A rolling bearing voiceprint anomaly detection method based on MDE-SVDD

Abstract

The present invention discloses a method for abnormal detection of rolling bearing acoustic fingerprints based on MDE-SVDD, which relates to the technical field of abnormal detection of rolling bearings and solves the technical problem of low efficiency of abnormal detection of rolling bearings under strong noise. The key points of its technical solution are to collect the sound signals of normal rolling bearings in operation, extract MFCC features and send them to the support vector data description based on Mahalanobis distance weighting (MDE-SVDD) for abnormal detection. Through the method of Mahalanobis distance weighting, the problems that the operation data of rolling bearings are vulnerable to noise interference and SVDD does not consider the data density distribution are solved. This method can effectively extract fault features from vibration signals with different intensities of noise, has good diagnostic performance, higher accuracy, fewer training parameters, fast convergence speed and good robustness.

Description

Technical Field

[0001] This application relates to the technical field of rolling bearing anomaly detection, and particularly to a rolling bearing voiceprint anomaly detection method based on MDE-SVDD. Background Art

[0002] With the continuous development of science and technology, mechanical equipment has become a key component in many production processes, and rotating mechanical equipment is very important for the development of the manufacturing industry. Rolling bearings are key components in large rotating machinery. During the operation of the equipment, rolling bearings bear huge loads and are prone to failures. Data shows that more than 40% of the failures in rotating machinery are bearing failures. Once a failure occurs, it will cause the entire system to shut down, having a huge impact on the safe operation of the production process. Therefore, when diagnosing faults in rolling bearings, if anomalies can be detected in advance and adjusted, it is of great significance for preventing accidents. Sound signals have advantages such as non-stop installation, non-contact measurement, convenient signal acquisition, and mature processing methods compared to vibration signals. In recent years, with the continuous development of voiceprint recognition technology, this technology in the field of speech has also been widely applied in the field of fault diagnosis, but there is still a large gap in its application in the field of anomaly detection.

[0003] Rolling bearing anomaly detection methods have already had a certain development, but currently still face many challenges. For example: traditional features need to be manually extracted, and different features need to be extracted for faults under different working conditions, which will require more expert experience. Traditional deep learning networks can only utilize single features and cannot extract deeper features of vibration signals, resulting in low diagnostic efficiency. In addition, under strong noise conditions, fault features are submerged by noise, making it difficult to extract effective feature information and unable to efficiently complete rolling bearing anomaly detection. Summary of the Invention

[0004] This application provides a rolling bearing voiceprint anomaly detection method based on MDE-SVDD, and its technical objective is to achieve rolling bearing anomaly detection by only using normal data to train the model under strong noise.

[0005] The above technical objective of this application is achieved through the following technical solutions:

[0006] A rolling bearing voiceprint anomaly detection method based on MDE-SVDD includes:

[0007] S1: Collect the sound signals of the rolling bearing under different operating states, obtain the original sound data of the rolling bearing, and add strong noise signals to the original sound data to obtain sound signal data with different signal-to-noise ratios;

[0008] S2: Segment the sound signal data samples. Divide the normal samples into a training set and a test set at a ratio of 3:1, and put all the abnormal state samples into the test set;

[0009] S3: Extract the Mel-frequency cepstral coefficients (MFCCs) of the training set and the test set respectively, and calculate the Mahalanobis distance weighting coefficients for the training set;

[0010] S4: Construct a speaker recognition model based on the SVDD model, and introduce the Mahalanobis distance weighting coefficients into the speaker recognition model to obtain the MDE-SVDD model. Use the MFCCs extracted from the training set as features and input them into the MDE-SVDD model to train and update the parameters of the MDE-SVDD model until the parameters of the MDE-SVDD model converge to end the training, and obtain the trained MDE-SVDD model;

[0011] S5: Input the MFCCs extracted from the test set into the trained MDE-SVDD model for anomaly detection to obtain the detection results.

[0012] The beneficial effects of this application are as follows:

[0013] This application designs an MDE-SVDD model suitable for anomaly detection of rolling bearings under strong noise, realizes automatic extraction of fault features under strong noise, can be better applied to anomaly detection of rolling bearings under strong noise, has higher accuracy, fewer training parameters, strong anti-noise performance, and good robustness.

