A CNN-based rotating machinery local fault signal sparse time-frequency representation method

Abstract

The application discloses a kind of CNN-based rotating machinery local fault signal sparse time-frequency representation method, comprising the following steps: step S1, utilize rotating machinery local fault signal response model and signal preprocessing to construct time-frequency spectrum data set for network model training;Step S2, construct a convolutional neural network model and train the model;Step S3, the rotating speed signal of acquisition equipment end is collected and the fault characteristic time-frequency ridge of each order when fault occurs in different positions is calculated, simultaneously the vibration signal of acquisition equipment end is collected and the time-frequency spectrum sample that can be used for the input of convolutional neural network model is obtained by signal preprocessing;Step S4, the trained convolutional neural network model is used to extract sparse feature to time-frequency spectrum sample, and the sparse time-frequency representation of the sample is obtained;Step S5, extract fault characteristic time-frequency ridge and compare each order fault characteristic time-frequency ridge one by one, further determine the fault occurrence position, to complete the high-precision fault diagnosis of rotating machinery.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rotating machinery fault diagnosis, and particularly relates to a rotating machinery local fault signal sparse time-frequency representation method based on CNN. BACKGROUND

[0002] The key to the state monitoring and fault diagnosis of rotating machinery under non-stationary working conditions lies in accurately depicting the variation law of the characteristic frequency of the rotating machinery with time. Time-frequency analysis methods are widely used in the analysis of non-stationary vibration signals of rotating machinery. Due to the uncertainty principle, it is difficult to obtain satisfactory time-frequency energy concentration by using short-time Fourier transform, especially when the collected vibration signals are mixed with a large amount of noise. The dispersion of time-frequency energy easily leads to large deviation of the extracted fault time-frequency feature line, which cannot meet the diagnostic requirements.

[0003] Since the ideal time-frequency representation of the rotating machinery fault signal is sparse in nature, people have begun to try to improve the sparsity of the rotating machinery fault signal in the time-frequency domain from the perspective of sparse representation, so as to improve the time-frequency energy concentration of the time-frequency representation of the rotating machinery fault signal. The method combining sparse representation and time-frequency representation is also called sparse time-frequency representation method. (Hou F, Selesnick I, Chen J, et al. Fault diagnosis for rolling bearings under unknown time-varying speed conditions with sparse representation [J]. Journal of Sound and Vibration, 2021, 494: 115854.) The sparse time-frequency representation of the local fault signal of the rolling bearing is performed and the fault diagnosis is completed. This method uses the inverse short-time Fourier transform operator to establish a general sparse time-frequency representation problem based on the l2-l1 norm, and obtains the sparse time-frequency representation of the local fault signal of the bearing through the Iterated Soft-thresholding Algorithm (ISTA). However, the fault feature time-frequency ridge line obtained by this method under the interference of noise is often not continuous, which is not conducive to the accurate fault diagnosis of rotating machinery under variable speed conditions. SUMMARY

[0004] The application aims at the problem that the existing sparse time-frequency representation method cannot simultaneously consider the time-frequency sparse characteristics and the time-frequency ridge continuous characteristics of the vibration signal under noise interference, and proposes a sparse time-frequency representation method for rotating machinery local fault signal based on CNN. The method of the application firstly constructs a time-frequency spectrum dataset by using a rotating machinery local fault signal response model, solves the problem that the labels of the measured vibration signal required for network training are difficult to obtain in actual engineering. Then a convolutional neural network model is constructed and trained, and finally the trained convolutional neural network model is used to extract sparse features of the time-frequency representation of the rotating machinery local fault signal. The obtained time-frequency representation has the characteristics of high time-frequency energy concentration, high time-frequency sparsity and continuous time-frequency ridge, and can be used for high-precision fault diagnosis of rotating machinery.

[0005] The application is implemented by at least one of the following technical solutions.

