Gearbox sound signal fault identification method, device, equipment and medium

Abstract

The invention relates to a gearbox sound signal fault identification method, device and equipment and a medium, and the method comprises the steps: carrying out the preprocessing of a to-be-identified gearbox sound signal, and obtaining the preprocessed to-be-identified gearbox sound signal; extracting a Mel-frequency cepstral coefficient of the preprocessed to-be-identified gearbox sound signal, and obtaining an MFCC static feature and an MFCC dynamic feature of the gearbox sound signal according to the Mel-frequency cepstral coefficient; a sparse self-encoding SAE network is adopted to extract self-learning features of the preprocessed gearbox sound signals; and inputting the MFCC static features, the MFCC dynamic features and the self-learning features into a gearbox fault recognition model to complete fault recognition of the to-be-recognized gearbox sound signals. According to the method, the characteristics capable of representing typical faults of the gearbox are effectively extracted, and the accuracy and robustness of the gearbox fault diagnosis method are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of rotating machinery fault diagnosis, and in particular to a gearbox acoustic signal fault identification method, device, equipment and medium. Background Art

[0002] Gearboxes are important rotating mechanical structures and are widely used in various mechanical fields such as transmission systems and generator sets. Due to long-term extreme working conditions such as high speed and heavy load, its key parts will inevitably produce various faults, causing unnecessary downtime and even casualties. As the industrialization of gearboxes becomes higher and higher, the working environment is also relatively harsh, so the use of non-contact acoustic sensors to collect gearbox acoustic signals for feature extraction and fault identification has certain value, and is also a means to supplement vibration monitoring and diagnostic methods. Effective acoustic diagnostic methods can solve the problem of gear health status diagnosis in noisy environments, and can identify various fault conditions in real time.

[0003] Gearboxes have many parts, the gears mesh frequently, and the signals are coupled with each other, which leads to complex fault signals. It is difficult to extract fault feature information using traditional methods. With the advent of the artificial intelligence era, more effective detection methods have been provided to the field of fault diagnosis. Deep neural networks can make timely predictions of the operating status of gearboxes through continuous self-learning and training, and have higher accuracy than fault diagnosis methods that rely on human experience. Drawing on the research results in speech recognition, the commonly used sound feature extraction method, Mel-frequency cepstral coefficients, can accurately extract feature information from language signals by simulating the auditory system of the human ear. However, there is a big difference between the acoustic signal and the speech signal of the gearbox, and it is difficult to accurately extract effective feature signals directly using the MFCC algorithm.

[0004] Therefore, this paper improves the Mel frequency cepstrum based on the characteristic distribution of gearbox acoustic signals, and proposes a gearbox acoustic signal fault identification method based on improved MFCC and self-learning features. By effectively extracting the improved Mel frequency cepstrum coefficients and self-learning features of the gearbox acoustic signal, the health status of the gearbox can be monitored in time, the accuracy and robustness of the gearbox fault diagnosis method can be increased, and a technical research basis can be provided for condition-based maintenance of the gearbox to avoid safety hazards and economic losses. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a gearbox acoustic signal fault identification method, device, equipment and medium, which can effectively extract features that can characterize typical gearbox faults and increase the accuracy and robustness of the gearbox fault diagnosis method.

[0006] The technical solution adopted by the present invention to solve the technical problem is: to provide a gearbox acoustic signal fault identification method, comprising the following steps:

[0007] Preprocessing the gearbox sound signal to be identified to obtain a preprocessed gearbox sound signal to be identified;

[0008] Extract the Mel frequency cepstrum coefficients of the preprocessed gearbox sound signal to be identified, and obtain the MFCC static features and MFCC dynamic features of the gearbox sound signal according to the Mel frequency cepstrum coefficients; the MFCC static features include the first 2-13 Mel frequency cepstrum coefficients and short-time energy of the gearbox sound signal; the MFCC dynamic features include the first-order difference features of two adjacent Mel frequency cepstrum coefficients and the ZMFCC features; the ZMFCC features are expressed as: ZMFCC=MFCC+MFCC μ +3σ(MFCC), where ZMFCC is the ZMFCC feature, MFCC represents the Mel frequency cepstral coefficient, MFCC μ is the mean value of the Mel frequency cepstral coefficients, σ(MFCC) is the variance of the Mel frequency cepstral coefficients;

[0009] A sparse autoencoder (SAE) network is used to extract the self-learning features of the preprocessed gearbox acoustic signal;

[0010] The MFCC static features, MFCC dynamic features and self-learning features are input into a gearbox fault recognition model to complete fault recognition of the gearbox acoustic signal to be recognized.

[0011] The preprocessing of the gearbox acoustic signal to be identified to obtain the preprocessed gearbox acoustic signal to be identified specifically includes:

[0012] Decentralizing the gearbox acoustic signal to be identified to remove DC component interference;

[0013] The signal-to-noise ratio of the gearbox acoustic signal to be identified after decentralized processing is enhanced based on partial coherence analysis;

[0014] The gearbox acoustic signal to be identified after signal-to-noise ratio enhancement is subjected to PCA whitening processing to remove the strong correlation between signals and obtain the preprocessed gearbox acoustic signal to be identified.

