Rolling bearing fault soundprint recognition method based on power regularized cepstral coefficients

Abstract

The present invention discloses a rolling bearing fault soundprint recognition method based on power regularized cepstral coefficients, which relates to the field of measurement technology and solves the technical problem of low bearing fault recognition accuracy under noise interference. The key point of the technical solution is to use improved power regularized cepstral coefficient features to recognize rolling bearing fault sound signals. Compared with other commonly used features, the calculation amount is only slightly increased, but it has a higher recognition accuracy; at the same time, the improved measure can still maintain a high recognition accuracy in a noisy environment, and has the advantage of strong robustness. In addition, the present application will not affect the normal operation of the rolling bearing, and the sound signal sensor can be installed without stopping the machine or modifying the original system, and on this basis, the rolling bearing fault recognition is realized, which is more efficient.

Description

Technical Field

[0001] The present application relates to the field of measurement technology, and in particular to a rolling bearing fault soundprint recognition method based on power regularized cepstral coefficients. Background Art

[0002] Due to long-term operation under heavy load and high speed environment, rolling bearings have become one of the most prone to failure in rotating mechanical equipment. Therefore, it is of great significance to diagnose rolling bearing faults. At present, the vibration analysis method commonly used in bearing fault diagnosis has the disadvantages of inconvenient sensor installation and inability to accurately identify faults in the early stage of faults. Compared with vibration signals, sound signals have the advantages of non-stop installation, non-contact measurement, convenient signal acquisition, and mature processing methods. In recent years, with the continuous development of voiceprint recognition technology, this technology in the field of speech has also been widely used in the field of fault diagnosis. There have been certain research results on bearing fault diagnosis by analyzing the sound signals of bearings, but most of the research is carried out in the laboratory without considering the noise interference problem in actual engineering. During the actual operation of the bearing, there will be noise interference from other components or the outside world, which poses a severe challenge to the realization of rolling bearing fault diagnosis under this situation. Summary of the invention

[0003] The present application provides a rolling bearing fault soundprint recognition method based on power regularized cepstral coefficients, the technical purpose of which is to improve the accuracy of bearing sound signal recognition under noise interference, thereby improving the accuracy of bearing fault recognition.

[0004] The above technical objectives of this application are achieved through the following technical solutions:

[0005] A rolling bearing fault soundprint recognition method based on power regularized cepstral coefficients, comprising:

[0006] The sound signals of the rolling bearing in a fault state and a normal state are collected, each sound signal is assigned a corresponding label, and the sound signals with labels are obtained, and the sound signals with labels are divided into a training set and a test set;

[0007] Preprocessing the training set, and then extracting features from the preprocessed training set to obtain improved power regularized cepstral coefficient features;

[0008] Construct a voiceprint recognition model, input the improved power regularized cepstral coefficient feature into the voiceprint recognition model to update the parameters, and stop updating when the objective function of the voiceprint recognition model meets the iteration termination condition to obtain the voiceprint model;

[0009] The test set is preprocessed and its improved power regularized cepstral coefficient features are extracted and then input into the voiceprint model. The labels obtained by the voiceprint model are compared with the real labels to obtain the recognition accuracy.

[0010] When the recognition accuracy reaches a preset threshold, the rolling bearing fault voiceprint is recognized through the voiceprint model, otherwise the voiceprint model is continuously trained until its recognition accuracy reaches a preset threshold.

[0011] The beneficial effects of the present application are as follows: the present application uses the improved power regularization cepstral coefficient feature to identify the fault sound signal of the rolling bearing, which has only a small increase in the amount of calculation compared to other commonly used features, but has a higher recognition accuracy; at the same time, the improvement measure can still maintain a high recognition accuracy in a noisy environment, and has the advantage of strong robustness. Moreover, the present application will not affect the normal operation of the rolling bearing, and the sound signal sensor can be installed without stopping the machine or modifying the original system, and on this basis, the fault identification of the rolling bearing can be realized, which is more efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A flowchart of the method described in this application;

[0013] Figure 2 It is a schematic diagram of frame division in sound preprocessing;

[0014] Figure 3 This is a flow chart for extracting the improved power regularized cepstral coefficient features described in this application;

[0015] Figure 4 Schematic diagram of the components of the Gammatone filter. DETAILED DESCRIPTION

[0016] The technical solution of the present application will be described in detail below with reference to the accompanying drawings.

