, accurately extract the abnormal sound features of roller wear, and provide a reliable basis for intelligent diagnosis of roller wear status.

[0031] 2. The use of array-type acoustic sensors to collect sound signals, combined with fiber optic transmission, improves the quality of the original signal and its anti-interference ability, making it suitable for the harsh environmental conditions of non-coal mines.

[0032] 3. Through preprocessing steps such as adaptive filtering and dynamic threshold peak suppression, environmental noise and sudden interference are effectively eliminated, providing a clean signal foundation for feature extraction.

[0033] 4. By integrating time domain, frequency domain, and multi-dimensional features, the sound characteristics of roller wear can be fully captured, avoiding the problem of incomplete feature information in a single domain and improving the distinguishability of features.

[0034] 5. By selecting the optimal feature subset through feature importance assessment and redundancy analysis, the feature dimensionality is reduced, the complexity of the subsequent diagnostic model is lowered, and the effectiveness and stability of the features are ensured at the same time. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] Example 1

[0038] As attached Figure 1 The method shown is a feature extraction method for abnormal wear sound of conveyor rollers in non-coal mines. The feature extraction method includes the following steps:

[0039] S1. Construct an anti-interference sound acquisition system by deploying array-type sound sensors in the idler group area of ​​belt conveyors in non-coal mines to collect sound signals during the operation of the idlers.

[0040] The array of acoustic sensors consists of 4-8 sensors, which are evenly distributed on the frame on both sides of the roller assembly. The acquisition frequency range is 20Hz-20kHz, and each sensor acquires data synchronously. The acquired data is transmitted to the data processing terminal via optical fiber.

[0041] S2. Preprocess the acquired audio signal, including adaptive filtering for noise reduction, signal segmentation, and suppression of abnormal spikes;

[0042] The adaptive filtering denoising described in S2 employs an adaptive filter based on the minimum mean square error criterion. It uses the sound signals of adjacent idlers as reference noise to denoise the sound signal of the target idler. The filter output formula is: y(n) = w T The error signal formula is: (n)x(n)

[0043] e(n) = d(n) - y(n), and the weight update formula is: w(n+1) = w(n) + μe(n)x(n), where y(n) represents the output signal of the filter at time n, w(n) is the weight vector used to weight the input signal at time n, and w... T (n) is the transpose of the weight vector, x(n) represents the input signal at time n, e(n) is the error signal at time n, d(n) is the expected signal at time n, i.e. the ideal output signal, w(n+1) is the weight vector updated at time n+1, and μ is the step size parameter used to control the magnitude of the weight update.

[0044] The signal segmentation adopts the sliding window method, with a window length of 50-200ms and a sliding step size of 1 / 4-1 / 2 of the window length;

[0045] The abnormal spike suppression identifies and replaces abrupt spikes in the signal by setting a dynamic threshold. The dynamic threshold is 3-5 times the root mean square (RMS) value of the current window signal. The RMS value is calculated using the following formula: Where RMS represents the root mean square value of the signal in the current window, N is the number of signal samples in the current window, and x i It is the value of the i-th signal sample within the window, where i = 1, 2, ..., N;

[0046] S3. A multi-domain fusion feature extraction strategy is adopted to extract time-domain features, frequency-domain features, and time-frequency-domain features from the preprocessed audio signal;

[0047] The time-domain features described in S3 include the signal's peak factor, kurtosis, impulse factor, margin factor, and waveform factor. The peak factor is calculated using the following formula: The formula for calculating kurtosis is: The formula for calculating the impulse factor is: The formula for calculating the margin factor is: The formula for calculating the waveform factor is:

[0048] Where x max This represents the peak value of the signal.

