Acoustic defect detection method based on narrowband screening feature fusion technology

Abstract

The invention relates to an acoustic defect detection method based on a narrow-band screening feature fusion technology, and the method comprises the steps: S1, signal collection and labeling, S2, narrow-band screening, S3, multi-dimensional feature extraction, S4, feature fusion, and S5, defect detection: according to the fusion index distribution of a healthy sample and a defect sample, configuring a dynamic threshold value, and comparing the signal fusion index with a threshold value in real time to judge the state of the equipment. According to the method, self-correlation spectrum kurtosis analysis is carried out on the collected sound signals to obtain the optimal fault frequency band, multi-dimensional features are extracted after signal band-pass filtering, the most important change trend is captured through dimension reduction, defect distinguishing is carried out through threshold value configuration, the defects of mechanical equipment are effectively detected on the basis of a small number of data samples, and the detection accuracy is improved. The actual engineering background is met.

Description

Technical Field

[0001] The present invention belongs to the field of mechanical equipment condition monitoring and fault diagnosis, and specifically relates to an acoustic defect detection method based on narrowband screening and feature fusion, which is applicable to non-contact online detection of equipment such as rotating machinery and bearings. Background Art

[0002] Mechanical equipment plays an important role in modern industry, and in recent years, the monitoring of its condition has received increasing attention. Vibration and noise respond quickly to faults and can reflect rich information of mechanical systems. However, vibration sensors need to be installed in a contact manner and are not suitable for batch detection scenarios. As a non-contact detection method, acoustic-based defect detection is increasingly widely used in the field of mechanical fault diagnosis.

[0003] Traditional health condition detection methods include methods based on qualitative models and methods based on quantitative models. Methods based on qualitative models rely on complete prior knowledge and expert experience, and have poor applicability and scalability; methods based on quantitative models can extract fault features well, but rely on accurate mechanism models.

[0004] With the development of artificial intelligence technology, many intelligent defect detection algorithms have emerged. Such methods do not rely on complete expert knowledge and accurate mechanism models, but through learning and analysis of a large amount of acoustic signal data, extract deep features of the data, and establish a mapping relationship between the data and the equipment state. However, the equipment is prone to interference during operation, and fault information is inevitably submerged, making it difficult to extract effective features; in addition, in practical applications, it is difficult to obtain a large number of complete fault samples, and directly using the signal as the input of the neural network will increase the model calculation amount.

[0005] In summary, the existing technologies mainly have the following deficiencies:

[0006] (1) Traditional methods have poor generalization ability and rely too much on complete prior knowledge and expert experience;

[0007] (2) Artificial intelligence technology has high computational time and space complexity and has high requirements for hardware conditions;

[0008] (3) The accuracy of model training depends on the accuracy of training data labels, and in practical applications, it is difficult to obtain a large number of complete fault samples, and the robustness is weak.

[0009] Traditional defect detection methods rely on expert experience or complex models and have poor generalization ability. For example, Patent CN109596354A optimizes filtering parameters through adaptive frequency band merging, but does not solve the problem of multi-domain feature fusion; Patent CN113340995A selects frequency bands using spectral kurtosis and F_score, but does not combine dimensionality reduction technology. Summary of the Invention

[0010] Aiming at the deficiencies of the existing technology and the problems faced in engineering applications, the present invention proposes an acoustic defect detection method based on a feature fusion technology with narrowband screening. Signals are collected through acoustic sensors and a four-channel data acquisition system, and then autocorrelation spectral kurtosis analysis is performed to obtain the center frequency and bandwidth of the optimal resonance band. Then, the signal is band-pass filtered, and multi-dimensional features of the filtered signal are extracted. By reducing the dimension, the most important change trend in the data is captured to indicate the defect state of the device, realizing the effective detection of mechanical device defects based on a small number of data samples, which is in line with the actual engineering background.