[0017] like Figure 1 As shown, the rolling bearing fault soundprint recognition method based on power regularized cepstral coefficients described in the present application is characterized by comprising:

[0018] S1: Sound signals of rolling bearings in faulty and normal states are collected, each sound signal is assigned a corresponding label, and labeled sound signals are obtained. The labeled sound signals are divided into a training set and a test set.

[0019] Generally, the sound signals of rolling bearings under three fault states and one normal state are collected. The fault states include inner ring fault, outer ring fault and inner and outer ring fault of the rolling bearing. The sound samples are 2s in length, and different types of bearing sound data are assigned different labels as a domain. For example, the label of the sound signal under the inner ring fault state is inner ring fault.

[0020] S2: Preprocess the training set, and then extract features from the preprocessed training set to obtain improved power regularized cepstral coefficient features.

[0021] Specifically, preprocessing includes pre-emphasis, framing and windowing. Pre-emphasis is because the general sound signal will have a phenomenon that the higher the frequency, the more serious the attenuation, so it is necessary to use a high-pass filter to compensate for the high-frequency component, which is expressed as:

[0022]

[0023] Wherein, α represents the pre-emphasis coefficient; S(n) represents the original signal; Indicates the sound signal after pre-emphasis.

[0024] The sound signal is an unsteady signal, but the characteristic changes are small in a short time range (10-20ms), so it can be regarded as a locally stable random signal. Dividing the sound signal into many segments according to time is called framing. The two most important parameters for framing are frame length and frame interval. The length of a frame is called the frame length, which is 25ms in this application; the interval between two adjacent frames is called the frame interval, which is 40% of the frame length in this application. A set of samples with a sampling frequency of 44.1kHz and a duration of 2s can obtain 197 frames after preprocessing. The framing diagram is shown in the figure. Figure 2 shown.

[0025] There is a high degree of discontinuity at both ends of the framed signal. In order to avoid the Gibbs effect of local peaks in the spectrum when directly performing Fourier transform, it is necessary to use windowing processing on the framed signal, assigning a larger weight to the central signal and a weight close to 0 to the window edge signal. This application uses a Hamming window. For a discrete signal x[n] of length N, the weight w[n] corresponding to the window function can be expressed as:

[0026]

[0027] The sampling rate of sound signals is high, and even if it is only a few seconds, there are tens of thousands of points. In order to facilitate research, it is necessary to compress the complex data into a vector of fixed dimension as a feature. In view of the differences between the characteristics of rolling bearing sound signals and human voice signals, the Gammatone filter component and the medium-duration power calculation module are improved on the basis of the original PNCC (power regularized cepstral coefficient feature), and the first-order difference and second-order difference are added to the calculation results to enhance dynamic continuity as an improved PNCC feature.

[0028] Improve the PNCC extraction process such as Figure 3 As shown in the figure, they are: FFT transformation, Gamma filtering, medium time processing, time-frequency normalization, average power normalization, power function transformation, DCT discrete cosine transform, first-order difference and second-order difference. The specific calculation process includes:

[0029] (1) Perform fast Fourier transform (FFT) on the preprocessed sound signal to obtain the spectrum, which is expressed as:

[0030]

[0031] (2) The frequency of the bearing sound signal is also significantly distributed in the mid-high frequency segment compared to the frequency of the human voice signal. Therefore, compared with the normalized filter component during the original PNCC extraction, using the original channel component of the Gammatone filter can increase the weight of the mid-high frequency component and is more suitable for the processing of the bearing sound signal. The original channel component of the Gammatone filter is as follows: Figure 4 As shown. Through a set of original component Gammatone filters, the short-time spectrum power P[m,l] is obtained, where m and l represent the number of frames and channels respectively, which is specifically expressed as:

[0032]

[0033] Where k represents the index of N, m represents the number of frames, and l represents the number of channels; H l (k) represents a Gammatone filter.