[0049] The S3 mid-frequency domain characteristics include the power spectrum peak value, centroid frequency, mean square frequency, and frequency variance obtained through Fast Fourier Transform (FFT). The FFT formula is as follows:

[0050] Where k = 0, 1, ..., N′-1, the formula for calculating the centroid frequency is: The formula for calculating the mean square frequency is:

[0051] The formula for calculating frequency variance is: Where X(k) is the value of the frequency domain signal after discrete Fourier transform at the k-th frequency point, N′ is the number of sampling points, and j is the imaginary unit, and j 2 =-1, f i For the i-th frequency point, P(f i ) represents the power spectral density value at the i-th frequency point, and M represents the number of frequency points.

[0052] In S3, the time-frequency domain features are extracted using wavelet packet decomposition and Mel frequency cepstral coefficients. Wavelet packet decomposition uses the db4 wavelet basis with 4-6 decomposition levels. The energy proportion of each frequency band, wavelet energy entropy, and wavelet singular entropy are calculated. The formula for calculating the frequency band energy is as follows:

[0053] x i (t) represents the signal in the i-th frequency band, and the frequency band energy percentage is: d represents the number of decomposition layers, E i This represents the energy percentage of the i-th frequency band, i.e., the energy of the i-th frequency band in all 22 frequency bands. d The proportion of total energy in each frequency band; the formula for calculating wavelet energy entropy is:

[0054] Wavelet singular entropy is obtained by performing singular value decomposition on the signal to obtain the singular value λ. i The formula for calculating wavelet singular entropy is: Where r is the number of non-zero singular values.

[0055] The Mel frequency cepstral coefficients are obtained by performing Mel filtering, logarithmic energy calculation, and discrete cosine transform on the signal. The calculation formula is as follows:

[0056] Where L is the number of Mel filter banks, m is the order of the cepstral coefficients, and the first 12-16 order coefficients are taken, and X(k′) is the amplitude value of the signal at frequency point k′ after Mel filter bank processing;

[0057] S4. Through feature importance assessment and redundancy analysis, the optimal feature subset sensitive to the wear state of the idler roller is selected;

[0058] In S4, feature importance assessment uses a feature importance ranking method based on random forests. The reduction in Gini index for each feature during the classification process is calculated to evaluate the feature's ability to distinguish the wear state of the idler rollers. The Gini index calculation formula is as follows: Where p k Let K be the probability of class k, where K is the total number of classes.

[0059] Redundancy analysis was performed using the Pearson correlation coefficient, calculated as follows:

[0060] Where, x′ i For the i-th sample value of the first feature, Let y′ be the mean of the first feature. i For the i-th sample value of the second feature, The mean of the second characteristic;

[0061] Redundant features with an absolute correlation coefficient greater than 0.8 are removed. A recursive feature elimination algorithm is used to gradually remove the least important features until the optimal feature subset is obtained.

[0062] Example 2

[0063] Taking the wear detection of idler rollers of a belt conveyor in an iron mine as an example, the method described in this invention is used to extract the abnormal sound characteristics of idler roller wear.

[0064] Anti-interference sound acquisition: Six waterproof and dustproof sound sensors are evenly distributed on the frames on both sides of the belt conveyor roller group. The model is a certain brand of noise-canceling microphone, and the acquisition frequency is set to 16kHz. Each sensor achieves synchronous acquisition through a synchronous controller. The acquired sound signal is transmitted to the data processing server via optical fiber.

[0065] Signal preprocessing: An adaptive filter based on the minimum mean square error criterion is used to reduce noise in the target idler signal, with the sound signal of adjacent normal idlers as reference noise, and a step size factor μ = 0.01. A sliding window method is used for segmentation, with a window length of 100ms and a sliding step size of 25ms. The dynamic threshold is set to 4 times the root mean square value of the current window signal, and the root mean square value is calculated according to the formula...

[0066] Calculate, identify, and replace abnormal spikes;

[0067] Multi-domain fusion feature extraction: Extracting time-domain features includes the signal's peak factor, kurtosis, impulse factor, margin factor, and waveform factor. The peak factor is calculated using the following formula:

[0068] The formula for calculating kurtosis is: The formula for calculating the impulse factor is: The formula for calculating the margin factor is: The formula for calculating the waveform factor is: Where x max The peak value of the signal;

[0069] Frequency domain features are obtained through Fast Fourier Transform (FFT), where the FFT is calculated according to the formula...