[0014] This application solves the problems that the operation data of rolling bearings are vulnerable to noise interference and the SVDD does not consider the data density distribution through the method of Mahalanobis distance weighting. This method can effectively extract fault features from vibration signals with different intensities of noise and has good diagnostic performance. Brief Description of the Drawings

[0015] Figure 1 is a flow chart of the method described in this application;

[0016] Figure 2 is a schematic structural diagram of the test bench in the embodiment of this application;

[0017] Figure 3 is the time-domain waveform diagram and frequency-domain diagram of the original sound data in the embodiment of this application;

[0018] Figure 4 is a flow chart for extracting the Mel-frequency cepstral coefficient features in the embodiment of this application;

[0019] Figure 5 is a schematic diagram of the training results of the MDE-SVDD model in the embodiment of this application. Detailed Embodiments

[0020] The technical solution of the present application will be described in detail below with reference to the accompanying drawings.

[0021] As Figure 1 shown, the method for abnormal rolling bearing acoustic fingerprint detection based on MDE-SVDD of the present application includes:

[0022] S1: Collect the sound signals of the rolling bearing under different operating states, obtain the original sound data of the rolling bearing, and add strong noise signals to the original sound data to obtain the sound signal data at different signal-to-noise ratios.

[0023] In the embodiment of the present application, the sound signal is collected through a test bench, as shown in Figure 2 . This test bench has the advantages of simple equipment, flexible installation, easy signal acquisition, and non-contact measurement. It will not affect the normal operation of the rolling bearing, and can install the sound signal sensor without shutting down the machine or modifying the original system, and can be better applied to complex industrial scenarios.

[0024] Collect the sound signals of the rolling bearing under five operating states, namely normal operation, outer ring crack, inner ring crack, inner and outer ring cracks, and rolling element failure, through the test bench. In this embodiment, the rolling bearing model is NSK-uc210.

[0025] The drive motor of the transmission chain system runs at a speed of 900 r / min, the sampling frequency is designed to be 22,050 Hz, and the sound signal of each type of bearing operation is collected in groups of 5120 points at 25 ms. The data of the normally operating bearing is continuously collected for 350 s, and the data of each type of abnormal sample is continuously collected for 200 s. Figure 3 It is the time-frequency diagram of the collected original sound data.

[0026] Due to the complex industrial site environment, there is usually noise interference when collecting sound signals. To further verify the model performance, Gaussian white noise is added to the collected sound signals to construct the noisy sound signal data with signal-to-noise ratios of 10, 0, and -5 dB respectively.

[0027] S2: Perform sample segmentation on the sound signal data, divide the normal samples into a training set and a test set according to 3:1, and put all the abnormal state samples into the test set.

[0028] S3: Extract the Mel-frequency cepstral coefficients of the training set and the test set respectively, and calculate the Mahalanobis distance weighting coefficients of the training set.

[0029] The extraction process of the Mel-frequency cepstral coefficients MFCC is as shown in Figure 4As shown, they are pre-emphasis, frame windowing, fast Fourier transform, and logarithmic operation and discrete cosine transform after passing through the Mel filter bank in sequence. In this embodiment, processing is performed with a frame length of 25 ms (5120) and a frame shift of 2048. For normal samples, a total of 3769 groups of 13-dimensional Mel cepstral coefficients are obtained, and for each of the 4 types of abnormal samples, 2154 groups of samples are obtained.

[0030] Specifically, the extraction of Mel cepstral coefficients includes:

[0031] S311: Perform preprocessing of pre-emphasis and frame windowing on the sound signal data, expressed as:

[0032]

[0033] S(n) = s(n)·h(n);

[0034]

[0035] Among them, α represents the pre-emphasis coefficient; s(n) represents the sound signal before preprocessing; represents the sound signal after preprocessing;

[0036] S312: Perform a fast Fourier transform on the preprocessed signal to obtain the spectrum X(k), expressed as:

[0037]

[0038] S313: Square X(k) to obtain the energy spectrum, and then filter it using M MEL band-pass filter banks; among them, the transfer function of the m-th filter is expressed as:

[0039]

[0040] Among them, f(m) represents the center frequency of the triangular filter;

[0041] S314: Calculate the logarithmic energy of the MEL band-pass filter bank, then the logarithmic energy of the m-th filter is expressed as:

[0042]

[0043] S315: Perform a discrete cosine transform on the logarithmic energy to obtain the Mel cepstral coefficients, expressed as:

[0044]

[0045] Among them, M is the number of filters, that is, the dimension of the Mel cepstral coefficient feature.

[0046] S4: Construct a voiceprint recognition model based on the SVDD model, introduce the Mahalanobis distance weighting coefficient into the voiceprint recognition model to obtain the MDE-SVDD model, use the Mel cepstral coefficients extracted from the training set as features and input them into the MDE-SVDD model, train the MDE-SVDD model and update its parameters until the parameters of the MDE-SVDD model converge to end the training, and obtain the trained MDE-SVDD model.

