Acoustic emission signal denoising and feature extraction method and system suitable for deep roadway

Abstract

The application discloses a noise reduction and feature extraction method and system of acoustic emission signals suitable for deep roadways, and relates to the technical field of safety monitoring of deep mineral resource exploitation. The specific steps of the method are as follows: acoustic emission sensor arrays are arranged along the deep roadways, multi-source data is collected and converted into digital signals for storage; the type of interference is identified through wavelet packet decomposition and environmental correction algorithm; self-adaptive noise reduction is achieved by using complexity-sensitive punishment algorithm; time domain and frequency domain features are extracted and coupling parameters are calculated, time-frequency domain features are obtained by using Hilbert-Huang transform; finally, dimensionality reduction is achieved through principal component analysis, and mutual information entropy screening is performed, a feature vector is constructed and transmitted to a safety warning system; the application improves the processing accuracy and reliability of acoustic emission signals through multi-source signal collection, interference identification and self-adaptive noise reduction modules; the signal characteristics are fully described by using algorithm optimization, key features are output after feature optimization, real-time monitoring and early warning are realized by using dynamic damage vector algorithm, and the safety of deep roadway construction is ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of safety monitoring of deep mineral resources exploitation, in particular to a method and system for acoustic emission signal noise reduction and feature extraction suitable for deep roadway. BACKGROUND

[0002] With the extension of underground engineering to deep part, the stability monitoring of deep roadway surrounding rock becomes a key link to ensure construction safety. As a dynamic monitoring method, acoustic emission technology can capture the elastic wave signals generated by rock mass rupture and reflect the damage state of surrounding rock in real time. However, the deep roadway environment is complex, and the acoustic emission signals are easily disturbed by multi-source interference such as mechanical vibration, electromagnetic radiation and environmental noise, resulting in low signal-to-noise ratio and blurred features. In addition, the damage evolution of surrounding rock has nonlinear and multi-scale characteristics, and traditional signal processing methods cannot balance noise reduction effect and feature integrity, leading to lagging or false monitoring results. Therefore, it is of great practical significance to develop an acoustic emission signal processing technology suitable for the complex environment of deep roadway to improve the monitoring accuracy of surrounding rock stability and prevent disasters.

[0003] The traditional acoustic emission signal processing technology has many limitations: first, the interference identification relies on fixed threshold or single feature, which cannot adapt to the dynamic change of multi-source interference scene, resulting in incomplete noise removal or loss of effective signals; second, the noise reduction method mainly uses static filtering or empirical mode decomposition, without considering the local mutation characteristics of signals, which may introduce distortion in non-stationary signal processing; third, the feature extraction focuses on a single dimension in time or frequency domain, ignoring the coupling features in time-frequency domain, and the feature optimization lacks damage correlation screening, resulting in high feature dimension and low calculation efficiency, which cannot be directly used for surrounding rock damage grade evaluation. These defects make it difficult for traditional technology to meet the real-time and accurate monitoring needs of deep roadway. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art and provide a method and system for acoustic emission signal noise reduction and feature extraction suitable for deep roadway. The method comprises the following steps: acquiring acoustic emission and environmental data through a multi-source signal acquisition module, using an environmental correction energy entropy and dynamic feature matching algorithm to realize accurate identification of multi-source interference, combining an adaptive hybrid noise reduction mechanism to suppress noise while preserving signal mutation characteristics, using a time-frequency linkage feature algorithm and Hilbert-Huang transform to extract multi-dimensional coupling features, and constructing a low-dimensional damage feature vector through principal component analysis and mutual information entropy screening.

[0005] To solve the above technical problems, the present application provides the following technical solutions: on the one hand, a method for acoustic emission signal noise reduction and feature extraction suitable for deep roadway, the specific steps of which are as follows:

[0006] Multi-source signal acquisition: Along the axis of the deep roadway, an array of acoustic emission sensors is arranged to collect multi-source signals. After processing by the signal conditioning circuit, the signals are converted into digital signals and stored locally. This provides a multi-dimensional data source for subsequent interference identification, noise reduction, and feature extraction.

[0007] Multi-source interference identification: The digital signals are preprocessed and wavelet packet decomposed. The weighted wavelet packet energy entropy is calculated by the environmental correction energy entropy algorithm combined with environmental parameters. The peak frequency and statistical characteristics of each frequency band are extracted. The similarity is calculated by the dynamic feature matching algorithm combined with the interference feature database to determine the interference type.

[0008] Adaptive hybrid noise reduction: Based on the weighted wavelet packet energy entropy, the adaptive penalty factor of improved variational mode decomposition is determined by the complexity-sensitive penalty algorithm. The target components are selected by traversing different modal number decomposed signals. The real-time signal-to-noise ratio and signal mutation value of the target components are calculated. The Kalman filter gain is dynamically adjusted by the mutation response filtering algorithm. The sampling point filtering window is set to filter the noise reduction signal.

[0009] Multi-dimensional feature extraction: Time domain and frequency domain features are extracted from the noise reduction signal. The coupling feature parameters are calculated by the time-frequency linkage feature algorithm. The time-frequency domain features are extracted by the Hilbert-Huang transform.

[0010] Feature optimization and output: The extracted features are reduced by principal component analysis, and the principal components with cumulative contribution rate exceeding 90% are retained. The mutual information entropy of the principal components and the surrounding rock damage grade is calculated, and the features with mutual information entropy value greater than 0.6 are retained. The feature vector is constructed by the dynamic damage vector algorithm and transmitted to the safety warning system. The results can be visualized and stored.

[0011] Contribution rate calculation: The contribution rate of a single principal component is: where: is the contribution rate of the th principal component, is the eigenvalue of the th principal component, is the total number of principal components, and the cumulative contribution rate is the sum of the first principal components, and the principal components with cumulative contribution rate ≥ 90% are retained.