[0034] (3) Medium-time processing: When the original PNCC calculates the medium-time power, the average power of the m frames before and after a frame is used as the short-time power of the current frame. The sound signal of mechanical equipment is more stable than the human voice signal. By increasing the number of frames when calculating the medium-time power, that is, increasing the m value, the sound signal characteristics of the bearing are adapted. For example, when m is 3, M=2m+1=7, then the average power of the 7 frames before and after the current frame is increased as the short-time power of the current frame. Then, after asymmetric noise suppression, time masking, and weight smoothing, the result is obtained.

[0035] (4) Time-frequency normalization yields T[m,l], expressed as:

[0036]

[0037] (5) The average power is normalized to obtain U[m,l], which is expressed as:

[0038]

[0039] (6) The power function transformation simulates the nonlinear characteristics of human ear's frequency perception and obtains V[m,l], which is expressed as:

[0040] V[m,l]=U[m,l] 1 / 15 ; (7)

[0041] (7) Finally, after discrete cosine transform (DCT), the first 13 parameters are taken to obtain W[m,l], and the first-order difference and second-order difference are performed on it and then concatenated with the original value to obtain the final 39-dimensional improved PNCC feature. The first-order difference method is shown in formula (8), and the second-order difference is performed on the basis of the first-order difference.

[0042]

[0043] S3: Construct a voiceprint recognition model, input the improved power regularized cepstral coefficient features into the voiceprint recognition model for parameter update, and stop updating when the objective function of the voiceprint recognition model meets the iteration termination condition to obtain the voiceprint model.

[0044] Specifically, the classifier of the above voiceprint recognition model is the Support Vector Machine (SVM) algorithm. The idea of ​​SVM is to establish a partitioning hyperplane in the sample space to separate samples of different categories so as to maximize the sum of the distances from heterogeneous support vectors to the hyperplane. The kernel function of this application uses rbf, the penalty coefficient C is 1, and the kernel function coefficient gamma is 0.00039.

[0045] S4: Preprocess the test set and extract its improved power regularized cepstral coefficient features, then input them into the voiceprint model. Compare the labels obtained by the voiceprint model with the real labels to obtain the recognition accuracy.

[0046] S5: When the recognition accuracy reaches a preset threshold, the rolling bearing fault voiceprint is recognized through the voiceprint model, otherwise the voiceprint model is continuously trained until its recognition accuracy reaches a preset threshold.

[0047] As a specific implementation, this application will compare and analyze the improved PNCC features with several other different features. The comparative experiment selects three features that are widely used in the field of sound signals: Mel-Frequency Cepstral Coefficients (MFCC), Gammatone Cepstral Coefficients (GTCC), and original power regularized cepstral coefficients PNCC. Among them, the three feature preprocessings all use a 25ms frame length and a 10ms frame interval; a group of 2s sound samples can be divided into 197 frames, each frame obtains a total of 13-dimensional parameters, and the cascade of each frame feature obtains a total of 2561-dimensional parameters.

[0048] The recognition accuracy results of the four different features on the pure sound signal test set are shown in Table 1. It can be seen that when there is no noise interference, the recognition accuracy of the four features can reach 100% or close to 100%, and the fault type can be well identified. The improved PNCC of this application has a certain improvement over the original PNCC.

[0049] Table 1. Recognition accuracy of pure sound signals

[0050] feature Recognition accuracy / % MFCC 100.00 GTCC 99.00 PNCC 98.00 Improved PNCC 100.00

[0051] In order to verify the recognition accuracy of this application under noise interference, noise is superimposed on the pure sound signal to simulate the real environment. Table 2 shows the average recognition rate of four different features facing five types of noise under four different signal-to-noise ratios. As can be seen from Table 2, the improved PNCC feature has a similar recognition rate when the signal-to-noise ratio is 20dB, which is equal to or close to 100% compared with other features; when the signal-to-noise ratio is below 20dB, the improved PNCC recognition accuracy is higher and has stronger robustness. The average value of the improved PNCC at each signal-to-noise ratio is 13.35% higher than that of the original PNCC.