[0006] A sparse time-frequency representation method for rotating machinery local fault signal based on CNN, comprising the following steps:

[0007] S1, constructing a time-frequency spectrum dataset for network model training by using a rotating machinery local fault signal response model and signal preprocessing;

[0008] S2, constructing a convolutional neural network model and training the model;

[0009] S3, collecting the rotating speed signal of the equipment end and calculating the fault feature time-frequency ridge of each order when the fault occurs at different positions, and collecting the vibration signal of the equipment end and obtaining the time-frequency spectrum sample which can be used as the input of the convolutional neural network through signal preprocessing;

[0010] S4, using the trained convolutional neural network model to extract sparse features of the time-frequency spectrum sample, and obtaining the sparse time-frequency representation of the sample;

[0011] S5, extracting the fault feature time-frequency ridge in the sparse time-frequency representation of the sample in step S4 and comparing with the fault feature time-frequency ridge of each order obtained in step S3 one by one, further determining the fault position, and completing the high-precision fault diagnosis of the rotating machinery.

[0012] Further, the step S1 specifically comprises:

[0013] S11, establishing an ideal local fault impact simulation signal according to the rotating machinery local fault signal response model, and the rotating machinery local fault signal response model is as follows:

[0014]

[0015] In the formula, f dj represents the jth order resonance frequency of the system, and ζ jthe jth order damping ratio of the system, b ij the signal amplitude of the ith pulse response under the jth order resonance, T i the time at which the ith pulse occurs, u(t) represents a unit step response function, I is the total number of pulse responses, J is the total order of system resonance caused by the fault, and t is time;

[0016] S12, Gaussian white noise is added to the ideal local fault impact simulation signal to obtain a noisy signal;

[0017] S13, the noise-free signal and the noisy signal are respectively preprocessed by signal segmentation, Hilbert demodulation and envelope spectrum extraction, and time-frequency transformation and spectrum matrix interception, and the time-frequency spectrum obtained by preprocessing the noisy signal is taken as the input, and the time-frequency spectrum obtained by preprocessing the noise-free signal is taken as the label, to complete the construction of the data set.

[0018] Further, the purpose of Hilbert demodulation and envelope spectrum extraction in step S1 is to obtain a signal component with the fault characteristic frequency of the bearing or gear as the main frequency component.

[0019] Further, the time-frequency transformation adopted in step S1 can be any time-frequency transformation method such as short-time Fourier transform, wavelet transform and Wigner-Ville transform.

[0020] Further, step S2 specifically includes:

[0021] S21, the convolutional neural network model is composed of a dilated convolution function, a batch normalization function and an activation function, wherein the dilated convolution function can not only be used to extract signal features, but also can greatly increase the receptive field of the convolution kernel without increasing the number of hyperparameters, preventing all convolution pixels of individual convolution operations from falling into the noise region; the batch normalization function can make the input sample features have the same mean and variance, which will reduce the sensitivity of model parameters and be conducive to robust training of the convolutional neural network model; and the activation function can make the convolutional neural network have nonlinear data fitting capability, thereby improving the feature extraction capability of the convolutional neural network model;

[0022] S22, the step size of each dilated convolution function is 1, and the kernel size and padding size parameters of the dilated convolution kernel are obtained according to the following formula:

[0023] F = 2P + 1

[0024] In the formula: F represents the kernel size of the dilated convolution kernel, and P represents the padding size of the dilated convolution kernel;

[0025] S23, the dilated rate parameter of the dilated convolution kernel is obtained according to the following formula:

[0026] max{r m+1-2r m ,2r m -r m+1 ,r m}≤K m

[0027] In the formula: r m K represents the dilation rate of the m-th dilated convolution layer. m This indicates the kernel size of the dilated convolution kernel;

[0028] S24. The constructed convolutional neural network model is trained using the dataset constructed in step S1 to obtain a trained convolutional neural network model.

[0029] Furthermore, step S3 specifically includes:

[0030] S31. Acquire the vibration acceleration signal of the equipment, and perform high-pass filtering, Hilbert demodulation and envelope spectrum extraction, time-frequency transformation and truncation of the analysis spectrum matrix in sequence to obtain the time spectrum suitable for the input of the convolutional neural network model.