[0015] The extracting of the Mel frequency cepstrum coefficients of the preprocessed gearbox sound signal to be identified, and obtaining the MFCC static features and MFCC dynamic features of the gearbox sound signal according to the Mel frequency cepstrum coefficients, specifically includes:

[0016] Pre-emphasize the pre-processed gearbox acoustic signal;

[0017] The pre-emphasized gearbox acoustic signal is framed and windowed by a sliding window method to obtain a plurality of frames of windowed processed signals;

[0018] Perform fast Fourier transform on the windowed signal of each frame to obtain the spectrum of each frame, and take the modulus square of the spectrum of each frame to obtain the power energy spectrum of the signal of each frame;

[0019] The power energy spectrum of each frame signal is converted by a Mel-scale triangular filter bank to obtain the logarithmic energy spectrum of each frame signal;

[0020] Using discrete cosine transform to remove the order features with low correlation in the logarithmic energy spectrum, and obtaining the first 13 Mel-frequency cepstrum coefficients;

[0021] The first 2-13 Mel-frequency cepstral coefficients and short-time energy are used as MFCC static features of the gearbox sound signal;

[0022] The first-order difference features and ZMFCC features of two adjacent Mel-frequency cepstral coefficients are calculated based on the first 13 Mel-frequency cepstral coefficients obtained, and the first-order difference features and ZMFCC features of two adjacent Mel-frequency cepstral coefficients obtained are used as the MFCC dynamic features of the gearbox acoustic signal.

[0023] The power energy spectrum of each frame signal is converted by a set of Mel-scale triangular filter groups to obtain the logarithmic energy spectrum of each frame signal, specifically including:

[0024] Determine the lowest frequency and the highest frequency of each frame signal, and the number of Mel filters in the Mel-scaled triangular filter bank;

[0025] Convert the lowest and highest Mel frequencies through the conversion relationship between physical frequency and Mel frequency;

[0026] Calculate the distance between the center Mel frequencies of two connected Mel filters;

[0027] Convert each center Mel frequency into a frequency, and calculate the subscript of the midpoint of the fast Fourier transform corresponding to the frequency;

[0028] The logarithmic energy of each filter output is calculated to obtain the logarithmic energy spectrum of each frame signal.

[0029] The sparse autoencoder SAE network is trained in the following way:

[0030] Preprocessing the collected gearbox sound signal;

[0031] The preprocessed gearbox acoustic signal is subjected to fast Fourier transform, and the transformed gearbox acoustic signal is cut into multiple spectrum frame sequences using a sliding window, and the root mean square feature is extracted from each spectrum frame sequence;

[0032] The obtained root mean square features are used to train the sparse autoencoder SAE network, so that the trained sparse autoencoder SAE network can extract self-learning features from the root mean square features.

[0033] The gearbox fault identification model is constructed based on the CNN network and trained in the following way:

[0034] Collecting gearbox acoustic signals under different fault modes, and dividing the obtained gearbox acoustic signals into a training set and a test set;

[0035] Extracting Mel frequency cepstral coefficients of the gearbox sound signals in the training set and the test set, and obtaining MFCC static features and MFCC dynamic features of the gearbox sound signals in the training set and the test set according to the Mel frequency cepstral coefficients;

[0036] Using a sparse autoencoder (SAE) network to extract self-learning features of the gearbox acoustic signals in the training set and the test set;

[0037] The gearbox fault identification model is trained using the MFCC static features, MFCC dynamic features and self-learning features of the gearbox sound signal in the training set. After the training is completed, the gearbox fault identification model is tested using the MFCC static features, MFCC dynamic features and self-learning features of the gearbox sound signal in the test set. If the test requirements are met, the training of the gearbox fault identification model is completed. If the test requirements are not met, the gearbox fault identification model continues to be trained until the test requirements are met.

[0038] The technical solution adopted by the present invention to solve the technical problem is: to provide a gearbox acoustic signal fault identification device, comprising:

[0039] A preprocessing module, used for preprocessing the gearbox sound signal to be identified to obtain the preprocessed gearbox sound signal to be identified;

[0040] The first extraction module is used to extract the Mel frequency cepstrum coefficients of the preprocessed gearbox sound signal to be identified, and obtain the MFCC static features and MFCC dynamic features of the gearbox sound signal according to the Mel frequency cepstrum coefficients; the MFCC static features include the first 2-13 order Mel frequency cepstrum coefficients and short-time energy of the gearbox sound signal; the MFCC dynamic features include the first-order difference features of two adjacent Mel frequency cepstrum coefficients and the ZMFCC features; the ZMFCC features are expressed as: ZMFCC=MFCC+MFCC μ +3σ(MFCC), where ZMFCC is the ZMFCC feature, MFCC represents the Mel frequency cepstral coefficient, MFCC μ is the mean value of the Mel frequency cepstral coefficients, σ(MFCC) is the variance of the Mel frequency cepstral coefficients;

[0041] A second extraction module is used to extract the self-learning features of the pre-processed gearbox acoustic signal by using a sparse autoencoder (SAE) network;

[0042] The identification module is used to input the MFCC static features, MFCC dynamic features and self-learning features into the gearbox fault identification model to complete the fault identification of the gearbox acoustic signal to be identified.

[0043] The technical solution adopted by the present invention to solve its technical problem is: to provide an electronic device, including a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned gearbox acoustic signal fault identification method when executing the computer program.

[0044] The technical solution adopted by the present invention to solve its technical problem is: providing a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned gearbox acoustic signal fault identification method are implemented.