[0012] Mutual information entropy calculation: Based on the principal component and surrounding rock damage grade samples, the mutual information entropy is calculated according to the formula: where: is the mutual information entropy of the principal component and the surrounding rock damage grade , is the marginal probability of the principal component taking the value of . is the damage level take the edge probability of the value, is the principal component take and the joint probability of the damage level take the principal component of the reserved .

[0013] Further, in the multi-source signal collection, the signal transmission lines of adjacent sensors of the acoustic emission sensor array adopt a star topology structure to be connected to a signal collection box, the signal collection box is internally provided with a multi-channel synchronous collection unit, and an anti-aliasing filter is independently configured for each channel.

[0014] Still further, in the multi-source interference identification, the mathematical expression of the environmental correction energy entropy algorithm is: wherein, is the weighted wavelet packet energy entropy, is the total number of frequency bands after wavelet packet decomposition, is the energy of the i-th wavelet packet decomposition frequency band, is the signal energy of the j-th frequency band, is the environmental correction weight, is the environmental correction weight corresponding to the frequency band, the environmental correction weight is determined through laboratory simulation of a typical environment of a deep roadway, standard acoustic emission signals under different environmental parameters are collected, the distortion degree of a single environmental parameter on a signal is calculated , the weight of the single environmental parameter is calculated according to the formula, the formula is: , is the type of environmental parameter, is the total number of environmental parameters participating in the calculation, is the weight of the i-th environmental parameter, , the values of different frequency bands and different environmental parameters are different.

[0015] Still further, in the multi-source interference identification, the interference feature database comprises three types of basic interference feature templates and a dynamic updating module: the basic interference feature templates include mechanical vibration templates, electromagnetic radiation templates and environmental noise templates, wherein the mechanical vibration templates store vibration feature parameters in the frequency band of 50-200 Hz; the electromagnetic radiation templates store electromagnetic feature parameters in the frequency band of 1-5 kHz, including the pulse width, repetition frequency and spectrum distribution of high-voltage cables and lighting equipment; and the environmental noise templates store random noise feature parameters in a wide frequency band, including the energy distribution and amplitude probability density of wind flow disturbance and rock mass friction.

[0016] Still further, in the multi-source interference identification, the mathematical expression of the dynamic feature matching algorithm is: ,in, The measured signal is the first The first in the frequency band component and interference characteristic database Similarity of interfering features, The k-th frequency band component of the measured signal. For the k-th type of interference feature in the interference feature database, For covariance calculation, The measured signal is the first Frequency band components standard deviation It is the first Interference-like features standard deviation , These are the weighting coefficients, and , The measured signal is the first Frequency band components peak frequency, It is the first Interference-like features Peak frequency, comparing overall similarity Find the size of the comprehensive similarity and then determine the type of interference the corresponding interference feature belongs to.

[0017] Furthermore, in the adaptive hybrid noise reduction, the mathematical expression for the complexity-sensitive penalty algorithm is: ,in, It is an adaptive penalty factor for improving variational mode decomposition. Basic penalty factor, The weighted wavelet packet energy entropy of the current signal. This represents the historical maximum energy entropy at this monitoring point. Adjustment factor, adjustment factor The rules for determining and adjusting the value are as follows: The value range is 0.1-0.5, used to balance the accuracy and stability of signal decomposition. The specific value is adjusted according to the surrounding rock fissure ratio: when the surrounding rock fissure ratio... hour When the surrounding rock fracture ratio is 25%, Adjustment criteria: If the signal-to-noise ratio after decomposition is <15dB, increase γ; if the peak signal loss is >10%, decrease γ, until the signal-to-noise ratio is ≥15dB and the peak retention rate is ≥90%.

[0018] Furthermore, in the adaptive hybrid noise reduction, the mathematical expression of the mutation response filtering algorithm is: ,in, yes Kalman filter gain at time t. This is the mutation sensitivity coefficient, ranging from 0.3 to 0.8, when mechanical vibration interference is the primary disturbance. When electromagnetic radiation interference is the main factor This is used to control the sensitivity of the filter to sudden changes in signal. This is the reference signal-to-noise ratio (SNR), which is the average SNR of 10 sets of acoustic emission signals collected when the monitoring point is idle and free from interference. It is used to calibrate the real-time SNR. This is the initial Kalman filter gain, with values ​​ranging from 0.1 to 0.3 as the initial filter parameters. If the signal distortion after filtering is greater than 5%, then decrease it. If noise suppression is insufficient, it will increase. , Rise time, physically defined as the time it takes for an acoustic emission signal to rise from 10% of its peak amplitude to 90% of its peak amplitude. It refers to the duration, which physically means the total time from when the acoustic emission signal starts at 10% of its peak amplitude at the rising edge to when it ends at 10% of its peak amplitude at the falling edge. yes The signal amplitude at any given time is the real-time acquired acoustic emission signal voltage value, used to calculate the signal abrupt change value. , yes The signal mutation value at time.

[0019] Furthermore, in the multi-dimensional feature extraction, the expression for the time-frequency linkage feature algorithm is: ,in, These are time-frequency linkage coupling characteristic parameters. Peak amplitude, Peak frequency, For the rising time, For duration.

[0020] Furthermore, the mathematical expression for the dynamic damage vector algorithm is: ,in, It is a dynamic damage vector. These are time-frequency linkage coupling characteristic parameters. The total energy of the signal. It is the duration of the noise-reduced signal. It is a frequency variable. For the marginal spectral density function, For maximum analysis frequency, This refers to the damage level of the surrounding rock, with a value ranging from 1 to 5. No damage; Minor injury; Moderate damage; Severe damage; Extremely severe damage; It is a time-frequency linkage coupling characteristic parameter Damage level of surrounding rock The mutual information entropy is calculated in the same way as the mutual information entropy of principal components and surrounding rock damage level.