[0011] To achieve the above object, the technical solution of the present invention: an acoustic defect detection method based on a feature fusion technology with narrowband screening, comprising the following steps:

[0012] S1. Signal acquisition and tagging: Collect the sound signal during the operation of the device through an acoustic sensor and mark it as a healthy sample or a defective sample;

[0013] S2. Narrowband screening: Perform autocorrelation spectral kurtosis analysis on the signal to determine the center frequency and bandwidth of the optimal resonance band, and extract the signal of the target frequency band through a band-pass filter;

[0014] S3. Multi-dimensional feature extraction: Extract 26-dimensional features of the signal from the time domain and the frequency domain respectively, including the peak-to-peak value in the time domain, the kurtosis factor, the center frequency of the frequency domain, and the mean square frequency;

[0015] S4. Feature fusion: Reduce the dimension of the multi-dimensional features through principal component analysis, and select the first principal component as the fusion index;

[0016] S5. Defect detection: Configure a dynamic threshold according to the distribution of the fusion indexes of the healthy samples and the defective samples, and compare the signal fusion index with the threshold in real time to determine the device state.

[0017] Further, in the step S2, the autocorrelation spectral kurtosis analysis is specifically: decompose the signal based on the maximum repeated discrete wavelet packet transform, calculate the unbiased autocorrelation kurtosis value of the squared envelope of each layer of the decomposed signal, and select the frequency band corresponding to the maximum kurtosis value as the optimal resonance band.

[0018] Further, the autocorrelation spectral kurtosis analysis includes the following steps:

[0019] Perform maximum repeated discrete wavelet packet transform decomposition on the signal and calculate the unbiased autocorrelation function of the squared envelope:

[0020] [[ID=3�]]

[0021] In the formula, τ = q / f s is the delay factor, fs is the sampling frequency, X is the squared envelope of the decomposed signal, M is the total length of the signal, q = 0, 1, ..., M - 1;

[0022] Calculate the spectral kurtosis value of the signal within the frequency band:

[0023]

[0024] Select the frequency band corresponding to the maximum spectral kurtosis as the optimal resonance frequency band.

[0025] Furthermore, the step S2 further includes: performing band-pass filtering using a Butterworth filter, and its parameters include: passband attenuation αp, stopband attenuation αs, upper passband cut-off frequency ωp, lower stopband cut-off frequency ωs, passband cut-off frequency f P = 4000Hz, stopband cut-off frequency f s = 5500Hz.

[0026] Furthermore, in the step S3, the time-domain features include 13 indicators, and the frequency-domain features include 13 indicators, specifically including peak-to-peak value, impulse factor, waveform factor, frequency-domain centroid frequency, and frequency standard deviation.

[0027] Furthermore, in the step S4, the principal component analysis includes: performing standardization processing on the feature matrix, calculating the covariance matrix, and selecting the first principal component with a cumulative contribution rate greater than 85% as the fusion index.

[0028] Furthermore, in the step S5, the dynamic threshold configuration includes: plotting a line graph of the fusion indices of healthy samples and defective samples, and the user manually selects the threshold according to the line spacing.

[0029] Furthermore, in the step S5, the defect determination condition is: when the fusion index is greater than the threshold, it is determined to be healthy, otherwise it is determined to be defective.

[0030] Furthermore, the installation position of the acoustic sensor is 10 cm away from the mechanical equipment, and the trigger signal controls the start and end of the acquisition, and 65536 data points are acquired each time.

[0031] Furthermore, the method is realized through the cooperation of the upper computer and the lower computer: the lower computer is responsible for signal acquisition, feature calculation, and threshold comparison, and the upper computer is responsible for threshold configuration, data storage, and visualization.

[0032] The beneficial effects of the present invention are:

[0033] The present invention relates to fields such as narrowband screening, feature extraction, data dimensionality reduction, software design, etc. Specifically, it is a method that utilizes the time-frequency features of sound data, realizes acquisition and calculation by the lower computer, and performs configuration distribution, data storage, and display by the upper computer software. It is applicable to most electromechanical equipment, and the specific effects are:

[0034] (1) Strong generalization ability: In the threshold configuration mode, it can select data of specific entries for different models of mechanical equipment, enabling personalized setting of the equipment threshold and enhancing the generalization ability.