[0070] The calculation, and then the calculation of the centroid frequency, is performed using the following formula:

[0071] The formula for calculating the mean square frequency is: The formula for calculating frequency variance is: Where X(k) is the value of the frequency domain signal after discrete Fourier transform at the k-th frequency point, N′ is the number of sampling points, and j is the imaginary unit, and j 2 =-1, f i For the i-th frequency point, P(f i ) represents the power spectral density value at the i-th frequency point, and M represents the number of frequency points;

[0072] The signal is decomposed into 5-level wavelet packets using the db4 wavelet basis, according to the formula.

[0073] Calculate the energy of each frequency band, then according to Calculate the frequency band energy ratio, according to Calculate the wavelet energy entropy, obtain the singular values ​​through singular value decomposition, and then... Calculate the wavelet singular entropy;

[0074] The signal is processed by Mel filtering (banking, L=40), logarithmic energy calculation, and discrete cosine transform, according to the formula. Calculate and take the cepstral coefficients of the first 16 Mel frequencies;

[0075] Feature selection and optimization: The Gini index reduction for each feature was calculated using the Random Forest algorithm (with 100 trees). (and calculation), assess the importance of features; use Pearson correlation coefficient (based on...)

[0076] Features with an absolute correlation coefficient greater than 0.8 were removed by calculation; a recursive feature elimination algorithm was used to select 10 optimal features, including kurtosis, wavelet energy entropy, and 8th order Mel frequency cepstral coefficient, as input features for roller wear condition diagnosis.

[0077] Through the above embodiments, the abnormal sound features of idler roller wear were successfully extracted. The test results showed that the accuracy of this feature set in idler roller wear diagnosis reached more than 95%, which verified the effectiveness of the method of the present invention.

[0078] Finally: The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for extracting the features of abnormal wear sound of conveyor rollers in non-coal mines, characterized in that: The feature extraction method includes the following steps: S1. Construct an anti-interference sound acquisition system by deploying array-type sound sensors in the idler group area of ​​belt conveyors in non-coal mines to collect sound signals during the operation of the idlers. S2. Preprocess the acquired audio signal, including adaptive filtering for noise reduction, signal segmentation, and suppression of abnormal spikes; S3. A multi-domain fusion feature extraction strategy is adopted to extract time-domain features, frequency-domain features, and time-frequency-domain features from the preprocessed audio signal; S4. Through feature importance assessment and redundancy analysis, the optimal feature subset sensitive to the wear state of the idler roller is selected.

2. The method for extracting the features of abnormal wear sound of conveyor rollers in non-coal mines according to claim 1, characterized in that: The array of acoustic sensors consists of 4-8 sensors, which are evenly distributed on the frame on both sides of the roller assembly. The acquisition frequency range is 20Hz-20kHz, and each sensor acquires data synchronously. The acquired data is transmitted to the data processing terminal via optical fiber.

3. The method for extracting the features of abnormal wear sound of conveyor belt rollers in non-coal mines according to claim 2, characterized in that: The adaptive filtering and denoising in S2 employs an adaptive filter based on the minimum mean square error criterion. It uses the sound signals of adjacent idlers as reference noise to denoise the sound signal of the target idler. The filter output formula is: y(n) = w T The error signal formula is: e(n)=d(n)-y(n), and the weight update formula is: w(n+1)=w(n)+μe(n)x(n), where y(n) represents the output signal of the filter at time n, w(n) is the weight vector used to weight the input signal at time n, and w T (n) is the transpose of the weight vector, x(n) represents the input signal at time n, e(n) is the error signal at time n, d(n) is the expected signal at time n, i.e. the ideal output signal, w(n+1) is the weight vector updated at time n+1, and μ is the step size parameter used to control the magnitude of the weight update. The signal segmentation adopts the sliding window method, with a window length of 50-200ms and a sliding step size of 1 / 4-1 / 2 of the window length; The abnormal spike suppression identifies and replaces abrupt spikes in the signal by setting a dynamic threshold. The dynamic threshold is 3-5 times the root mean square (RMS) value of the current window signal. The RMS value is calculated using the following formula: Where TMS represents the root mean square value of the signal in the current window, N is the number of signal samples in the current window, and x i It is the value of the i-th signal sample within the window, where i = 1, 2, ..., N.