[0021] On the other hand, a system for acoustic emission signal denoising and feature extraction suitable for deep tunnels includes:

[0022] Multi-source signal acquisition module: An array of acoustic emission sensors is deployed along the axial direction of the deep tunnel to convert the acquired signals into digital signals for storage after conditioning;

[0023] Multi-source interference identification module: preprocesses and decomposes digital signals into wavelet packets, calculates weighted wavelet packet energy entropy based on environmental parameters, extracts peak frequencies and statistical features of each frequency band, calculates similarity based on interference feature database, and determines interference type;

[0024] Preprocessing of digital signals includes the following steps: Baseline correction: The signal baseline is fitted using a 5th-order polynomial, with the following formula: ,in, for The baseline signal at time [time]. For time, For polynomial fitting coefficients, outlier removal: based on the 3σ criterion, outliers exceeding the limit are removed. The pulse, where: The average value of the signal amplitude. The standard deviation of the signal amplitude; normalization: scaling the signal amplitude to the interval [-1, 1], the formula is: ,in: for The signal amplitude after time normalization. for The original signal amplitude at time. This represents the maximum amplitude of the original signal. This represents the minimum amplitude of the original signal;

[0025] The adaptive hybrid noise reduction module determines the adaptive penalty factor based on the weighted wavelet packet energy entropy, decomposes the signal and filters the target components, calculates the real-time signal-to-noise ratio and signal mutation value, dynamically adjusts the Kalman filter gain and filters to obtain the noise-reduced signal;

[0026] Multi-dimensional feature extraction module: Extracts time-domain and frequency-domain features from the denoised signal, calculates coupling feature parameters, and uses Hilbert-Huang transform to extract time-frequency domain features;

[0027] Feature optimization and output module: Performs principal component analysis to reduce dimensionality and mutual information entropy to filter the extracted features, constructs a dynamic damage vector and transmits it to the safety early warning system, and visualizes the results and stores the data.

[0028] Compared with the prior art, the acoustic emission signal denoising and feature extraction method and system suitable for deep roadway have the following beneficial effects:

[0029] Firstly, the multi-source signal acquisition module comprehensively captures the acoustic emission signals and environmental parameters in the roadway, providing a rich data basis for subsequent analysis.

[0030] Secondly, the time-frequency coupling characteristics of the signal are deeply mined through the time-frequency linkage feature algorithm, and the time-frequency domain features are further extracted through the Hilbert-Huang transform, realizing a comprehensive characterization of the signal characteristics.

[0031] Other advantages, objects and features of the present application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following or can be learned from practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced as follows.

[0033] Figure 1 The acoustic emission signal denoising and feature extraction method suitable for deep roadway has the following steps:

[0034] Figure 2 The acoustic emission signal denoising and feature extraction system suitable for deep roadway has the following module interaction flow chart. DETAILED DESCRIPTION

[0035] To further illustrate the technical means and effects taken by the present application to achieve the predetermined object of the invention, the specific embodiments, structures, features and effects thereof according to the present application are described in detail below in conjunction with the drawings and preferred embodiments.

[0036] Embodiment one:

[0037] The embodiment proposes a noise reduction and feature extraction method for acoustic emission signals in deep tunnels. The specific steps of the method are as follows:

[0038] Multi-source signal acquisition: Acoustic emission sensor arrays are arranged along the axis of deep tunnels. The collected multi-source signals are processed by a signal conditioning circuit and then converted into digital signals for local storage, providing multi-dimensional data sources for subsequent interference identification, noise reduction, and feature extraction.

[0039] Multi-source interference identification: The digital signals are preprocessed and wavelet packet decomposed. The weighted wavelet packet energy entropy is calculated by an environmental correction energy entropy algorithm combined with environmental parameters. The peak frequency and statistical characteristics of each frequency band are extracted. The similarity is calculated by a dynamic feature matching algorithm combined with the interference feature database to determine the interference type.

[0040] The specific process of extracting the peak frequency and statistical characteristics of each frequency band is as follows: For each frequency band component obtained by wavelet packet decomposition, a sliding window method is used to traverse the frequency band signal waveform. The window length is set to 5-8 times the period corresponding to the center frequency of the frequency band. Local peaks in each window are identified by comparing the signal amplitudes within the window. The amplitudes and corresponding time instants of all local peaks are calculated to obtain the mean and standard deviation of the local peaks. Local peaks exceeding the mean plus 2 times the standard deviation are determined as valid peaks of the frequency band. The frequency corresponding to the valid peak is the peak frequency of the frequency band. For each frequency band component, time-domain statistical parameters are calculated, including the mean, variance, peak factor, and pulse factor of the frequency band signal. Frequency-domain statistical parameters are calculated by converting the frequency band signal to the frequency domain using fast Fourier transform. The center frequency, frequency bandwidth, and spectral flatness of the frequency domain signal are calculated. For frequency bands with obvious periodic interference characteristics, additional periodic characteristics of the signal are extracted. The main period of the signal is determined by autocorrelation analysis, and the number of pulses and pulse intervals within the main period are calculated. The peak frequency, time-domain statistical parameters, frequency-domain statistical parameters, and periodic characteristics of each frequency band are integrated to form a feature set for the frequency band, which is used for subsequent matching analysis with the interference feature database.

[0041] Adaptive hybrid noise reduction: Based on the weighted wavelet packet energy entropy, an adaptive penalty factor for improved variational mode decomposition is determined by a complexity-sensitive penalty algorithm. The signal is decomposed into different modes, and the target component is selected. The real-time signal-to-noise ratio and signal mutation value of the target component are calculated. The Kalman filter gain is dynamically adjusted by a mutation response filtering algorithm, and the sampling point filter window is set to filter the noise signal.