[0035] (2) Fast detection: Without relying on a large domain knowledge base or complex neural network training, accurate and fast equipment defect detection is achieved through low-complexity calculations such as feature extraction and PCA dimensionality reduction.

[0036] (3) High robustness of the detection algorithm: The frequency band with the richest defect information is obtained through narrowband screening, and one-dimensional features with rich meanings are obtained through data fusion, increasing the anti-interference ability of the algorithm and improving the robustness of the detection. Description of the Drawings

[0037] Figure 1 is the flowchart of the method of the present invention;

[0038] Figure 2 is the autocorrelation spectral kurtosis analysis result;

[0039] Figure 3 is the index distribution after PCA feature fusion. Detailed Embodiments

[0040] As Figure 1 shown, an acoustic defect detection method based on narrowband screening feature fusion technology of the present invention includes the following steps:

[0041] Step 1: Signal acquisition and tagging

[0042] Collect the sound signal during equipment operation through a data acquisition device and a sound sensor, and the start and end of the acquisition are automatically controlled by the trigger signal. And calibrate the collected small batch of sample signals, which are divided into healthy samples and defect samples.

[0043] Step 2: Narrowband screening

[0044] Defect features are easily masked by noise and low-frequency effects. Therefore, before extracting features from the sound signal, narrowband screening needs to be carried out first to obtain the frequency band with relatively rich defect information, which is beneficial to improving the accuracy of the state diagnosis model. Use the autocorrelation spectral kurtosis method, use the maximum correlation kurtosis as the measurement index, and combine the squared envelope spectrum, comprehensively consider the valuable frequency components and their corresponding energy distributions, and identify the frequency region where the energy is concentrated when the mechanical equipment has defects, and screen out a suitable analysis frequency band for the acoustic defect detection method.

[0045] Step 3: Multi-dimensional feature extraction

[0046] To accurately achieve defect detection of mechanical equipment, it is necessary to effectively and comprehensively extract key feature information from signals. Commonly used features include time-domain features and frequency-domain features. The time-domain features of signals include dimensional features such as mean, standard deviation, effective value, root mean square value, etc. and dimensionless features such as kurtosis, waveform factor, impulse coefficient, margin factor, skewness, etc.; the frequency-domain features of signals obtain the global frequency distribution of signals through Fourier transform, including centroid frequency, mean square frequency, frequency standard deviation, etc.

[0047] The feature description in a single domain is inevitably one-sided. By extracting and fusing multi-domain feature quantities of signals, the limitations of single-domain feature extraction can be effectively overcome, making the extraction of feature parameters more comprehensive, with a small amount of calculation, and applicable to defect detection tasks with high efficiency, speed, and strong generalization ability.

[0048] Step 4: Feature fusion

[0049] Taking multi-dimensional features of signals can describe the characteristics of signals in all aspects, but a larger number of feature parameter dimensions will increase the space cost, and more redundant data or repeated information will also affect the recognition effect and increase the computational burden. Therefore, it is particularly important to achieve effective fusion of multi-dimensional features.

[0050] Principal component analysis (PCA) is used to achieve multi-dimensional feature fusion. Principal component analysis is a data dimensionality reduction technique that projects data onto a new coordinate system through a linear transformation. The basis vectors of the new coordinate system are determined by the eigenvectors of the covariance matrix of the original data. Each basis vector corresponds to a principal component, representing the largest variance in the data. In the present invention, the first principal component is selected as the fused index. In this way, while retaining the original feature attributes and physical meanings of the data after dimensionality reduction, the feature dimension and data redundancy are reasonably reduced, and the detection effect is improved.