[0045] Beneficial Effects

[0046] Due to the adoption of the above technical scheme, the present invention has the following advantages and positive effects compared with the prior art: the present invention extracts the MFCC static features and improved dynamic features of the gearbox acoustic signal, which not only eliminates the influence of redundant data, but also compensates for the difference between the static information and dynamic information of the acoustic signal, and effectively improves the accuracy and robustness of gearbox fault identification; the present invention uses the SAE model to deeply mine the acoustic signal to characterize the self-learning features of the signal, reduces the dependence on artificial features, and improves the fault detection and generalization capabilities of the deep learning model; the present invention combines the improved MFCC features and the self-learning features to establish a feature system that can characterize typical gearbox faults, thereby improving the gearbox fault identification and condition-based maintenance capabilities, which is of great significance to improving the safety performance and maintenance efficiency of aviation gearboxes. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of a method for identifying a gearbox acoustic signal fault according to a first embodiment of the present invention;

[0048] Figure 2 Time domain diagram and spectrum diagram of normal acoustic signal of gearbox;

[0049] Figure 3 Time domain diagram and spectrum diagram of the normal acoustic signal of the gearbox after decentralization;

[0050] Figure 4 Time domain diagram and spectrum diagram of normal acoustic signal of gearbox after partial coherence analysis;

[0051] Figure 5 Time domain diagram and spectrum diagram of normal sound signal of gearbox after PCA whitening;

[0052] Figure 6 Comparison of MFCC static features of broken gear and normal sound signals;

[0053] Figure 7 Improved MFCC dynamic feature comparison diagram of gear tooth breakage and normal sound signals;

[0054] Figure 8 Comparison of MFCC static features of gear erosion and normal acoustic signals;

[0055] Fig. 9 Improved MFCC dynamic feature comparison of gear erosion and normal acoustic signals;

[0056] Fig.10 Comparison of MFCC static features of gear crack and normal acoustic signals;

[0057] Fig.11 Improved MFCC dynamic feature comparison diagram of gear crack and normal acoustic signal;

[0058] Fig.12 Comparison of MFCC static characteristics of bearing outer ring wear and normal acoustic signals;

[0059] Fig.13 Improved MFCC dynamic feature comparison diagram of bearing outer ring wear and normal acoustic signal;

[0060] Fig.14 Comparison of MFCC static characteristics of bearing inner ring wear and normal acoustic signals;

[0061] Fig.15 Improved MFCC dynamic feature comparison diagram of bearing inner ring wear and normal acoustic signal;

[0062] Fig.16 Comparison of MFCC static characteristics of cage fracture and normal acoustic signals;

[0063] Fig.17 Improved MFCC dynamic feature comparison diagram of cage fracture and normal acoustic signal;

[0064] Fig.18 Self-learning features and scatter plots of different fault modes;

[0065] Fig.19 Schematic diagram of the confusion matrix of typical fault recognition in the test set based on the CNN model. DETAILED DESCRIPTION

[0066] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0067] The first embodiment of the present invention relates to a method for identifying a gearbox acoustic signal fault. The method first preprocesses the gearbox acoustic signal to be identified to obtain a preprocessed gearbox acoustic signal. Then, the MFCC static features and MFCC dynamic features of the gearbox acoustic signal are extracted based on the improved Mel frequency cepstrum. Then, the self-learning features of the gearbox acoustic signal are mined using SAE deep training. Finally, the CNN-based gearbox fault identification is carried out by combining the MFCC static features, MFCC dynamic features and self-learning features. Figure 1 As shown, the specific steps include:

[0068] Step 1: preprocess the gearbox acoustic signal to be identified to obtain the preprocessed gearbox acoustic signal to be identified. The preprocessing in this step includes decentralization, signal-to-noise ratio enhancement based on partial coherence analysis, and PCA whitening processing.

[0069] In this embodiment, a normal acoustic signal of a gearbox is used as an example for explanation. Preprocessing the acquired acoustic signal of the gearbox can be obtained by the following steps:

[0070] Step A: First, the gearbox acoustic signal is decentralized to remove DC component interference. The decentralized processing can be expressed as:

[0071]

[0072] where x(t)=(x1(t),x2(t),…,x n (t)) T ,t=1,2,...,N, N is the number of samples, x i (t) is the gearbox sound signal, This is the gearbox sound signal after centralized processing.

[0073] Step B: Use partial coherence analysis to enhance the signal-to-noise ratio of the decentered gearbox acoustic signal. The partial coherence analysis can be expressed as:

[0074]

[0075] Among them, X1 is the spectrum of the acoustic signal x1, X2 is the spectrum of the acoustic signal x2, S 12 is the frequency spectrum of the cross-correlation function of the acoustic signal x1 and the acoustic signal x2, S 11is the spectrum of the autocorrelation function of the acoustic signal x1, and X is the signal of the acoustic signal x1 enhanced based on partial coherence analysis.

[0076] Step C: Perform PCA whitening processing on the enhanced acoustic signal to remove the strong correlation between signals. The PCA whitening processing can be expressed as:

[0077]

[0078] Where X∈R n*m is the enhanced gearbox sound signal, n is the dimension, m is the number of samples, U is the orthogonal matrix, Λ is the diagonal matrix, X pca is the preprocessed gearbox sound signal.

[0079] As an example, the time domain diagram and spectrum diagram of the normal sound signal of the gearbox in this embodiment are as follows: Figure 2 As shown in the figure, the time domain diagram and spectrum diagram of the normal sound signal of the gearbox after decentralization are as follows: Figure 3 As shown in the figure, the time domain diagram and spectrum diagram of the normal sound signal of the gearbox after partial coherence analysis are as follows Figure 4 As shown in the figure, the time domain diagram and spectrum diagram of the normal sound signal of the gearbox after PCA whitening are as follows: Figure 5 As shown. The obvious components in the spectrum of the normal acoustic signal of the gearbox are the actual meshing frequency of the first-stage parallel gear of 979Hz and its 2 times, and the actual meshing frequency of the second-stage parallel gear of 1215Hz and its 2 times. The gear meshing frequency and its 2 times can be clearly seen from the normal acoustic signal of the gearbox and the decentralized spectrum; on the basis of decentralization, the normal acoustic signal is subjected to partial coherence analysis to enhance the signal characteristics. From the time domain diagram of the normal acoustic signal based on partial coherence analysis, it can be seen that the time domain amplitude is significantly increased, and the gear meshing frequency and its multiple amplitude are increased in the spectrum diagram; on the basis of partial coherence analysis, the normal acoustic signal is whitened to remove the correlation of the signal. From the time domain diagram and spectrum diagram of the normal acoustic signal subjected to whitening, it can be seen that the time domain amplitude and the amplitude of the meshing frequency of the gear are significantly enhanced, highlighting the gear signal characteristics.