[0042] The specific process of calculating the real-time signal-to-noise ratio and the signal mutation value of the target component is as follows: for the real-time signal-to-noise ratio calculation, first, a smooth section without obvious signal mutation in the target component is intercepted as a noise reference section, the amplitude distribution of the signal in this section is counted, and the noise amplitude is determined; then, an effective section containing signal characteristics is selected, the peak amplitude of the signal in the effective section is counted; the ratio of the peak amplitude of the effective section to the amplitude of the noise reference section is taken as the real-time signal-to-noise ratio; for the signal mutation value calculation, the signal amplitudes of the target component are extracted in time sequence, the signal amplitude at the current time is subtracted from the signal amplitude at the previous time, and the absolute value is taken, which is the signal mutation value at the current time; the signal mutation value at each time point is continuously calculated to form a signal mutation sequence, and the time interval is consistent with the sampling period of the target component;

[0043] Multi-dimensional feature extraction: time domain and frequency domain features are extracted from the denoised signal; the time domain and frequency domain features are integrated to calculate the coupling feature parameters through the time-frequency linkage feature algorithm; the Hilbert-Huang transform is used to extract the time-frequency domain features;

[0044] The specific steps of extracting the time-frequency domain features by the Hilbert-Huang transform are as follows: the denoised signal is subjected to empirical mode decomposition, the local extreme points in the denoised signal are first identified, the local maximum points are connected to form an upper envelope line through cubic spline interpolation, and the local minimum points are connected to form a lower envelope line; the mean values of the upper and lower envelope lines are calculated, and the denoised signal is subtracted from the mean value to obtain a preliminary modal component; it is judged whether the component meets the condition of the intrinsic mode function, if not, the component is taken as a new signal to be decomposed, and the above steps of extreme point identification, envelope fitting and mean value calculation are repeated; when the intrinsic mode function meeting the condition is obtained, the component is separated from the denoised signal, and the remaining signal is taken as a new signal to be decomposed, and the above decomposition process is repeated until the remaining signal cannot be decomposed into the intrinsic mode function, obtaining a plurality of intrinsic mode components and one residual component; the Hilbert transform is performed on each intrinsic mode component to calculate the instantaneous amplitude and instantaneous frequency of each component; the time-frequency distribution of each intrinsic mode component is constructed with time as the horizontal coordinate, frequency as the vertical coordinate, and instantaneous amplitude as the amplitude; the time-frequency distributions of all intrinsic mode components are superimposed to obtain the overall time-frequency distribution of the denoised signal; the overall time-frequency distribution is integrated according to the time dimension to obtain the marginal spectral density; the instantaneous frequency and instantaneous amplitude corresponding to each time in the overall time-frequency distribution are extracted as the time-frequency domain features;

[0045] Feature optimization and output: the extracted features are reduced in dimension through principal component analysis, and the principal components whose cumulative contribution rate exceeds 90% are retained; the mutual information entropy of the principal components and the surrounding rock damage grade is calculated, and the features whose mutual information entropy value is greater than 0.6 are retained; the feature vector is constructed through the dynamic damage vector algorithm and transmitted to the safety warning system, and the results can be visualized and the data can be stored;

[0046] The specific process of extracting features by principal component analysis dimension reduction and calculating mutual information entropy to screen features is as follows: all extracted feature parameters are collected, including time domain features, frequency domain features, coupling feature parameters and time-frequency domain features, to form a feature matrix, each feature parameter in the feature matrix is standardized to eliminate the influence of different dimensions on the analysis results, the covariance matrix of the standardized feature matrix is calculated, the eigenvalues and corresponding eigenvectors of the covariance matrix are obtained by eigenvalue decomposition, the eigenvectors are sorted in descending order of eigenvalues, the contribution rate of each eigenvalue to the total sum of all eigenvalues is calculated, and the cumulative contribution rate is calculated, the eigenvectors with high eigenvalues are selected in turn, until the cumulative contribution rate of the selected eigenvectors exceeds 90%, and the features corresponding to these eigenvectors are determined as the principal components after dimension reduction; a surrounding rock damage grade sample set is established, which contains principal component data corresponding to different damage grades, and the number of samples of each damage grade is not less than 50 groups; for each principal component, the mutual information entropy between it and the surrounding rock damage grade is calculated, the probability distribution of the principal component under different damage grades is counted, and the mutual information entropy value is calculated according to the probability distribution; compare the mutual information entropy values of each principal component, retain the principal components with mutual information entropy values greater than 0.6, and eliminate the principal components with mutual information entropy values less than or equal to 0.6, to form the key features finally used to construct the feature vector.

[0047] In the multi-source signal collection, the layout of the acoustic emission sensor array is as follows: each group of acoustic emission sensors contains 3 radial sensors and 2 axial sensors, the radial sensors are installed at the center of the roadway vault and 1.5 m away from the floor on the left and right sides, and the axial sensors are installed at the hance positions 3 m and 5 m behind the heading face; the sensors are connected to the surrounding rock surface through an epoxy resin-based coupling agent, a stainless steel protective shell is provided outside the sensor, and the gap between the protective shell and the surrounding rock is filled with sound-absorbing material; the signal transmission lines of adjacent sensors in the array are connected to the signal collection box in a star topology, the collection box is built-in with an 8-channel synchronous acquisition unit, the synchronization error between channels is less than or equal to 1 μs, and each channel is independently configured with an anti-aliasing filter with a cutoff frequency of 1.2 MHz.

[0048] Specifically, taking the deep coal mine roadway surrounding rock monitoring scene as an example.

[0049] In a -800 m level recovery roadway of a certain deep coal mine, an acoustic emission sensor array is arranged every 8 m along the axial direction of the roadway, each group of array contains 3 radial sensors and 2 axial sensors, wherein the radial sensors are installed at the center of the roadway vault and 1.5 m away from the floor on the left and right sides, and the axial sensors are installed at the hance positions 3 m and 5 m behind the heading face, the sensors are tightly connected to the surrounding rock surface through an epoxy resin-based coupling agent, a stainless steel protective shell is provided outside the sensor, and the gap between the protective shell and the surrounding rock is filled with sound-absorbing material to reduce environmental interference.