[0051] Step 5: Defect detection

[0052] Signals will vary due to different equipment models, working environments, etc. and are reflected in the feature values through calculation. Therefore, the present invention realizes the custom modification of defect thresholds for different mechanical equipment by developing a threshold configuration mode. The user collects a batch of data to participate in the threshold configuration and performs category calibration in the database. The system plots the fusion indexes of healthy samples and defect samples according to the label categories respectively, forming two different broken lines. At this time, the user can select the defect threshold between the broken lines.

[0053] Once a trigger signal is received, the system starts to collect sound data, calculates the multi-dimensional feature fusion index through the lower computer algorithm, and compares it with the set defect threshold. When the health determination condition is met, the device health signal is output; when the health determination condition is not met, the device defect signal is output.

[0054] Embodiment: An acoustic defect detection method based on a feature fusion technology with narrowband screening, comprising the following steps:

[0055] Step 1: Signal acquisition and tagging. Collect the sound signals during equipment operation through a data acquisition device and a sound sensor, and automatically control the start and end of the acquisition by a trigger signal. And calibrate the collected small batch of sample signals, dividing them into healthy samples and defective samples. According to the actual situation on site, place the sound sensor 10 cm away from the mechanical equipment.

[0056] Step 2: Narrowband screening. First, based on the maximum repetitive discrete wavelet packet transform, perform multi-layer decomposition processing on the signal to obtain signals of different layers. Then, calculate the unbiased autocorrelation based on the squared envelope of the decomposed signal, that is

[0057]

[0058] where f s is the sampling frequency; τ = q / f s is the delay factor; X is the squared envelope of the decomposed signal; M is the total length of the signal, q = 0, 1,..., M - 1.

[0059] Calculate the kurtosis value of the unbiased autocorrelation of the first half of all envelope signals. The formula for the kurtosis of the signal within the frequency band is

[0060]

[0061] The frequency band where the component corresponding to the maximum kurtosis value is located is the one sought. It effectively reduces the interference of non-periodic components on the actual fault frequency, thereby improving the accuracy of identifying the optimal frequency band.

[0062] The result of analyzing the data collected in Step 1 is as Figure 2 shown. The maximum kurtosis is 4.8, the center of the selected optimal resonance frequency band is 4864 Hz, and the bandwidth is 512 Hz. After comparing multiple sample data, select 4000 Hz - 5500 Hz as the analysis frequency band, and implement narrowband screening through a band-pass filtering algorithm in the lower computer. Set the Butterworth filter parameters, including the passband attenuation αp, the stopband attenuation αs, the upper passband cut-off frequency ωp, the lower stopband cut-off frequency ωs, etc., to obtain the corresponding Butterworth filter transfer function, and perform narrowband screening on the signal.

[0063] Step 3: Multi-dimensional feature extraction. For the discrete vibration time-domain signal x(n), n = 1, 2,..., N, where N is the number of sampling points, the extracted time-domain features are shown in the following table:

[0064] Table 1 Time-domain features

[0065]

[0066] Here, assume that S(k) is the spectrum of the signal x(n), where k = 1, 2, ..., K, and K is the number of spectral lines, and f k is the frequency value of the k-th spectral line. The extracted frequency-domain features are shown in the following table:

[0067] Table 2 Frequency-Domain Features

[0068]

[0069]

[0070] Among them, there are 13 time-domain indicators and 13 frequency-domain indicators, a total of 26 features.

[0071] Step 4: Feature fusion. First, standardize the feature value matrix X of the samples:

[0072]

[0073] where μ is the mean and σ is the standard deviation.

[0074] Then, calculate the covariance matrix C of the standardized data:

[0075]

[0076] where m is the number of samples.

[0077] Solve for the eigenvalues and corresponding eigenvectors of the covariance matrix C:

[0078] Cv i = λ i v i

[0079] where λ i is the eigenvalue and v i is the corresponding eigenvector.

[0080] According to the order of eigenvalues from largest to smallest, select the eigenvectors corresponding to the first k eigenvalues. These vectors form the eigenvector matrix V.