[0047] Specifically, the classifier of the MDE-SVDD model is the Mahalanobis distance weighted support vector data description (MDE-SVDD) algorithm. The idea of SVDD is to establish a dividing hypersphere in the sample space and minimize the radius of the hypersphere on the premise of enclosing normal samples. MDE-SVDD uses the Mahalanobis distance weighting coefficient on the basis of SVDD to assign different weights to samples according to the degree of outliers. In this application, the kernel function is rbf, the penalty coefficient C is taken as 0.5, and the kernel function coefficient gamma is taken as 0.01. The training results of the MDE-SVDD model are as Figure 5 shown.

[0048] Specifically, the SVDD model includes:

[0049] (1) Construct the objective function of the SVDD model, expressed as:

[0050]

[0051] where R represents the radius of the hypersphere; C represents the penalty factor; x (i) represents the sample point; ξ represents the slack variable;

[0052] (2) Use the Lagrange multiplier method to obtain the Lagrangian function, make the partial derivatives of the Lagrangian function with respect to R, a, and ξ i all equal to 0, substitute the results back into the Lagrangian function, and transform the solution of the above objective function into a dual problem, expressed as:

[0053]

[0054] where α i , γ i ≥0, both represent Lagrange multipliers;

[0055] (3) Introduce the kernel function K(x (i) , x (j) ) to replace the in the above formula, deform the dual problem, and obtain the solution of the objective function as:

[0056]

[0057] (4) Update the hypersphere center vector, the hypersphere radius R, and the Lagrange multipliers through the sequential minimal optimization method to minimize the objective function; when the degree of the maximum violation of the KKT condition reaches the convergence accuracy, stop the iteration.

[0058] The Mahalanobis distance weight coefficient is expressed as:

[0059]

[0060]

[0061] where D M represents the Mahalanobis distance, μ represents the mean, and ∑ -1 represents the inverse matrix of the covariance matrix; D M avr represents the average of the Mahalanobis distances from all samples to the mean; D M max and D M min respectively represent the maximum and minimum values in D M (x i ); ε represents the correction coefficient. In this embodiment, ε = 0.01.

[0062] Introduce the Mahalanobis distance weighting coefficient into the SVDD model to obtain the MDE - SVDD model. Then, the objective function of the MDE - SVDD model is expressed as:

[0063]

[0064] Correspondingly, the solution corresponding to the objective function of the MDE - SVDD model is expressed as:

[0065]

[0066] S5: Input the Mel - frequency cepstral coefficients extracted from the test set into the trained MDE - SVDD model for anomaly detection to obtain the detection result.

[0067] In order to verify the performance of the proposed MDE - SVDD model in this application, verify its recognition accuracy under different noises in 5 operating states including inner - ring crack, outer - ring crack, inner - and - outer - ring crack, rolling - element fault, and normal state.

[0068] Table 1 Detection results of each abnormal state under different signal - to - noise ratios

[0069]

[0070] The abnormal detection results for various operating states under different noise levels are shown in Table 1. The recognition rates of the model for normal / abnormal test samples are represented by accuracy / true negative rate respectively. In the case of no noise / slight noise (10 db), the MDE-SVDD model achieved a recognition rate close to 100%. Even in the strong noise background (-5 db), the detection accuracy was still close to 90%. For different types of abnormal samples, the MDE-SVDD model showed similar detection effects and there was no preference for a certain specific type of abnormality.

[0071] To verify the improvement effect of the MDE-SVDD model proposed in this application compared with the ordinary SVDD model, a comparative study on the abnormal detection performance of the ordinary SVDD model and AE (autoencoder) was carried out. The data of 4 abnormal states and normal state were mixed into the test set, and the average accuracy (Accuracy) of the three classification algorithms was compared. The parameters of the ordinary SVDD model were the same as those of the MDE-SVDD model; the number of input layer units of the autoencoder was 5120, and the number of hidden layer units was 25. The discrimination was made by taking a fixed threshold for the correlation coefficient between the reconstructed signal and the original signal. The results of the three models are shown in Table 2.

[0072] Table 2 Abnormal detection results of each model under different signal-to-noise ratios

[0073]

[0074] Using the MDE-SVDD model, the abnormal detection accuracy is higher than 90% under any noise condition. As the noise intensity increases, the gap in the abnormal detection accuracy between the MDE-SVDD model and the ordinary SVDD model is increasing. In the strong noise (-5 db) background, the accuracy of the MDE-SVDD model is relatively improved by 9.17% compared with the ordinary SVDD model, which indicates that the MDE-SVDD model has better anti-noise performance.