[0080] Step 2: extract the Mel frequency cepstrum coefficients of the preprocessed gearbox sound signal to be identified, and obtain the MFCC static features and MFCC dynamic features of the gearbox sound signal according to the Mel frequency cepstrum coefficients. The details are as follows:

[0081] Step 1: Pre-emphasize the pre-processed gearbox sound signal. Pre-emphasis is to let the signal pass through a high-pass filter to balance the intensity of high and low frequency components in the sound signal. The high-pass filter can be expressed as:

[0082] H(Z)=1-μZ -1

[0083] In the formula, μ is usually 0.95 or 0.97. In this embodiment, μ is 0.97.

[0084] Step ②: The pre-emphasized gearbox sound signal is framed and windowed by a sliding window, wherein the frame length is generally 10 to 40 ms, the frame shift is generally 1 / 2 or 3 / 4 of the frame length, and each frame is multiplied by a Hamming window to increase the continuity of the left and right ends of the frame. The signal after windowing is expressed as:

[0085] y i (n) = w(n)*x i (n)

[0086]

[0087] Among them, w(n) becomes the window function, x i (n) is the i-th frame signal, y i (i) is the windowed signal of the i-th frame, and N is the frame length.

[0088] As an example, in this implementation, the frame length is 10 ms and the frame shift is 5 ms.

[0089] Step ③: After windowing each frame of signal, perform fast Fourier transform to obtain the spectrum of each frame, and square the spectrum of each frame to obtain the power energy spectrum of each frame of signal. The signal fast Fourier transform formula is:

[0090]

[0091] Among them, Y i (k) is the signal after fast Fourier transform of the i-th frame.

[0092] The power energy spectrum formula is:

[0093]

[0094] Among them, G i (k) is the power energy spectrum of the i-th frame signal.

[0095] Step ④: Convert the power energy spectrum through the Mel scale triangular filter bank to obtain the logarithmic energy spectrum of each frame signal. Specifically:

[0096] Step A: Determine and analyze the lowest frequency (0 Hz), the highest frequency (fs / 2), and the number of Mel filters M in each frame signal, where M is usually 22-26. In this embodiment, M=23.

[0097] Step B: Convert the lowest and highest Mel frequencies through the conversion relationship between physical frequency and Mel frequency. The formula for converting physical frequency to Mel frequency is:

[0098] mel(f)=2595*log 10 (1+f / 700)

[0099] Where f is the physical frequency and Mel(f) is the Mel frequency.

[0100] Step C: Calculate the distance between the center Mel frequencies of two connected Mel filters. The center frequencies between the two Mel frequencies are equidistant. The transfer function of the triangular filter is defined as:

[0101]

[0102] in, m=1,2,…,M, M is the number of filters; k is the kth coefficient of the power energy spectrum after transformation; f(m) is the center frequency of the Mel bandpass filter, f1 and f b The lowest frequency and the highest frequency in the preset filter group frequency range, N is the number of sampling points.

[0103] Step D: Convert each center Mel frequency into a physical frequency f (non-uniformly spaced), and calculate the subscript of the FFT midpoint corresponding to the physical frequency.

[0104] Step E: Calculate the logarithmic energy of each filter group output. Logarithmic operations include taking absolute values ​​and log operations. Taking absolute values ​​only uses the amplitude value and ignores the influence of phase, because phase information is not very useful in speech recognition. Log operations are used to separate envelopes and details. Envelopes represent timbre, and details represent pitch. Obviously, speech recognition is to recognize timbre. In addition, human perception is proportional to the logarithm of the Mel frequency, which is exactly simulated using log.

[0105] Step 5: Use discrete cosine transform to remove the low-correlation order features and calculate the Mel-frequency cepstrum coefficients. Usually, the first 13 Mel-frequency cepstrum coefficients can effectively represent most of the signal information. Keep the 2nd to 13th cepstrum coefficients and add the short-time energy of the sound signal to form the MFCC static feature vector. The discrete cosine transform formula is:

[0106]

[0107] Wherein, S(m) is the logarithmic energy spectrum output by the Mel filter bank; p is the number of Mel frequency cepstrum coefficients, and M is the number of Mel filters.

[0108] The short-time energy formula is:

[0109]

[0110] Among them, x i(n) is the preprocessed acoustic signal of the i-th frame, N is the frame length, and E(i) is the short-time energy of the i-th frame signal.

[0111] Step ⑥: Extract the dynamic features of MFCC features and calculate the first-order differential features using two consecutive adjacent MFCC coefficients. At the same time, in order to balance high and low frequencies and improve the signal-to-noise ratio, this implementation proposes ZMFCC feature extraction based on MFCC. The calculation formula of the first-order differential feature is:

[0112]

[0113] Among them, ΔMFCC k Represents the first-order difference feature of the kth MFCC.

[0114] The calculation formula of ZMFCC is:

[0115]

[0116] Among them, MFCC represents Mel frequency cepstral coefficients, represents the mean value of each MFCC coefficient, and σ(MFCC) represents the variance of each MFCC coefficient.

[0117] The improved MFCC features can be formed by adding the MFCC static features to the first-order difference features and the ZMFCC features.