[0050] During the acquisition process, the environmental parameters such as temperature, humidity, air pressure in the tunnel, and the engineering data such as the advancing speed of the tunneling machine and the supporting pressure of the hydraulic support are synchronously acquired. The acoustic emission signals collected by the sensors are processed by the signal conditioning circuit, converted into digital signals, transmitted to the acquisition box through the signal transmission line connected in star topology, and stored locally after being processed by the 8-channel synchronous acquisition unit in the acquisition box. The synchronization error between channels is less than or equal to 1 μs, and each channel is independently configured with an anti-aliasing filter with a cutoff frequency of 1.2 MHz to ensure the accuracy and synchronization of signal acquisition. Finally, all digital signals are stored locally to provide a data basis for subsequent processing, as shown in Figure 1 .

[0051] Firstly, the stored digital signals are preprocessed to remove baseline drift and abnormal pulses in the signals. Then, the signals are decomposed into multiple frequency band components through wavelet packet decomposition. Combined with the environmental parameters such as temperature, humidity, and air pressure collected in the early stage, the weighted wavelet packet energy entropy is calculated through the environmental correction energy entropy algorithm. The mathematical expression of the environmental correction energy entropy algorithm is: wherein, is the weighted wavelet packet energy entropy, is the total number of frequency bands after wavelet packet decomposition, is the energy of the i-th wavelet packet decomposition frequency band, is the environmental correction weight, is the signal energy of the j-th frequency band, and wj is the environmental correction weight corresponding to the frequency band, which realizes the correction of signal energy distribution under different environmental influences.

[0052] For each frequency band component obtained by wavelet packet decomposition, the sliding window method is used to traverse the signal waveform, and the window length is set to 6 times the period corresponding to the center frequency of the frequency band. By comparing the signal amplitudes in the window, the local peaks in each window are identified, and the amplitudes and corresponding time instants of all local peaks are counted. The mean and standard deviation of the local peaks are calculated, and the local peaks exceeding the mean plus 2 times the standard deviation are determined as the effective peaks of the frequency band. The frequency corresponding to the effective peak is the peak frequency of the frequency band.

[0053] At the same time, the time domain statistical parameters and frequency domain statistical parameters of each frequency band component are calculated. For frequency bands with obvious periodic interference characteristics, the main period of the signal is determined through autocorrelation analysis, and the number of pulses and pulse intervals in the main period are counted. These parameters are integrated to form the feature set of each frequency band.

[0054] Finally, combined with the interference feature database, the similarity between the features of each frequency band of the measured signal and the interference features in the database is calculated through the dynamic feature matching algorithm. The mathematical expression of the dynamic feature matching algorithm is: wherein, is the i-th frequency band component of the measured signal, is the j-th frequency band component in the interference feature database. the similarity of the interference feature, is the kth frequency band component of the measured signal, is the kth interference feature in the interference feature database, is the covariance calculation, is the standard deviation of the kth frequency band component of the measured signal, is the standard deviation of the kth interference feature, , is the weight coefficient, and , is the peak frequency of the kth frequency band component of the measured signal, is the peak frequency of the kth interference feature, The size of the comprehensive similarity is compared, the maximum comprehensive similarity is found, and the corresponding interference feature belongs to an interference type. The interference type is determined according to the similarity result, such as mechanical vibration interference generated by a heading machine running, electromagnetic radiation interference generated by a high-voltage cable, and the like.

[0055] Based on the weighted wavelet packet energy entropy obtained in the multi-source interference identification stage, the adaptive penalty factor of the improved variational mode decomposition is determined through a complexity-sensitive penalty algorithm. The mathematical expression of the complexity-sensitive penalty algorithm is: wherein, is the adaptive penalty factor of the improved variational mode decomposition, is a basic penalty factor, is the weighted wavelet packet energy entropy of the current signal, is the historical maximum energy entropy of the monitoring point, is an adjustment coefficient, so as to balance the accuracy and stability of signal decomposition. Subsequently, the signal is decomposed by traversing 3-8 mode numbers. According to the energy distribution of each mode component and the interference feature, a target component containing effective information of rock acoustic emission is screened out.

[0056] The real-time signal-to-noise ratio and signal mutation value of the screened target component are calculated. When the real-time signal-to-noise ratio is calculated, the smooth section without obvious signal mutation in the target component is intercepted as a noise reference section, the noise amplitude of the section is counted, and then the peak amplitude of the section containing effective signal is counted. The ratio of the two is taken as the real-time signal-to-noise ratio. When the signal mutation value is calculated, the amplitude values of the target component are extracted in time sequence, the difference between the current time and the previous time is taken as the absolute value, and the signal mutation sequence is formed.

[0057] ​​​​​​​​According to the real-time signal-to-noise ratio and the signal mutation value, the Kalman filter gain is dynamically adjusted through the mutation response filtering algorithm, and the mathematical expression of the mutation response filtering algorithm is: wherein, is the Kalman filter gain at the moment t, is the initial gain, is the signal mutation value at the moment t, is the signal mean value, is the real-time signal-to-noise ratio, is the reference signal-to-noise ratio, is the mutation sensitivity coefficient, a filter window of 10 sampling points is set to filter the target component, the residual interference is effectively suppressed, and finally the denoised acoustic emission signal is obtained. The time-domain features and the frequency-domain features are extracted from the denoised signal, then the time-domain features and the frequency-domain features are integrated, the coupling feature parameters are calculated through the time-frequency linkage feature algorithm, and the expression of the time-frequency linkage feature algorithm is:

[0058] wherein, is the time-frequency linkage coupling feature parameter, is the peak amplitude, is the peak frequency, is the rise time, is the duration, which can comprehensively reflect the correlation characteristics of the signal in the time and frequency dimensions. Meanwhile, the time-frequency domain features are extracted by using the Hilbert-Huang transform: the denoised signal is first subjected to empirical mode decomposition, the local extreme points in the signal are identified, the local maximum points are connected to form the upper envelope line through cubic spline interpolation, the local minimum points are connected to form the lower envelope line, the mean values of the upper and lower envelope lines are calculated and subtracted from the original signal to obtain the preliminary modal component; it is judged whether the component meets the intrinsic mode function condition, if not, the above decomposition steps are repeated until a plurality of intrinsic mode components and one residual component are obtained.