[0031] S32. Acquire the input shaft speed signal of the equipment, analyze the possible fault locations of the bearings and gears based on the mechanical structure of the equipment, and calculate the time-frequency ridges of each order of fault characteristics under various possible fault locations. The formula for calculating the fault characteristic frequency at each instant in the time-frequency ridge is as follows:

[0032] f c (t)=n·c·f r (t), t∈[t s ,t e ]

[0033] In the formula: f c (t) represents the instantaneous fault characteristic frequency at a possible location at time t, where c is the fault characteristic frequency coefficient corresponding to that location, determined by the mechanical structure of the component, and f r (t) represents the instantaneous rotational frequency of the axis of rotation at that position at time t. s t is the starting time point for vibration signal acquisition. e The point at which vibration signal acquisition ends is n, and n is the order of the instantaneous fault characteristic frequency. Generally, calculating the first few orders is sufficient to meet the fault diagnosis requirements.

[0034] Furthermore, the purpose of using high-pass filtering in step S3 is to filter out interference from non-stationary modulation low-frequency components present in the actual signal.

[0035] Further, in the step S4, the convolutional neural network model trained in the step S2 is used to extract sparse features from the time-frequency spectrum of the vibration signal obtained in the step S3, to obtain a sparse time-frequency representation of the vibration signal.

[0036] Further, the step S5 specifically includes:

[0037] S51, extracting fault feature time-frequency ridge lines of each order in the sparse time-frequency representation of the vibration signal by using maximum value processing;

[0038] S52, comparing the extracted fault feature time-frequency ridge lines with the fault feature time-frequency ridge lines at the possible fault positions calculated in the step S3, and when the curve rate error of the compared ridge lines at a certain order is less than the allowable error value, the fault occurrence position can be further determined, and in addition, when the curve rate error of the compared ridge lines at multiple orders is less than the allowable error value, the result of the fault diagnosis will be more reliable. The calculation formula of the curve rate error is as follows:

[0039]

[0040] In the formula, e n is the curve rate error of the ridge line of the nth order, N c is the total number of time points of each ridge line, τ p is the time corresponding to the pth point, f n (τ p ) is the ridge frequency of the ridge line of the nth order at the time τ p , f b (τ p ) is the fundamental frequency of the fault feature frequency at the time τ p corresponding to the fault position.

[0041] The present application has the following advantages and effects compared with the prior art:

[0042] (1) The convolutional neural network model trained by using the proposed data set has good generalization, and can be directly used for sparse time-frequency feature extraction of the time-frequency representation of the actual rotating machinery local fault signal under different working conditions after one training;

[0043] (2) Compared with the traditional sparse time-frequency representation method, the proposed method has better noise resistance, and can simultaneously consider the time-frequency sparse characteristics and ridge line continuity of the vibration signal;

[0044] (3) The present application combines convolutional neural network and time-frequency representation, uses the feature extraction capability of the convolutional neural network and the time-frequency representation to obtain a sparse time-frequency representation with more concentrated time-frequency energy, and overcomes the discontinuity defect of the time-frequency ridge line in the traditional sparse time-frequency representation, to complete the rotating machinery local fault diagnosis with higher precision. Attached Figure Description

[0045] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the illustrations used in the embodiments will be briefly described below. These illustrations are part of this application, but are merely non-limiting examples used to demonstrate the inventive concept and not to limit the scope of the invention.

[0046] Figure 1 This is a flowchart illustrating the implementation of a sparse time-frequency representation method for local fault signals in rotating machinery based on CNN, as provided in an embodiment of the present invention.