[0118] As an example, in this implementation, MFCC static features and MFCC dynamic features are extracted for gearbox acoustic signals of different fault modes, including normal state, gear tooth breakage, gear erosion, gear cracks, bearing outer ring wear, bearing inner ring wear, and cage fracture. The sampling frequency is 51200Hz, the acquisition time is 60s, the frame length is 512, the frame shift is 256, and the number of samples obtained by frame splitting and windowing is 200. MFCC static features and MFCC dynamic features are extracted for each sample of each fault mode, and the MFCC static features of gear tooth breakage and normal acoustic signals are compared. Figure 6 As shown in the figure, the MFCC dynamic characteristics of the gear tooth breaking signal and the normal sound signal are compared. Figure 7 As shown in Figure 2, the MFCC static features of gear erosion and normal acoustic signals are compared. Figure 8 As shown in Figure 2, the MFCC dynamic characteristics of gear erosion and normal sound signals are compared. Fig. 9 As shown in Figure 2, the MFCC static features of gear crack and normal sound signal are compared. Fig.10 As shown in Figure 2, the MFCC dynamic characteristics of gear crack and normal sound signals are compared. Fig.11 As shown in Figure 2, the MFCC static characteristics of the bearing outer ring wear and normal sound signals are compared. Fig.12As shown in Figure 2, the MFCC dynamic characteristics of the bearing outer ring wear and normal sound signals are compared. Fig.13 As shown in the figure, the MFCC static characteristics of the bearing inner ring wear and normal sound signal are compared. Fig.14 As shown in Figure 2, the MFCC dynamic characteristics of the bearing inner ring wear and normal sound signals are compared. Fig.15 As shown in the figure, the MFCC static characteristics of the cage fracture and normal acoustic signals are compared. Fig.16 As shown in the figure, the MFCC dynamic characteristics of the cage fracture and normal acoustic signals are compared. Fig.17 As shown, samples 1-200 are MFCC features of normal signals, and samples 201-400 are MFCC features of fault signals.

[0119] This implementation method extracts the MFCC static features and improved dynamic features of the gearbox acoustic signal, which not only eliminates the influence of redundant data, but also compensates for the difference between the static information and dynamic information of the acoustic signal, effectively improving the accuracy and robustness of gearbox fault identification.

[0120] Step 3, using a sparse autoencoder SAE network to extract the self-learning features of the preprocessed gearbox acoustic signal. The sparse autoencoder SAE network in this step is trained in the following way: preprocessing the collected gearbox acoustic signal; performing a fast Fourier transform on the preprocessed gearbox acoustic signal, and using a sliding window to cut the transformed gearbox acoustic signal into multiple spectrum frame sequences, and extracting the root mean square features of each spectrum frame sequence; using the obtained root mean square features to train the sparse autoencoder SAE network, so that the trained sparse autoencoder SAE network can extract self-learning features from the root mean square features. Specifically as follows:

[0121] First, the SAE network is initialized and the coding layer and hidden layer network parameters are set. The SAE network settings in this embodiment are shown in Table 1.

[0122] Table 1 SAE network model setting table

[0123]

[0124]

[0125] Then, the preprocessed gearbox sound signal is transformed by FFT, the spectrum of the first 12000 Hz is cut into multiple spectrum frame sequences using a sliding window, and the root mean square (RMS) feature is extracted from each spectrum frame to achieve feature compression. In this embodiment, the spectrum frame length of the sliding window cutting is 100, and the frame shift is 100.

[0126] Finally, the SAE network model is trained based on RMS feature compression, and the essential representation of the RMS compression feature of the input signal is deeply learned through the encoding-decoding process. After the training is completed, the encoding part and the feature part of the SAE network model are retained, and the self-learning features of the gearbox sound signal can be adaptively obtained.

[0127] In this embodiment, the gearbox acoustic signal with different fault modes is subjected to self-learning feature extraction through the SAE network, including normal state, gear tooth breakage, gear erosion, gear crack, bearing outer ring wear, bearing inner ring wear and cage fracture. The sampling frequency is 51200Hz, and the acquisition time is 60s. The gearbox acoustic signal with different fault modes is subjected to RMS feature compression, and the length of the signal after feature compression is 512. The RMS compression features under different fault modes of the gearbox are input into the SAE network for training, and the self-learning features under different fault modes of the gearbox are deeply mined. For the convenience of visualization, the time domain diagram and scatter diagram of the self-learning features are shown as follows: Fig.18 As shown in the figure, it can be seen that the self-learning features of the seven types of acoustic signals, including normal gearbox, broken gear, cracked gear, eroded gear, inner ring wear of bearing, outer ring wear of bearing and broken cage, are obviously different. The SAE network model learns the deep features to distinguish these seven types of acoustic signals. The two-dimensional scatter plot also shows that the self-learning feature classification of these seven types of acoustic signals is very obvious, and the same fault features are relatively concentrated, which is conducive to the subsequent research on typical fault identification of gearboxes.

[0128] This implementation method uses the SAE network model to deeply mine the acoustic signal to characterize the self-learning features of the signal, reduce the dependence on artificial features, and improve the fault detection and generalization capabilities of the deep learning model.