[0059] The Hilbert transform is performed on each intrinsic mode component to calculate the instantaneous amplitude and the instantaneous frequency, the time-frequency distribution of each component is constructed with time as the horizontal coordinate, frequency as the vertical coordinate and instantaneous amplitude as the amplitude, the overall time-frequency distribution is obtained by superimposing the time-frequency distributions of all components, and the marginal spectral density is obtained by integrating the overall time-frequency distribution in the time dimension, and finally the instantaneous frequency and the instantaneous amplitude of each moment in the overall time-frequency distribution are extracted as the time-frequency domain features.

[0060]

[0061] ​The extracted time domain features, frequency domain features, coupling feature parameters and time-frequency domain features are collected to form a feature matrix. Each feature parameter in the matrix is normalized to eliminate the influence of different dimensions. The covariance matrix of the normalized feature matrix is calculated. The eigenvalues and corresponding eigenvectors are obtained by eigenvalue decomposition. The contribution rate and cumulative contribution rate of each eigenvalue are calculated. The features corresponding to the eigenvectors with cumulative contribution rate exceeding 90% are selected as the principal components after dimension reduction.

[0062] A sample set of surrounding rock damage levels is established. For each principal component, the probability distribution under different damage levels is calculated. The correlation between the principal component and the surrounding rock damage level is calculated by mutual information entropy. The principal components with mutual information entropy greater than 0.6 are retained as key features.

[0063] Finally, the feature vector is constructed by the dynamic damage vector algorithm. The mathematical expression of the dynamic damage vector algorithm is: wherein, is the dynamic damage vector, is the time-frequency coupling feature parameter, is the total energy of the signal, is the duration of the noise-reduced signal, is the frequency variable, is the marginal spectral density function, is the mutual information entropy of the feature parameter and the damage level, is the maximum analysis frequency. The feature vector is transmitted to the coal mine roadway safety warning system. The surrounding rock damage features, interference identification results and noise reduction effects are displayed through a visual interface. All processed data is stored to provide a basis for subsequent roadway support adjustment and safety risk assessment.

[0064] In summary, the method embodiment of the present application completes the acoustic emission signal processing in the -800m deep coal mine roadway scene according to the steps. First, a specific sensor array is arranged every 8m along the roadway axis. Multi-source signals are synchronously collected and standardized. Then, the pre-processing and wavelet packet decomposition are performed. The environmental correction energy entropy algorithm and the dynamic feature matching algorithm are used to identify interference. Next, the complexity-sensitive penalty algorithm and the sudden response filtering algorithm are used to realize adaptive hybrid noise reduction. Then, multi-dimensional features are extracted. Finally, the principal component analysis dimension reduction and mutual information entropy screening and optimization are used to construct the dynamic damage vector, which is transmitted to the warning system to provide reliable method support for roadway safety assessment and support adjustment.

[0065] Embodiment two:

[0066] Metal mine deep roadway stability monitoring scene.

[0067] In a transport roadway of a certain metal mine at -1200m level, an acoustic emission sensor array is arranged along the axial direction of the roadway every 5-10m (according to the geological conditions of the roadway, 5m interval is arranged in the fault fracture zone, and 10m interval is arranged in the complete rock mass region). The sensor installation mode, protection measures and signal transmission structure of each array are consistent with the method embodiment, that is, three radial sensors are respectively located at the center of the vault and the left and right sides 1.5m high, two axial sensors are located at the arch waist 3m and 5m behind the head, the sensors are connected to the surrounding rock through the epoxy resin-based coupling agent, the external stainless steel protective shell is filled with sound-absorbing material, and the signal is transmitted to the acquisition box containing 8-channel synchronous acquisition unit and 1.2MHz cutoff frequency anti-aliasing filter through star topology structure.

[0068] The module can synchronously collect acoustic emission signals, environmental parameters such as temperature, humidity and air pressure in the roadway, and engineering data such as tunneling progress and anchor support pressure. After the collected analog signals are converted into digital signals by the conditioning circuit, they are stored in the local data server, supporting real-time calling and historical backtracking of data, such as Figure 2 .

[0069] The multi-source interference identification module receives the digital signals transmitted by the multi-source signal acquisition module, first executes a preprocessing program to remove baseline drift and outliers, and then starts a wavelet packet decomposition function to decompose the signals into multiple frequency band components. Combined with the temperature, humidity and air pressure data transmitted by the environmental parameter acquisition sub-module, the environmental correction energy entropy algorithm is called to calculate the weighted wavelet packet energy entropy, realizing the quantitative correction of environmental interference.

[0070] Subsequently, the module automatically executes a feature extraction process for each frequency band component: local peak values are identified and peak frequencies are determined by the sliding window method, time domain statistical parameters and frequency domain statistical parameters are calculated, and the main period, pulse number and interval are additionally extracted for the periodic interference frequency band to form a frequency band feature set.

[0071] The module has a built-in interference feature database containing mechanical vibration templates, electromagnetic radiation templates and environmental noise templates adapted to the metal mine scene, and has a dynamic updating function. New interference signal templates can be supplemented according to newly collected interference signals.

[0072] Finally, the module calculates the similarity between the measured signal features and the database templates by a dynamic feature matching algorithm, determines the interference type according to the preset similarity threshold, and generates an interference identification report to mark the main interference source and its affected frequency band.

[0073] Similarity threshold: the threshold is set , as the similarity determination critical value. After 50 known interference samples are calibrated, the determination process is as follows: the similarity between the measured signal and the mechanical vibration, electromagnetic radiation and environmental noise interference feature templates is calculated . 、 , is the similarity with the mechanical vibration template, is the similarity with the electromagnetic radiation template, is the similarity with the environmental noise template, the maximum similarity is taken If , it is determined as corresponding type of interference; if , it is determined as unknown interference, and the dynamic updating of the interference feature database is triggered.