[0047] Figure 2 This is a flowchart of the data set construction method in this invention;

[0048] Figure 3 This is a time-domain waveform diagram of the original vibration signal collected in the experiment in this embodiment of the invention;

[0049] Figure 4 This is a schematic diagram of the time-frequency spectrum sample obtained after signal preprocessing of the original vibration signal in an embodiment of the present invention;

[0050] Figure 5 This is an input shaft speed curve collected in an experiment in this embodiment of the invention;

[0051] Figure 6 This is a schematic diagram of the time spectrum obtained by the method proposed in the embodiments of the present invention;

[0052] Figure 7 This is a schematic diagram of the time spectrum obtained by the sparse time-frequency representation comparison method based on short-time Fourier transform in an embodiment of the present invention; Detailed Implementation

[0053] To more clearly describe the objectives, technical solutions, and advantages of the embodiments of the present invention, they will be described in detail below with reference to the accompanying drawings. It should be noted that the embodiments described herein are only a part of the embodiments of the present invention, and not all of them. Those skilled in the art can obtain other embodiments based on the embodiments described herein without creative effort, and these embodiments all fall within the protection scope of the present invention.

[0054] The implementation flow of a sparse time-frequency representation method for local fault signals in rotating machinery based on CNN provided in this embodiment of the invention is as follows: Figure 1 As shown.

[0055] This invention will be further explained using a gear in a five-speed automotive transmission that experiences a partial malfunction as the research object. It includes the following steps:

[0056] S1, constructing a time-frequency spectrum data set for network model training by using a rotating machinery local fault signal response model and signal preprocessing. The construction process is as shown in Figure 2 The step specifically includes:

[0057] S11, establishing an ideal local fault impact simulation signal according to a rotating machinery local fault signal response model, and the rotating machinery local fault signal response model is as shown in the following formula:

[0058]

[0059] In the formula, f dj represents the jth order resonance frequency of the system, ζ j represents the jth order damping ratio of the system, b ij represents the signal amplitude of the ith pulse response under the jth order resonance, T i represents the time when the ith pulse occurs, u(t) represents a unit step response function, I is the total number of pulse responses, J is the total order of system resonance caused by faults, and t is a time sequence;

[0060] S12, adding Gaussian white noise to the ideal local fault impact simulation signal to obtain a noisy signal;

[0061] S13, respectively pre-processing the noise-free signal and the noisy signal, taking the time-frequency spectrum obtained by pre-processing the noisy signal as input and the time-frequency spectrum obtained by pre-processing the noise-free signal as label, and completing the construction of the time-frequency spectrum data set.

[0062] In some embodiments of the present application, the preprocessing includes: sequentially performing signal segmentation, Hilbert demodulation and envelope spectrum extraction, and time-frequency transformation on the noise-free signal and the noisy signal, respectively. Wherein, the Hilbert demodulation and envelope spectrum extraction are adopted to obtain a signal component with a rotating machinery (such as a bearing or a gear) fault characteristic frequency as the main frequency component.

[0063] In some embodiments of the present application, the time-frequency transformation can be any kind of time-frequency transformation, such as Fourier transformation, wavelet transformation, and Wigner transformation.

[0064] S2, constructing a convolutional neural network model by using neural network components such as hole convolution, batch normalization and activation function, and training the model, specifically including:

[0065] S21, the components of the convolutional neural network model include three kinds of cavity convolution function, batch normalization function and activation function. Among them, the cavity convolution function can not only be used to extract signal features, but also can greatly increase the receptive field of the convolution kernel without increasing the amount of hyperparameters, preventing all convolution pixels of individual convolution operations from falling into the noise region; the batch normalization function can make the input sample features have the same mean and variance, which will reduce the sensitivity of model parameters and be beneficial to robustly train the convolutional neural network model; and the activation function can make the convolutional neural network have nonlinear data fitting capability, thereby improving the feature extraction capability of the convolutional neural network model;

[0066] S22, the step of each cavity convolution function is 1, and the kernel size and padding size parameters of the cavity convolution kernel are obtained according to the following formula:

[0067] F=2P+1

[0068] In the formula, F represents the kernel size of the cavity convolution kernel, and P represents the padding size of the cavity convolution kernel;

[0069] S23, the dilution rate parameter of the cavity convolution kernel is obtained according to the following formula:

[0070] max{r m+1 -2r m ,2r m -r m+1 ,r m}≤K m

[0071] In the formula, r m represents the dilution rate of the mth layer cavity convolution, K m represents the kernel size of the cavity convolution kernel;

[0072] S24, the constructed convolutional neural network model is trained by using the time-frequency spectrum data set constructed in step S1, and a trained convolutional neural network model is obtained.