[0129] Step 4: Input the MFCC static features, MFCC dynamic features and self-learning features into the gearbox fault recognition model to complete the fault recognition of the gearbox acoustic signal to be recognized. In this step, the gearbox fault recognition model is constructed based on the CNN network and is trained in the following manner: collecting gearbox sound signals under different fault modes, and dividing the obtained gearbox sound signals into a training set and a test set; extracting the Mel-frequency cepstral coefficients of the gearbox sound signals in the training set and the test set, and obtaining the MFCC static features and MFCC dynamic features of the gearbox sound signals in the training set and the test set according to the Mel-frequency cepstral coefficients; using a sparse autoencoder SAE network to extract the self-learning features of the gearbox sound signals in the training set and the test set; using the MFCC static features, MFCC dynamic features and self-learning features of the gearbox sound signals in the training set to train the gearbox fault recognition model, after the training is completed, using the MFCC static features, MFCC dynamic features and self-learning features of the gearbox sound signals in the test set to test, if the test requirements are met, the training of the gearbox fault recognition model is completed, if the test requirements are not met, the gearbox fault recognition model continues to be trained until the test requirements are met.

[0130] In this embodiment, when building a CNN network model, the network parameters of the reshape layer, convolution layer, pooling layer and fully connected layer are set. The CNN network model consists of two alternating convolution layers and pooling layers, a fully connected layer and an output layer. In the training stage, the acoustic features are first learned by the convolution layer and downsampled by the pooling layer. The feature learning of convolution and pooling is repeated to extract high-level features. In the final output layer, Softmax regression is used as a classifier to identify the health status of the gearbox based on the deep features learned from the acoustic features. The parameters of the CNN network model in this embodiment are shown in Table 2.

[0131] Table 2 CNN model settings

[0132] Network Layer Number of nodes Activation Function Is BN padding Reshape layer Convolutional layer 1 32 ReLU yes same Convolutional Layer 2 16 ReLU yes same Fully connected layer 8 ReLU no Fully connected layer 7 ReLU no

[0133] During training, the gearbox acoustic signal is divided into a training set and a test set. The improved MFCC features and self-learning features of the training set are extracted and input into the CNN network model for feature iterative learning. The fault recognition accuracy of the test set is calculated based on the trained CNN network model. When the recognition accuracy meets the requirements, the trained CNN network model is used as the final gearbox fault recognition model. The fault recognition accuracy calculation formula is:

[0134]

[0135] In this implementation, 8 groups of signals are collected for the 7 modes of normal gearbox, broken gear, gear erosion, gear crack, bearing outer ring wear, bearing inner ring wear and cage fracture. The sampling frequency of each group is 51200Hz, the sampling time is 60s, and each group of data is divided into 1s. The first 4 groups are used as training data sets, and the last 4 groups are used as test data sets. The number of samples for each mode is shown in Table 3. The improved MFCC features and self-learning features are extracted for each sample, the training set data is loaded and the CNN neural network is trained. The cross entropy is used as the loss function, the optimizer is adam, the learning rate is 0.0005, the number of iterations is 1100, and the Dropout ratio is 0.5. The classification and recognition performance of CNN is evaluated using the test set. The fault classification confusion matrix of the test set is shown in the following figure. Fig.19 As shown in the figure. From the fault classification confusion matrix of the test set, it can be seen that among the seven modes of gearbox normal, gear tooth breakage, gear erosion, gear crack, bearing outer ring wear, bearing inner ring wear and cage fracture, cage fracture, inner ring fault, outer ring fault, normal, gear tooth breakage and gear crack can be 100% identified, but 13.8% (33 samples) of the gear erosion data are identified as normal, 1.7% (4 samples) are identified as bearing outer ring fault, and 1.3% (3 samples) are identified as gear crack fault. The overall accuracy of typical gearbox fault identification is as high as 97.73%.

[0136] Table 3 Data partition table

[0137] normal Gear tooth breakage Gear Crack Gear erosion Inner race fault Outer ring failure Cage fracture Total number of samples Training set 280 240 240 240 240 240 280 1760 Test Set 280 240 240 240 240 240 280 1760

[0138] It is not difficult to find that this implementation method combines improved MFCC features and self-learning features to establish a feature system that can characterize typical gearbox faults, thereby improving the gearbox fault identification and condition-based maintenance capabilities, which is of great significance to improving the safety performance and maintenance efficiency of aviation gearboxes.

[0139] A second embodiment of the present invention relates to a gearbox acoustic signal fault identification device, comprising:

[0140] A preprocessing module, used for preprocessing the gearbox sound signal to be identified to obtain the preprocessed gearbox sound signal to be identified;

[0141] The first extraction module is used to extract the Mel frequency cepstrum coefficients of the preprocessed gearbox sound signal to be identified, and obtain the MFCC static features and MFCC dynamic features of the gearbox sound signal according to the Mel frequency cepstrum coefficients; the MFCC static features include the first 2-13 Mel frequency cepstrum coefficients and short-time energy of the gearbox sound signal; the MFCC dynamic features include the first-order difference features of two adjacent Mel frequency cepstrum coefficients and the ZMFCC features; the ZMFCC features are expressed as: Among them, ZMFCC is the ZMFCC feature, MFCC represents the Mel frequency cepstral coefficient, is the mean value of the Mel frequency cepstral coefficients, σ(MFCC) is the variance of the Mel frequency cepstral coefficients;

[0142] A second extraction module is used to extract the self-learning features of the pre-processed gearbox acoustic signal by using a sparse autoencoder (SAE) network;

[0143] The identification module is used to input the MFCC static features, MFCC dynamic features and self-learning features into the gearbox fault identification model to complete the fault identification of the gearbox acoustic signal to be identified.

[0144] The preprocessing module comprises:

[0145] A decentralization unit, used for performing decentralization processing on the gearbox acoustic signal to be identified to remove DC component interference;

[0146] A partial coherence analysis unit, used for enhancing the signal-to-noise ratio of the gearbox acoustic signal to be identified after decentralized processing based on partial coherence analysis;

[0147] The PCA whitening unit is used to perform PCA whitening processing on the gearbox acoustic signal to be identified after the signal-to-noise ratio is enhanced, remove the strong correlation between the signals, and obtain the pre-processed gearbox acoustic signal to be identified.