[0074] The adaptive hybrid noise reduction module receives the weighted wavelet packet energy entropy and the interference determination result output by the multi-source interference recognition module. First, the complexity-sensitive penalty algorithm is called, and the adaptive penalty factor of the improved variational mode decomposition is calculated according to the weighted wavelet packet energy entropy of the current signal and the historical maximum energy entropy of the monitoring point, to ensure that the decomposition process adapts to the current signal complexity.

[0075] The variational mode decomposition program is started, and the signal is decomposed by traversing 3-8 modal numbers. Combined with the interference recognition result, the target component containing effective information of the surrounding rock acoustic emission without main interference is selected. The real-time signal-to-noise ratio and signal mutation value of the target component are calculated: the stable segment is automatically intercepted as the noise reference segment, and the real-time signal-to-noise ratio is obtained by comparing the peak amplitude of the effective segment; the absolute value of the amplitude difference of adjacent time series is calculated to generate a signal mutation sequence.

[0076] According to the real-time signal-to-noise ratio and the signal mutation value, the mutation response filtering algorithm is used to dynamically adjust the Kalman filtering gain, and the filtering window of 5-15 sampling points is set to filter the target component. Finally, the acoustic emission signal after noise reduction is output, and the filtering parameters and noise reduction effect evaluation indexes are recorded.

[0077] The multi-dimensional feature extraction module receives the noise reduction signal from the adaptive hybrid noise reduction module. First, the time domain feature extraction submodule calculates the peak value, mean value, variance, peak factor, pulse factor and other parameters of the signal; the frequency domain feature extraction submodule performs fast Fourier transform on the signal to obtain the center frequency, frequency bandwidth, spectral flatness and other frequency domain parameters.

[0078] The time-frequency coupling feature algorithm is called to integrate the time domain and frequency domain feature calculation coupling feature parameters, which comprehensively reflect the collaborative characteristics of the signal in the time and frequency dimensions. At the same time, the Hilbert-Huang transform submodule is started, and the intrinsic mode components and residual components are obtained according to the empirical mode decomposition process. The Hilbert transform is performed on each intrinsic mode component to obtain the instantaneous amplitude and instantaneous frequency, and the time-frequency distribution is constructed and superimposed to calculate the marginal spectral density. Finally, the instantaneous frequency, instantaneous amplitude and time domain, frequency domain, coupling features are extracted as a multi-dimensional feature set and stored in the feature database.

[0079] The feature optimization and output module reads a multi-dimensional feature set from a feature database, constructs a feature matrix and completes standardization processing, calls a principal component analysis algorithm, calculates a covariance matrix, eigenvalues and eigenvectors, selects principal components with a cumulative contribution rate exceeding 90% according to the eigenvalue order, establishes a metal mine surrounding rock damage grade sample set, calculates the correlation degree of the principal components and the damage grade through mutual information entropy, and retains key features with a mutual information entropy value greater than 0.6.

[0080] The feature vector is constructed through a dynamic damage vector algorithm, is transmitted to a metal mine deep roadway stability monitoring platform, when the surrounding rock damage risk reflected by the feature vector exceeds a preset threshold, the platform automatically triggers a warning signal, meanwhile, the module displays a signal comparison chart before and after noise reduction, a time-frequency distribution chart, a damage feature parameter trend chart and the like through a visual terminal, and stores all processing data to a cloud database, supports remote query and historical data tracing of mine management personnel, and provides systematic support for roadway support scheme optimization and safety production decision-making.

[0081] In summary, the system embodiment is applied to a metal mine -1200m transport roadway, through the cooperative work of the five modules, the multi-source signal acquisition module flexibly arranges sensors according to geological conditions and acquires and stores data; the multi-source interference identification module processes signals and determines interference in combination with the database; the adaptive hybrid noise reduction module dynamically adjusts parameters to complete noise reduction; the multi-dimensional feature extraction module extracts various features; and the feature optimization and output module optimizes features and then transmits to the monitoring platform, and can also visually display results and store data, thereby providing comprehensive and systematic support for mine roadway stability monitoring and safety production decision-making.

[0082] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above with a preferred embodiment, it is not intended to limit the present application. Any person skilled in the art can make some changes or modifications to the above disclosed technical content to obtain equivalent embodiments with equivalent changes, without departing from the technical solution of the present application. Any modification, change, equivalent change and modification of the above embodiments, which does not depart from the technical solution of the present application, is still within the scope of the present application.

Claims

1. A method for acoustic emission signal noise reduction and feature extraction applicable to deep tunnels, characterized in that, The specific steps of this method are as follows: Multi-source signal acquisition: An array of acoustic emission sensors is deployed along the axial direction of the deep tunnel. The acquired multi-source signals are processed by a signal conditioning circuit and converted into digital signals for local storage. Multi-source interference identification: The digital signal is preprocessed and decomposed into wavelet packets. The weighted wavelet packet energy entropy is calculated using the environment-corrected energy entropy algorithm, combined with environmental parameters. The mathematical expression for the environment-corrected energy entropy algorithm is: ,in, It is the weighted wavelet packet energy entropy. It is the total number of frequency bands after wavelet packet decomposition. Let i be the energy of the frequency band decomposed from the i-th wavelet packet. It is the signal energy of the j-th frequency band. Adjust the weights for the environment. This refers to the environmental correction weights corresponding to this frequency band. The environmental correction weights are determined through laboratory simulation of typical environments in deep tunnels. Standard acoustic emission signals under different environmental parameters are collected, and the distortion of the signal by a single environmental parameter is calculated. The weights of individual environmental parameters are calculated using the following formula: , x is the type of environmental parameter, m is the total number of environmental parameters involved in the calculation, w x The weights of environmental parameters of type x are w for different frequency bands and different environmental parameters. i w j Different values ​​are taken; the peak frequency and statistical characteristics of each frequency band are extracted, and the similarity is calculated by combining the interference feature database with a dynamic feature matching algorithm to determine the type of interference; Adaptive Hybrid Denoising: Based on weighted wavelet packet energy entropy, an adaptive penalty factor for the improved variational mode decomposition is determined using a complexity-sensitive penalty algorithm. The mathematical expression of the complexity-sensitive penalty algorithm is as follows: Where S is the adaptive penalty factor for the improved variational mode decomposition, and S0 is the basic penalty factor. H is the weighted wavelet packet energy entropy of the current signal. max The historical maximum energy entropy of the monitoring point, To adjust the coefficients, the signals are decomposed by different modes and the target components are selected. The real-time signal-to-noise ratio and signal mutation value are calculated for the target components. The Kalman filter gain is dynamically adjusted by the mutation response filtering algorithm, and the sampling point filtering window is set to obtain the denoised signal. Multi-dimensional feature extraction: Extracting time-domain and frequency-domain features from the denoised signal; integrating time-domain and frequency-domain features to calculate coupled feature parameters using a time-frequency linkage feature algorithm; and extracting time-frequency domain features using Hilbert-Huang transform. Feature optimization and output: The extracted features are dimensionality reduced by principal component analysis, and principal components with a cumulative contribution rate of more than 90% are retained; the mutual information entropy between the principal components and the surrounding rock damage level is calculated, and features with a mutual information entropy value greater than 0.6 are retained; feature vectors are constructed by dynamic damage vector algorithm, transmitted to the safety early warning system, and the results are visualized and the data is stored.