[0073] In some embodiments of the present application, the cavity convolution is used in combination with batch normalization and activation function respectively, and Table 1 is the hyperparameters of the proposed convolutional neural network model. Among them, D represents the cavity convolution function, B represents the batch normalization function, L represents the LeakyReLU activation function, and R represents the ReLU activation function.

[0074] Table 1 Hyperparameters of convolutional neural network model

[0075]

[0076] S3, collect the rotating speed signal of the equipment end and calculate the fault characteristic frequency ridge of each order when faults occur at different positions, and collect the vibration signal of the equipment end and obtain the time-frequency spectrum sample that can be used as the input of the convolutional neural network model through signal preprocessing.

[0077] This step specifically includes:

[0078] S31, collect the mechanical vibration signal from the equipment end of the experimental platform, sequentially perform high-pass filtering, Hilbert demodulation, envelope spectrum extraction, time-frequency transformation (in some embodiments of the application, short-time Fourier transform is adopted) and spectrum matrix preprocessing on the signal, to obtain the time-frequency spectrum suitable for the input of the convolutional neural network model. The high-pass filtering is used to filter out the interference of non-stationary modulation low-frequency components in the actual signal.

[0079] In some embodiments of the application, the collected vibration signal is as shown in Figure 3 , and the time-frequency spectrum obtained after preprocessing is as shown in Figure 4 .

[0080] S32, collect the input shaft rotating speed signal from the equipment end of the experimental platform, which is performed simultaneously with the step S31. In some embodiments of the application, the collected rotating speed signal is as shown in Figure 5 , analyze the possible fault positions of the bearing and gear of the equipment, and calculate the fault characteristic frequency ridge of each order under various possible fault positions, such as the instantaneous fault characteristic frequency when the inner ring, outer ring, rolling element and gear tooth of the bearing fail.

[0081] The fault characteristic frequency of each instantaneous point in the time-frequency ridge line is calculated according to the following formula:

[0082] f c (t)=n·c·f r (t),t∈[t s ,t e ]

[0083] In the formula, f c (t) represents the instantaneous fault characteristic frequency at t time under a possible position, c is the fault characteristic frequency coefficient corresponding to the position fault, which is determined by the mechanical structure of the rotating element, f r (t) is the instantaneous rotating frequency of the rotating shaft at t time where the position is located, t s is the starting time point of vibration signal collection, t e is the ending time point of vibration signal collection, and n is the order of the instantaneous fault characteristic frequency, and the first few orders can meet the fault diagnosis requirements.

[0084] S4, sparse feature extraction is performed on the time-frequency spectrum sample by using the trained convolutional neural network model in step S2, to obtain a sparse time-frequency representation of the time-frequency spectrum sample.

[0085] In some embodiments of the present application, the obtained sparse time-frequency representation of the local fault signal of the rotating machine is as shown in the following formula (3). Figure 6

[0086] To verify the performance of the time-frequency representation of the method of the present application, a sparse time-frequency representation method based on short-time Fourier transform is selected for comparison, and the time-frequency representation of the vibration signal obtained based on the comparison method is as shown in the following formula (4). Figure 7 Figure 6 and Figure 7 Compared with the above-mentioned traditional sparse time-frequency representation method, the sparse time-frequency representation method of the present application can better balance the time-frequency sparsity and time-frequency ridge continuity of the vibration signal, and thus can be more conducive to the accurate diagnosis of the rotating machine under variable speed conditions.