[0081] V = [v1, v2, ..., v k

[0082] Project the original data onto the selected principal components to obtain the data Y after dimensionality reduction:

[0083] Y = X standardized V

[0084] In this method, select the first principal component with a cumulative contribution rate greater than 85% as the fused index, as shown in the appendix Figure 3 ​As shown, in this way, on the premise of retaining the original feature attributes and physical meanings of the data after dimensionality reduction, the feature dimension and data redundancy are reasonably reduced, and the detection effect is improved.

[0085] Step 5: Defect detection. The user collects a batch of data to participate in threshold configuration and performs category calibration in the database. The system plots the fusion indicators of healthy samples and defective samples according to the label categories respectively, forming two different broken lines. At this time, the user can select the defect threshold between the broken lines. Once a trigger signal is received, the system starts to collect a 65,536-point sound signal and calls an algorithm to calculate the data fusion indicator. Compare it with the preset threshold. When the calculated indicator is greater than the corresponding threshold, it is judged as healthy, otherwise it is judged as defective. As Figure 3 shown, when the threshold is taken as 0, healthy devices and faulty devices can be well distinguished. The user can access the original data and health status of each item of data through the host computer software.

[0086] Example 1: Bearing defect detection

[0087] (1) Signal acquisition: Place the acoustic sensor 10 cm away from the bearing housing, collect acoustic signals with a sampling rate of 50 kHz, and mark 100 groups of healthy and inner ring defect samples each;

[0088] (2) Narrowband screening: Determine the best frequency band as 4864 Hz ± 256 Hz ( Figure 2 ) through autocorrelation spectral kurtosis analysis, and filter using a Butterworth filter;

[0089] (3) Feature extraction: Calculate the time-domain peak-to-peak value (Xpp = 12.5X_{pp} = 12.5Xpp = 12.5V) and the frequency-domain centroid frequency (F2 = 4900F_2 = 4900F2 = 4900 Hz);

[0090] (4) Feature fusion: After PCA dimensionality reduction, the cumulative contribution rate of the first principal component is 88%, the average value of the fusion indicator of healthy samples is 1.2, and that of defective samples is -0.8;

[0091] (5) Threshold setting: Select the threshold 0 in the host computer to distinguish between healthy and defective ( Figure 3 ), and the real-time detection false positive rate is 3.2%.

[0092] Example 2: Gearbox wear detection

[0093] (1) Parameter adjustment: Adjust the narrowband range to 3800 - 5200 Hz, and add a margin factor (Ce = 5.8C_e = 5.8Ce = 5.8) to feature extraction;

[0094] (2) Dynamic threshold: When the fusion indicator is lower than the threshold three times in a row, an alarm is triggered, and the missed detection rate is reduced to 2.1%.

[0095] The core innovation points of the acoustic defect detection method based on narrowband screening and feature fusion technology of the present invention:

[0096] 1. Narrowband screening: Locate the defect-sensitive frequency band (such as the center frequency of 4864 Hz and the bandwidth of 512 Hz) through autocorrelation spectral kurtosis analysis to suppress noise interference.

[0097] 2. Multidimensional feature fusion: Extract 26-dimensional features in the time domain and frequency domain, and reduce the dimension to 1-dimensional fusion index through PCA, reducing the computational complexity while retaining key information.

[0098] 3. Dynamic threshold configuration: Users can customize the threshold according to the actual data distribution to adapt to different device models and working conditions.

[0099] 4. Technical effects: The detection speed is increased by 30%, suitable for real-time monitoring scenarios; the accuracy rate remains above 90% when the sample size is insufficient (<50 groups); it has strong anti-interference ability, and the false alarm rate is lower than 5% under the environmental noise of 85 dB.