2. The method for acoustic emission signal denoising and feature extraction applicable to deep tunnels according to claim 1, characterized in that, In the multi-source signal acquisition, the signal transmission lines of adjacent sensors in the acoustic emission sensor array are connected to the signal acquisition box in a star topology. The acquisition box has a built-in multi-channel synchronous acquisition unit, and each channel is independently configured with an anti-aliasing filter.

3. The method for acoustic emission signal denoising and feature extraction applicable to deep tunnels according to claim 1, characterized in that, In the multi-source interference identification, the interference feature database includes three types of basic interference feature templates and a dynamic update module: the basic interference feature templates include mechanical vibration templates, electromagnetic radiation templates, and environmental noise templates. The mechanical vibration template stores vibration feature parameters in the 50-200Hz frequency band; the electromagnetic radiation template stores electromagnetic feature parameters in the 1-5kHz frequency band, including the pulse width, repetition frequency, and spectral distribution of high-voltage cables and lighting equipment; and the environmental noise template stores broadband random noise feature parameters, including the energy distribution and amplitude probability density of wind disturbance and rock friction.

4. The method for acoustic emission signal denoising and feature extraction applicable to deep tunnels according to claim 1, characterized in that, In the multi-source interference identification, the mathematical expression of the dynamic feature matching algorithm is: , of which S k X is the similarity between the k-th frequency band component of the measured signal and the k-th type of interference feature in the interference feature database. k Y is the k-th frequency band component of the measured signal. k Here, con(.) represents the k-th type of interference feature in the interference feature database, and covariance is calculated. It is the k-th frequency band component X of the measured signal. k standard deviation Y is the k-th type of interference feature k standard deviation , These are the weighting coefficients, and , It is the k-th frequency band component X of the measured signal. k peak frequency, Y is the k-th type of interference feature k The peak frequency is compared with the overall similarity S. k Find the size of the comprehensive similarity and then determine the type of interference the corresponding interference feature belongs to.

5. The method for acoustic emission signal denoising and feature extraction applicable to deep tunnels according to claim 1, characterized in that, In the adaptive hybrid noise reduction, the mathematical expression of the mutation response filtering algorithm is: , where K t Kt represents the Kalman filter gain at time t, and K0 represents the initial reference gain value of the Kalman filter. The value of the signal change at time t. SNR is the signal mean. t For real-time signal-to-noise ratio, SNR ref For reference signal-to-noise ratio, This represents the mutation sensitivity coefficient.

6. The method for acoustic emission signal denoising and feature extraction applicable to deep tunnels according to claim 1, characterized in that, In the multi-dimensional feature extraction, the expression for the time-frequency linkage feature algorithm is: , of which F tc It is a time-frequency linkage coupling characteristic parameter, A peak f is the peak amplitude. peak For the peak frequency, t r For the rise time, T dur For duration.

7. The method for acoustic emission signal denoising and feature extraction applicable to deep tunnels according to claim 1, characterized in that, The mathematical expression for the dynamic damage vector algorithm is: ,in, It is a dynamic damage vector, F tc It is a time-frequency linkage coupling characteristic parameter, E total T represents the total energy of the signal. dur The duration of the noise-reduced signal is given by f, where f is the frequency variable and M(f) is the marginal spectral density function. f is the mutual information entropy between the feature parameters and the damage level. max This represents the maximum analysis frequency.

8. A system for acoustic emission signal denoising and feature extraction in deep tunnels, wherein the system is applicable to the acoustic emission signal denoising and feature extraction method for deep tunnels as described in any one of claims 1-7, characterized in that, The system includes: Multi-source signal acquisition module: An array of acoustic emission sensors is deployed along the axial direction of the deep tunnel to convert the acquired signals into digital signals for storage after conditioning; Multi-source interference identification module: preprocesses and decomposes digital signals into wavelet packets, calculates weighted wavelet packet energy entropy based on environmental parameters, extracts peak frequencies and statistical features of each frequency band, calculates similarity based on interference feature database, and determines interference type; The adaptive hybrid noise reduction module determines the adaptive penalty factor based on the weighted wavelet packet energy entropy, decomposes the signal and filters the target components, calculates the real-time signal-to-noise ratio and signal mutation value, dynamically adjusts the Kalman filter gain and filters to obtain the noise-reduced signal; Multi-dimensional feature extraction module: Extracts time-domain and frequency-domain features from the denoised signal, calculates coupling feature parameters, and uses Hilbert-Huang transform to extract time-frequency domain features; Feature optimization and output module: Performs principal component analysis to reduce dimensionality and mutual information entropy to filter the extracted features, constructs a dynamic damage vector and transmits it to the safety early warning system, and visualizes the results and stores the data.

Images