[0087] S5, the fault feature time-frequency ridge is extracted, and each order fault feature time-frequency ridge obtained in step S3 according to the calculation formula is compared one by one, when the curve rate error of the compared ridges is less than the allowable error value, the fault occurrence position can be further determined, thereby completing the high-precision fault diagnosis of the rotating machine.

[0088] This step specifically includes:

[0089] S51, the fault feature time-frequency ridge under each order in the sparse time-frequency representation of the mechanical vibration signal obtained in step S4 is extracted by using maximum value processing;

[0090] S52, the extracted fault feature time-frequency ridge is compared with the fault feature time-frequency ridge under the possible fault position calculated in step S3, when the curve rate error of the compared ridges under a certain order is less than the allowable error value, the fault occurrence position can be further determined, in addition, when the curve rate error of the compared ridges under multiple orders is less than the allowable error value, the result of the fault diagnosis will be more reliable. The calculation formula of the curve rate error is as follows:

[0091]

[0092] In the formula, e n is the curve rate error of the ridge under the n-th order, N c is the total number of time points of each ridge, τ p is the time corresponding to the p-th point, f n (τ p ) is the ridge frequency corresponding to the ridge under the n-th order at time τ p , f b (τ p ​​) is the base frequency of the fault feature frequency at the time τ p

[0093] In some embodiments of the present application, 1st, 2nd, 3rd and 5th order fault feature time-frequency ridges are extracted from the sparse time-frequency representation of the local fault signal of the rotating machine shown in FIG. 6, respectively, and compared, and it is known that the fifth output shaft gear has a local fault. Figure 6

[0094] In addition, the curve rate error between the fault feature time-frequency ridge extracted from the time-frequency spectrum obtained by the method of the present application and the ideal fault feature ridge is only 1.29%, which is far lower than the upper limit of the error of 5% allowed for rotating machine fault diagnosis, and therefore, it can be considered that the method of the present application can be used to complete high-precision rotating machine equipment fault diagnosis.

[0095] The serial numbers before the foregoing steps are for the convenience of description, and do not constitute a limitation on the order of the steps.

[0096] The above embodiment is an embodiment with better effects provided by the present application, but this does not exclude the existence of other implementation manners. As long as these manners do not deviate from the principles and spirits of the present application, any changes, adjustments, replacements, combinations or simplifications made to the present application should be regarded as equivalent alternative manners, and these alternative manners should also be regarded as within the protection scope of the present application.​​

Claims

1. A CNN-based method for sparse time-frequency representation of local fault signals of rotating machinery, characterized in that, The method comprises the following steps: S1, constructing a time-frequency spectrum data set for network model training by using a rotating machinery local fault signal response model and signal preprocessing; S2, constructing a convolutional neural network model and training the model; S3, collecting rotating speed signals at the equipment end and calculating fault feature time-frequency ridges of different orders when faults occur at different positions, simultaneously collecting vibration signals at the equipment end and obtaining time-frequency spectrum samples that can be used as inputs of the convolutional neural network through signal preprocessing; S4, extracting sparse features of the time-frequency spectrum samples by using the trained convolutional neural network model to obtain sparse time-frequency representations of the samples; S5, extracting fault feature time-frequency ridges in the sparse time-frequency representations of the samples in step S4 and comparing the fault feature time-frequency ridges with the fault feature time-frequency ridges obtained in step S3 one by one to further determine the fault position and complete high-precision fault diagnosis of the rotating machinery; In step S1, the following steps are specifically included: S11, establishing an ideal local fault impact simulation signal according to a rotating machinery local fault signal response model, and the rotating machinery local fault signal response model is shown in the following formula: In the formula: The system's first First resonant frequency, The system's first Damping ratio, Indicates the first The second resonance of the first order The signal amplitude of the sub-pulse response Indicates the first The time when the next pulse occurs Represents the unit step response function. This represents the total number of impulse responses. The total order of system resonance caused by the fault. For time; S12, adding Gaussian white noise to the ideal local fault impact simulation signal to obtain a noisy signal; S13, pre-processing the noise-free signal and the noisy signal respectively, and using the time-frequency spectrum obtained by pre-processing the noisy signal as the input and the time-frequency spectrum obtained by pre-processing the noise-free signal as the label to complete the construction of the time-frequency spectrum data set; In step S3, the following steps are specifically included: S31, collecting vibration acceleration signals of the equipment and pre-processing the vibration acceleration signals to obtain time-frequency spectrums suitable for inputs of the convolutional neural network model; S32, collecting input shaft rotating speed signals of the equipment, analyzing possible fault positions of bearings and gears according to the mechanical structure of the equipment, and calculating fault feature time-frequency ridges of different orders under various possible fault positions, and the fault feature frequency of each instantaneous point in the time-frequency ridge is calculated according to the following formula: In the formula: represents the instantaneous fault characteristic frequency of the possible position at the moment is the corresponding fault characteristic frequency coefficient of the position fault, is the instantaneous rotation frequency of the rotating shaft where the position is located at the moment is the starting time point of vibration signal collection, is the ending time point of vibration signal collection, is the order of the instantaneous fault characteristic frequency.​​ 2. The CNN-based local fault signal sparse time-frequency representation method of rotating machinery according to claim 1, characterized in that, The preprocessing in step S13 includes signal segmentation, Hilbert demodulation and envelope spectrum extraction, and time-frequency transformation and analysis spectrum matrix extraction.