[0148] The first extraction module comprises:

[0149] A pre-emphasis unit, used for pre-emphasis the pre-processed gearbox acoustic signal;

[0150] A framing and windowing unit is used to frame and window the pre-emphasized gearbox sound signal in a sliding window manner to obtain a plurality of frames of windowed processed signals;

[0151] A Fourier transform unit is used to perform a fast Fourier transform on the windowed signal of each frame to obtain the spectrum of each frame, and to obtain the power energy spectrum of each frame signal by taking the modulus square of the spectrum of each frame;

[0152] A conversion unit, used for converting the power energy spectrum of each frame signal through a Mel-scale triangular filter bank to obtain a logarithmic energy spectrum of each frame signal;

[0153] An extraction unit, used for removing the order features with low correlation in the logarithmic energy spectrum by discrete cosine transform, and obtaining the first 13 Mel-frequency cepstrum coefficients;

[0154] A static feature acquisition unit, used for taking the first 2-13 order Mel frequency cepstral coefficients and short-time energy as MFCC static features of the gearbox sound signal;

[0155] The dynamic feature acquisition unit is used to calculate the first-order difference feature and ZMFCC feature of two adjacent Mel-frequency cepstral coefficients based on the obtained first 13-order Mel-frequency cepstral coefficients, and use the obtained first-order difference feature and ZMFCC feature of two adjacent Mel-frequency cepstral coefficients as the MFCC dynamic feature of the gearbox sound signal.

[0156] The conversion unit comprises:

[0157] A determination subunit, used for determining the lowest frequency and the highest frequency of each frame signal, and the number of Mel filters in the Mel-scaled triangular filter bank;

[0158] A frequency conversion subunit, used for converting the lowest frequency and the highest frequency Mel frequency through a physical frequency and Mel frequency conversion relationship;

[0159] A first calculation subunit is used to calculate the distance between the central Mel frequencies of two connected Mel filters;

[0160] A conversion calculation subunit, used to convert each center Mel frequency into a frequency, and calculate the subscript of the fast Fourier transform midpoint corresponding to the frequency;

[0161] The second calculation subunit is used to calculate the logarithmic energy output by each filter to obtain the logarithmic energy spectrum of each frame signal.

[0162] The sparse autoencoder SAE network is trained in the following way:

[0163] Preprocessing the collected gearbox sound signal;

[0164] The preprocessed gearbox acoustic signal is subjected to fast Fourier transform, and the transformed gearbox acoustic signal is cut into multiple spectrum frame sequences using a sliding window, and the root mean square feature is extracted from each spectrum frame sequence;

[0165] The obtained root mean square features are used to train the sparse autoencoder SAE network, so that the trained sparse autoencoder SAE network can extract self-learning features from the root mean square features.

[0166] The gearbox fault identification model is constructed based on the CNN network and trained in the following way:

[0167] Collecting gearbox acoustic signals under different fault modes, and dividing the obtained gearbox acoustic signals into a training set and a test set;

[0168] Extracting Mel frequency cepstral coefficients of the gearbox sound signals in the training set and the test set, and obtaining MFCC static features and MFCC dynamic features of the gearbox sound signals in the training set and the test set according to the Mel frequency cepstral coefficients;

[0169] Using a sparse autoencoder (SAE) network to extract self-learning features of the gearbox acoustic signals in the training set and the test set;

[0170] The gearbox fault identification model is trained using the MFCC static features, MFCC dynamic features and self-learning features of the gearbox sound signal in the training set. After the training is completed, the gearbox fault identification model is tested using the MFCC static features, MFCC dynamic features and self-learning features of the gearbox sound signal in the test set. If the test requirements are met, the training of the gearbox fault identification model is completed. If the test requirements are not met, the gearbox fault identification model continues to be trained until the test requirements are met.

[0171] A third embodiment of the present invention relates to an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the gearbox acoustic signal fault identification method of the first embodiment when executing the computer program.

[0172] A fourth embodiment of the present invention relates to a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the gearbox acoustic signal fault identification method of the first embodiment are implemented.

[0173] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) that contain computer-usable program code.

[0174] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0175] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction method, which is implemented in the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0176] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0177] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for identifying gearbox acoustic signal faults, characterized in that: The following steps are involved: Preprocessing the gearbox sound signal to be identified to obtain the preprocessed gearbox sound signal to be identified; Extracting the Mel frequency cepstrum coefficients of the preprocessed gearbox sound signal to be identified, and obtaining the MFCC static features and MFCC dynamic features of the gearbox sound signal according to the Mel frequency cepstrum coefficients; The MFCC static features include the first 2-13 Mel-frequency cepstrum coefficients and short-time energy of the gearbox sound signal; the MFCC dynamic features include the first-order difference features of two adjacent Mel-frequency cepstrum coefficients and the ZMFCC feature; the ZMFCC feature is expressed as: ZMFCC = MFCC + MFCC μ +3σ(MFCC), where ZMFCC is the ZMFCC feature, MFCC represents the Mel frequency cepstral coefficient, MFCC μ is the mean value of the Mel frequency cepstral coefficients, σ(MFCC) is the variance of the Mel frequency cepstral coefficients; A sparse autoencoder (SAE) network is used to extract the self-learning features of the preprocessed gearbox acoustic signal; The MFCC static features, MFCC dynamic features and self-learning features are input into a gearbox fault recognition model to complete fault recognition of the gearbox acoustic signal to be recognized.