[0075] Those skilled in the art of this technology can understand that unless otherwise defined, all terms (including technical terms and scientific terms) used here have the same meaning as the general understanding of those of ordinary skill in the art to which this application belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted with an idealized or overly formal meaning unless defined as here.

[0076] Taking the above ideal embodiments of the present invention as an inspiration, through the above description, relevant staff can completely make various changes and modifications without departing from the technical idea of this invention. The technical scope of this invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A method for detecting abnormal sound patterns of rolling bearings based on MDE-SVDD, characterized in that, Including: S1: Collect the sound signals of the rolling bearing under different operating states, obtain the original sound data of the rolling bearing, and add strong noise signals to the original sound data to obtain sound signal data with different signal-to-noise ratios; S2: Perform sample segmentation on the sound signal data, divide the normal samples into a training set and a test set at a ratio of 3:1, and put all the abnormal state samples into the test set; S3: Extract the Mel-frequency cepstral coefficients of the training set and the test set respectively, and calculate the Mahalanobis distance weighting coefficients of the training set; S4: Construct a voiceprint recognition model based on the SVDD model, and introduce the Mahalanobis distance weighting coefficients into the voiceprint recognition model to obtain the MDE-SVDD model. Use the Mel-frequency cepstral coefficients extracted from the training set as features and input them into the MDE-SVDD model to train and update the parameters of the MDE-SVDD model until the parameters of the MDE-SVDD model converge and end the training to obtain the trained MDE-SVDD model; S5: Input the Mel-frequency cepstral coefficients extracted from the test set into the trained MDE-SVDD model for anomaly detection to obtain the detection results.

2. The method according to claim 1, characterized in that, In step S3, the extraction of the Mel-frequency cepstral coefficients includes: S311: Perform pre-emphasis and frame windowing preprocessing on the sound signal data, expressed as: S(n) = s(n)·h(n); Among them, α represents the pre-emphasis coefficient; s(n) represents the sound signal before preprocessing; represents the sound signal after preprocessing; S312: Perform a fast Fourier transform on the preprocessed signal to obtain the spectrum X(k), which is expressed as: S313: Square X(k) to obtain the energy spectrum, and then filter it using M MEL band-pass filter banks; where the transfer function of the m-th filter is expressed as: Among them, f(m) represents the center frequency of the triangular filter; S314: Calculate the logarithmic energy of the MEL band-pass filter bank, and the logarithmic energy of the m-th filter is expressed as: S315: Perform discrete cosine transform on the logarithmic energy to obtain the Mel-frequency cepstral coefficients, expressed as: where M is the number of filters, that is, the dimension of the Mel-frequency cepstral coefficient features.

3. The method according to claim 1, wherein The SVDD model includes: (1) Construct the objective function of the SVDD model, expressed as: where, R represents the radius of the hypersphere; C represents the penalty factor; x (i) represents the sample point; ξ represents the slack variable; (2) The Lagrangian function is obtained using the Lagrange multiplier method. Let the partial derivatives of the Lagrangian function with respect to R, a, and ξ i all be 0, and substitute the results back into the Lagrangian function to transform the solution of the above objective function into a dual problem, which is expressed as: where α i , γ i ≥ 0, both represent Lagrange multipliers; (3) Introduce the kernel function K(x (i) , x (j) ) to replace x in the above formula (i)T x (j) , and transform the dual problem to obtain the solution of the objective function as follows: (4) Update the hypersphere center vector, hypersphere radius R, and Lagrange multipliers through the sequential minimal optimization method to minimize the objective function; when the degree of satisfying the maximum violation of the KKT condition reaches the convergence accuracy, stop the iteration.

4. The method according to claim 1, wherein In step S3, the Mahalanobis distance weight coefficient is expressed as: Among them, D M represents the Mahalanobis distance, μ represents the mean value, and ∑ -1 represents the inverse matrix of the covariance matrix; D M avr represents the average of the Mahalanobis distances of all samples to the mean value; D M max and D M min respectively represent the maximum value and the minimum value in D M (x i ); ε represents the correction coefficient.

5. The method according to claim 4, wherein In step S4, introducing the Mahalanobis distance weighting coefficients into the SVDD model to obtain the MDE-SVDD model, the objective function of the MDE-SVDD model is expressed as: Correspondingly, the solution corresponding to the objective function of the MDE-SVDD model is expressed as:

6. The method according to claim 4, characterized in that ε = 0.01.

Images