4. The method for extracting the features of abnormal wear sound of conveyor rollers in non-coal mines according to claim 3, characterized in that: The time-domain features mentioned in S3 include the signal's peak factor, kurtosis, impulse factor, margin factor, and waveform factor. The peak factor is calculated using the following formula: The formula for calculating kurtosis is: The formula for calculating the impulse factor is: The formula for calculating the margin factor is: The formula for calculating the waveform factor is: Where x max This represents the peak value of the signal.

5. The method for extracting the features of abnormal wear sound of conveyor rollers in non-coal mines according to claim 4, characterized in that: The frequency domain features of S3 include the power spectrum peak value, centroid frequency, mean square frequency, and frequency variance obtained through Fast Fourier Transform (FFT). The FFT formula is as follows: Where k = 0, 1, ..., N′-1, the formula for calculating the centroid frequency is: The formula for calculating the mean square frequency is: The formula for calculating frequency variance is: Where X(k) is the value of the frequency domain signal after discrete Fourier transform at the k-th frequency point, N′ is the number of sampling points, and j is the imaginary unit, and j 2 =-1, f i For the i-th frequency point, P(f i ) represents the power spectral density value at the i-th frequency point, and M represents the number of frequency points.

6. The method for extracting the features of abnormal wear sound of conveyor rollers in non-coal mines according to claim 5, characterized in that: The time-frequency domain features in S3 are extracted using wavelet packet decomposition and Mel frequency cepstral coefficients. Wavelet packet decomposition uses the db4 wavelet basis with 4-6 decomposition levels. The energy proportion, wavelet energy entropy, and wavelet singular entropy of each frequency band are calculated. The formula for calculating the frequency band energy is as follows: x i (t) represents the signal in the i-th frequency band, and the frequency band energy percentage is: E represents the number of decomposition layers. i This represents the energy percentage of the i-th frequency band, i.e., the energy of the i-th frequency band in all 22 frequency bands. d The proportion of total energy in each frequency band; the formula for calculating wavelet energy entropy is: Wavelet singular entropy is obtained by performing singular value decomposition on the signal to obtain the singular value λ. i The formula for calculating wavelet singular entropy is: Where r is the number of non-zero singular values.

7. The method for extracting the features of abnormal wear sound of conveyor rollers in non-coal mines according to claim 6, characterized in that: The Mel frequency cepstral coefficients are obtained by performing Mel filtering bank processing, logarithmic energy calculation, and discrete cosine transform on the signal. The calculation formula is as follows: Where L is the number of Mel filter banks, m is the order of the cepstral coefficients, taking the first 12-16 orders, and X(k′) is the amplitude value of the signal at frequency point k′ after Mel filter bank processing.

8. The method for extracting the features of abnormal wear sound of conveyor rollers in non-coal mines according to claim 7, characterized in that: In step S4, the feature importance assessment uses a feature importance ranking method based on random forests. The reduction in the Gini index for each feature during the classification process is calculated to evaluate the feature's ability to distinguish the wear state of the idler rollers. The Gini index calculation formula is as follows: Where p k Let K be the probability of class k, where K is the total number of classes. Redundancy analysis was performed using the Pearson correlation coefficient, calculated as follows: Where, x′ i For the i-th sample value of the first feature, Let y′ be the mean of the first feature. i For the i-th sample value of the second feature, The mean of the second characteristic; Redundant features with an absolute correlation coefficient greater than 0.8 are removed. A recursive feature elimination algorithm is used to gradually remove the least important features until the optimal feature subset is obtained.

Images