3. The CNN-based local fault signal sparse time-frequency representation method of rotating machinery according to claim 2, characterized in that, The time-frequency transformation is any one of Fourier transformation, wavelet transformation and Wigner-Ville transformation.

4. The CNN-based local fault signal sparse time-frequency representation method of rotating machinery according to claim 1, characterized in that, The components of the convolutional neural network model in step S2 include a dilated convolution function, a batch normalization function and an activation function, wherein the dilated convolution function is used to extract signal features and increase the receptive field of the convolution kernel without increasing the number of model hyperparameters, the batch normalization function is used to make the input sample features have the same mean and variance, and the activation function is used to make the convolutional neural network have nonlinear data fitting capability.

5. The CNN-based local fault signal sparse time-frequency representation method of rotating machinery according to claim 4, characterized in that, The kernel size and padding size parameters of the dilated convolution kernel are obtained according to the following formula: In the formulae: represents the kernel size of the dilated convolution kernel, represents the padding size of the dilated convolution kernel; The dilatation rate parameter of the dilated convolution kernel is obtained according to the following formula: In the formula: represents the first represents the expansion rate of the layer cavity convolution, represents the kernel size of the cavity convolution kernel.

6. The CNN-based local fault signal sparse time-frequency representation method of rotating machinery according to claim 1, characterized in that, The preprocessing in step S31 includes high-pass filtering, Hilbert demodulation and envelope spectrum extraction, and time-frequency transformation and analysis spectrum matrix extraction.

7. The CNN-based rotating machinery local fault signal sparse time-frequency representation method according to claim 1, characterized in that, In step S4, the vibration signal time-frequency spectrum obtained in step S3 is subjected to sparse feature extraction by using the trained convolutional neural network model in step S2 to obtain a sparse time-frequency representation of the vibration signal.

8. The CNN-based local fault signal sparse time-frequency representation method of rotating machinery according to any one of claims 1-7, characterized in that, In step S5, the following steps are specifically included: S51, extracting the fault feature time-frequency ridge line under each order in the sparse time-frequency representation of the vibration signal; S52, comparing the extracted fault feature time-frequency ridge line with the fault feature time-frequency ridge line under the possible fault position calculated in step S3 order by order, when the curve rate error of the compared ridge line is less than the allowable error value under a certain order, the fault occurrence position can be further determined, wherein the calculation formula of the curve rate error is as follows: In the formula, is the extracted is the curve rate error of the ridge line of the order, is the total number of time points of each ridge line, is the time corresponding to the point, is the ridge frequency corresponding to the time of the ridge line of the order, is the fundamental frequency of the fault feature frequency at the time of the corresponding position.

Images