2. The gearbox acoustic signal fault identification method according to claim 1, characterized in that: The preprocessing of the gearbox sound signal to be identified to obtain the preprocessed gearbox sound signal to be identified specifically includes: Decentralizing the gearbox acoustic signal to be identified to remove DC component interference; The signal-to-noise ratio of the gearbox acoustic signal to be identified after decentralized processing is enhanced based on partial coherence analysis; The gearbox acoustic signal to be identified after signal-to-noise ratio enhancement is subjected to PCA whitening processing to remove the strong correlation between signals and obtain the preprocessed gearbox acoustic signal to be identified.

3. The gearbox acoustic signal fault identification method according to claim 1, characterized in that: The extracting of the Mel frequency cepstrum coefficients of the preprocessed gearbox sound signal to be identified, and obtaining the MFCC static features and MFCC dynamic features of the gearbox sound signal according to the Mel frequency cepstrum coefficients, specifically includes: Pre-emphasize the pre-processed gearbox acoustic signal; The pre-emphasized gearbox acoustic signal is framed and windowed by a sliding window method to obtain a plurality of frames of windowed processed signals; Perform fast Fourier transform on the windowed signal of each frame to obtain the spectrum of each frame, and take the modulus square of the spectrum of each frame to obtain the power energy spectrum of the signal of each frame; The power energy spectrum of each frame signal is converted by a Mel-scale triangular filter bank to obtain the logarithmic energy spectrum of each frame signal; Using discrete cosine transform to remove the order features with low correlation in the logarithmic energy spectrum, and obtaining the first 13 Mel-frequency cepstrum coefficients; The first 2-13 Mel-frequency cepstral coefficients and short-time energy are used as MFCC static features of the gearbox sound signal; The first-order difference features and ZMFCC features of two adjacent Mel-frequency cepstral coefficients are calculated based on the first 13 Mel-frequency cepstral coefficients obtained, and the first-order difference features and ZMFCC features of two adjacent Mel-frequency cepstral coefficients obtained are used as the MFCC dynamic features of the gearbox acoustic signal.

4. The gearbox acoustic signal fault identification method according to claim 3 is characterized in that: The power energy spectrum of each frame signal is converted by a set of Mel-scale triangular filters to obtain the logarithmic energy spectrum of each frame signal. Specifically include: Determine the lowest frequency and the highest frequency of each frame signal, and the number of Mel filters in the Mel-scaled triangular filter bank; Convert the lowest and highest Mel frequencies through the conversion relationship between physical frequency and Mel frequency; Calculate the distance between the center Mel frequencies of two connected Mel filters; Convert each center Mel frequency into a frequency, and calculate the subscript of the midpoint of the fast Fourier transform corresponding to the frequency; The logarithmic energy of each filter output is calculated to obtain the logarithmic energy spectrum of each frame signal.

5. The gearbox acoustic signal fault identification method according to claim 1, characterized in that: The sparse autoencoder SAE network is trained in the following way: Preprocessing the collected gearbox sound signal; The preprocessed gearbox acoustic signal is subjected to fast Fourier transform, and the transformed gearbox acoustic signal is cut into multiple spectrum frame sequences using a sliding window, and the root mean square feature is extracted from each spectrum frame sequence; The obtained root mean square features are used to train the sparse autoencoder SAE network, so that the trained sparse autoencoder SAE network can extract self-learning features from the root mean square features.

6. The gearbox acoustic signal fault identification method according to claim 1, characterized in that: The gearbox fault identification model is constructed based on the CNN network and trained in the following way: Gearbox sound signals under different fault modes are collected, and the obtained gearbox sound signals are divided into a training set and a test set; Mel-frequency cepstral coefficients of the gearbox sound signals in the training set and the test set are extracted, and MFCC static features and MFCC dynamic features of the gearbox sound signals in the training set and the test set are obtained according to the Mel-frequency cepstral coefficients; Using a sparse autoencoder (SAE) network to extract self-learning features of the gearbox acoustic signals in the training set and the test set; The gearbox fault identification model is trained using the MFCC static features, MFCC dynamic features and self-learning features of the gearbox sound signal in the training set. After the training is completed, the gearbox fault identification model is tested using the MFCC static features, MFCC dynamic features and self-learning features of the gearbox sound signal in the test set. If the test requirements are met, the training of the gearbox fault identification model is completed. If the test requirements are not met, the gearbox fault identification model continues to be trained until the test requirements are met.

7. A gearbox acoustic signal fault identification device, characterized in that: include: A preprocessing module, used for preprocessing the gearbox sound signal to be identified to obtain the preprocessed gearbox sound signal to be identified; A first extraction module is used to extract the Mel frequency cepstrum coefficients of the preprocessed gearbox sound signal to be identified, and obtain the MFCC static features and MFCC dynamic features of the gearbox sound signal according to the Mel frequency cepstrum coefficients; The MFCC static features include the first 2-13 Mel frequency cepstrum coefficients and short-time energy of the gearbox sound signal; The MFCC dynamic feature includes the first-order difference feature of two adjacent Mel-frequency cepstral coefficients and the ZMFCC feature; the ZMFCC feature is expressed as: Among them, ZMFCC is the ZMFCC feature, MFCC represents the Mel frequency cepstral coefficient, is the mean value of the Mel frequency cepstral coefficients, σ(MFCC) is the variance of the Mel frequency cepstral coefficients; A second extraction module is used to extract the self-learning features of the pre-processed gearbox acoustic signal by using a sparse autoencoder (SAE) network; The identification module is used to input the MFCC static features, MFCC dynamic features and self-learning features into the gearbox fault identification model to complete the fault identification of the gearbox acoustic signal to be identified.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the gearbox acoustic signal fault identification method as claimed in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the gearbox acoustic signal fault identification method as claimed in any one of claims 1 to 6 are implemented.

Images