Claims

1. An acoustic defect detection method based on a feature fusion technique with narrowband screening, characterized in that, It includes the following steps: S1. Signal acquisition and tagging: Collect the sound signals during equipment operation through a sound sensor and tag them as healthy samples or defective samples; S2. Narrowband screening: Conduct autocorrelation spectral kurtosis analysis on the signals to determine the center frequency and bandwidth of the optimal resonance band, and extract the target frequency band signals through a band-pass filter; S3. Multi-dimensional feature extraction: Extract 26-dimensional features of the signals from the time domain and frequency domain respectively, including time-domain peak-to-peak value, kurtosis factor, frequency-domain centroid frequency, and mean square frequency; S4. Feature fusion: Reduce the dimension of the multi-dimensional features through principal component analysis and select the first principal component as the fusion index; S5. Defect detection: Configure a dynamic threshold according to the distribution of the fusion indexes of healthy samples and defective samples, and compare the signal fusion index with the threshold in real time to determine the equipment status.

2. The acoustic defect detection method based on the feature fusion technology of narrowband screening according to claim 1, wherein In the step S2, the autocorrelation spectral kurtosis analysis is specifically as follows: Decompose the signal based on the maximum redundant discrete wavelet packet transform, calculate the unbiased autocorrelation kurtosis value of the squared envelope of each layer of decomposed signals, and select the frequency band corresponding to the maximum kurtosis value as the optimal resonance band.

3. The acoustic defect detection method based on the feature fusion technology of narrowband screening according to claim 2, wherein The autocorrelation spectral kurtosis analysis includes the following steps: Decompose the signal through the maximum redundant discrete wavelet packet transform and calculate the unbiased autocorrelation function of the squared envelope; where τ = q / f s is the delay factor, f s is the sampling frequency, X is the squared envelope of the decomposed signal, M is the total length of the signal, q = 0, 1,..., M - 1; Calculate the spectral kurtosis value of the signals within the frequency band; Select the frequency band corresponding to the maximum spectral kurtosis as the optimal resonance band.

4. The acoustic defect detection method based on the feature fusion technology of narrowband screening according to claim 1, wherein The step S2 further includes: performing band-pass filtering using a Butterworth filter, and its parameters include: passband attenuation αp, stopband attenuation αs, upper passband cut-off frequency ωp, lower stopband cut-off frequency ωs, passband cut-off frequency f P = 4000 Hz, stopband cut-off frequency f s = 5500 Hz.

5. The acoustic defect detection method based on narrowband screening feature fusion technology according to claim 1, characterized in that: In the step S3, the time-domain features include 13 indicators, and the frequency-domain features include 13 indicators, specifically including peak-to-peak value, pulse factor, waveform factor, frequency-domain centroid frequency, and frequency standard deviation.

6. The acoustic defect detection method based on the feature fusion technology of narrowband screening according to claim 1, wherein In the step S4, the principal component analysis includes: Standardize the feature matrix, calculate the covariance matrix, and select the first principal component with a cumulative contribution rate greater than 85% as the fusion index.

7. The acoustic defect detection method based on the feature fusion technology of narrowband screening according to claim 1, characterized in that In the step S5, the dynamic threshold configuration includes: Draw a line graph of the fusion indexes of healthy samples and defective samples, and the user manually selects the threshold according to the line spacing.

8. The acoustic defect detection method based on the feature fusion technology of narrowband screening according to claim 1, wherein In the step S5, the defect determination condition is: When the fusion index is greater than the threshold, it is determined to be healthy, otherwise it is determined to be defective.

9. The acoustic defect detection method based on the feature fusion technology of narrowband screening according to claim 1, characterized in that The installation position of the sound sensor is 10 cm away from the mechanical equipment. The trigger signal controls the start and end of the acquisition, and 65,536 data points are collected each time.

10. The acoustic defect detection method based on the feature fusion technology of narrowband screening according to any one of claims 1-9, characterized in that, The method is realized through the cooperation of the upper computer and the lower computer: The lower computer is responsible for signal acquisition, feature calculation, and threshold comparison, and the upper computer is responsible for threshold configuration, data storage, and visualization.

Images