[0052] In order to explore the effect of each improved method in the proposed improved PNCC feature, Table 3 shows the average recognition rate of five types of noise under different signal-to-noise ratios when only retaining the original components of the original Gammaone filter, only adjusting the total number of frames M for power calculation, and only the first-order and second-order differences. It can be seen from Table 3 that compared with the original PNCC, retaining the original components of the filter can improve the recognition rate at each signal-to-noise ratio; the number of frames M for power calculation during adjustment has a certain improvement in the recognition rate under low signal-to-noise ratio, and the average recognition rate is the highest when M is 7; the first-order and second-order differences can improve the recognition rate under high signal-to-noise ratio. Compared with using the three improved methods alone, the recognition rate at each signal-to-noise ratio is significantly improved when the three improved methods are used simultaneously.

[0053] Table 2 Average recognition accuracy of different features under different signal-to-noise ratios

[0054]

[0055] Table 3 Average recognition accuracy of different improved methods under different signal-to-noise ratios

[0056]

[0057] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art to which this application belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0058] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A rolling bearing fault soundprint recognition method based on power regularized cepstral coefficients, characterized in that: include: The sound signals of the rolling bearing in a fault state and a normal state are collected, each sound signal is assigned a corresponding label, and the sound signals with labels are obtained, and the sound signals with labels are divided into a training set and a test set; Preprocessing the training set, and then extracting features from the preprocessed training set to obtain improved power regularized cepstral coefficient features; Construct a voiceprint recognition model, input the improved power regularized cepstral coefficient feature into the voiceprint recognition model to update the parameters, and stop updating when the objective function of the voiceprint recognition model meets the iteration termination condition to obtain the voiceprint model; The test set is preprocessed and its improved power regularized cepstral coefficient features are extracted and then input into the voiceprint model. The labels obtained by the voiceprint model are compared with the real labels to obtain the recognition accuracy. When the recognition accuracy reaches the preset threshold, the voiceprint model is used to recognize the rolling bearing fault voiceprint, otherwise the voiceprint model is continuously trained until its recognition accuracy reaches the preset threshold; The extraction of the improved power regularized cepstral coefficient feature includes: subjecting the pre-processed sound signal to FFT transformation, Gamma filtering, medium time processing, time-frequency normalization, average power normalization, power function transformation, DCT, first-order difference and second-order difference in sequence, and finally obtaining the improved power regularized cepstral coefficient feature; The preprocessing includes pre-emphasis, framing and windowing; The pre-emphasis compensates for the high frequency component through a high-pass filter, which is expressed as: Wherein, α represents the pre-emphasis coefficient; S(n) represents the original signal; Indicates the sound signal after pre-emphasis; Framing is performed by frame length and frame interval; Windowing is implemented through a Hamming window, including: For a discrete signal x[n] of length N, the weight w[n] corresponding to the window function is expressed as: The Gamma filtering is implemented by a Gammatone filter. The weights of the mid- and high-frequency components are increased through the original component channel of the Gammatone filter to obtain the short-time spectrum power P[m,l], then P[m,l] is expressed as: Where k represents the index of N, m represents the number of frames, and l represents the number of channels; H l (k) represents a Gammatone filter; The medium time processing includes: increasing the number of frames in the medium time power calculation, and then performing asymmetric noise suppression, time masking and weight smoothing to obtain Among them, increasing the number of frames when calculating the mid-time power includes: when calculating the power of the current frame, taking the power average of the previous and next M frames as the power of the current frame, then the number of previous frames and next frames of the current frame are both (M-1) / 2.

2. The method according to claim 1, characterized in that The fault states include inner ring fault, outer ring fault and inner and outer ring fault of the rolling bearing; the sound signal is a sound sample with a time length of 2s.

3. The method according to claim 1, characterized in that The classifier of the voiceprint recognition model is a support vector machine.

Images