CN114897016A - Fan bearing fault intelligent diagnosis method based on multi-source frequency spectrum characteristics - Google Patents

Abstract

The invention provides a fan bearing fault intelligent diagnosis method based on multi-source frequency spectrum characteristics, which comprises the following steps of: collecting vibration signals-data preprocessing-data enhancement and standardization-training-testing. The method adopts various signal processing algorithms and simultaneously extracts the multi-source frequency domain characteristics of the vibration signal; by adopting a sample enhancement algorithm and performing data splicing on the samples after signal preprocessing, the characteristic expression capability of the label samples is enhanced; by adopting the idea of model fusion, the accuracy and the generalization of the whole classification effect can be effectively improved by the model fusion.

Description

Fan bearing fault intelligent diagnosis method based on multi-source frequency spectrum characteristics

Technical Field

The invention relates to a fan bearing fault intelligent diagnosis method based on multi-source frequency spectrum characteristics.

Background

Bearings are one of the key components in widespread use in wind turbine equipment. Bearings are susceptible to damage during machine operation due to overload, fatigue, wear, corrosion, and the like. In fact, more than 50% of rotating machine failures are related to bearing failures. The failure of the rolling bearing can cause the equipment to shake violently, stop the equipment, stop production and even cause personal injury and death. In general, early bearing weak faults are complex and difficult to detect. Therefore, monitoring and analysis of the condition of the bearing is very important, and it can find early weak failure of the bearing, and prevent the failure from causing loss.

The vibration characteristics due to mechanical faults are difficult to distinguish from the time domain signal, but there is a frequency signature in the vibration signal that corresponds to the fault. The traditional analysis method mainly utilizes signal processing algorithms such as wavelet transform, Hilbert transform demodulation algorithm, power spectrum analysis and the like to decompose the acquired vibration signals into different frequency band components from high frequency to low frequency. And extracting corresponding signal frequency spectrum characteristics on each component by adopting different signal processing algorithms according to bearing parameters such as rotating speed, bearing size, sampling rate and the like. And (4) sorting the features extracted from different components into feature vectors or matrixes, inputting the feature vectors or the matrixes into a machine learning algorithm or a deep learning algorithm for training, and performing classification diagnosis on the faults by using the trained model.

The existing fault intelligent diagnosis method adopts a single feature extraction method aiming at different fault types, and then classifies the different fault types by adopting a classification model aiming at the features. The prior method has the following problems: firstly, frequency domain and time domain characteristics shown by different faults are not consistent, and fault characteristics cannot be well extracted by adopting the same signal processing algorithm; secondly, label samples of bearing faults are few, and a reliable model is difficult to train in a deep learning algorithm; finally, a large amount of noise signals are often brought in by vibration signal acquisition, and the accuracy of the model is seriously influenced in the early period of failure.

Disclosure of Invention

In order to solve the technical problems, the invention provides an intelligent fan bearing fault diagnosis method based on multi-source frequency spectrum characteristics, and the intelligent fan bearing fault diagnosis method based on the multi-source frequency spectrum characteristics solves the problems that different faults in the prior art need to analyze different time-frequency characteristics of signals, label samples are few, the fault diagnosis accuracy is low due to the existence of noise, and the like.

The invention is realized by the following technical scheme.

The invention provides a fan bearing fault intelligent diagnosis method based on multi-source frequency spectrum characteristics, which comprises the following steps:

firstly, collecting vibration signals: collecting vibration signals of a fan rolling bearing in different states;

preprocessing data: preprocessing the vibration signal in the first step to obtain a diagnosis model training sample;

enhancing and standardizing data: performing data enhancement and standardization processing on a diagnostic model training sample to obtain a training set, a verification set and a test set;

fourthly, training: training a deep residual shrinkage network by using a training set to obtain a fault diagnosis model, and adjusting the fault diagnosis model by using a verification set;

testing: and (4) testing the test set by using the fault diagnosis model in the step (IV), and performing weighted average after obtaining the diagnosis result to obtain the final fault diagnosis result.

In the step I, the vibration signals in different states comprise normal, outer ring fault, inner ring fault and ball fault states.

The step II comprises the following steps:

(2.1) carrying out equalization processing and wavelet packet filtering on the vibration signals;

and (2.2) respectively carrying out Hilbert transform, power spectrum analysis and cepstrum analysis on the vibration signals processed in the step (2.1) to obtain a demodulation spectrum sample, a power spectrum sample and a cepstrum sample.

In the step (2.1), the vibration signal after the equalization is removed is subjected to wavelet filtering processing by adopting sym6 wavelet in Symlets family, and noise signals are removed.

In the step (2.2), the rotating frequency and the higher harmonic component of the fan bearing are obtained through Hilbert transform; diagnosing faults of the bearing rolling body and the outer ring through power spectrum analysis; and (4) reflecting the modulation frequency of the fault characteristics through cepstrum analysis, thereby diagnosing the fault type.

In the second step, the diagnosis model training sample comprises a demodulation spectrum sample, a power spectrum sample and a cepstrum sample.

The third step is specifically divided into the following steps:

(3.1) grouping the diagnostic model training samples according to labels;

(3.2) traversing all training samples in the group, randomly splicing each training sample with other training samples in the group, and acquiring enhanced data;

(3.3) carrying out standardization processing on the enhanced data;

and (3.4) dividing the training samples into a training set, a verification set and a test set.

In the fourth step, the obtained fault diagnosis model comprises a demodulation spectrum fault diagnosis model, a power spectrum diagnosis model and a cepstrum diagnosis model.

The invention has the beneficial effects that: extracting multi-source frequency domain characteristics of the vibration signals simultaneously by adopting various signal processing algorithms; by adopting a sample enhancement algorithm and performing data splicing on the samples after signal preprocessing, the characteristic expression capability of the label samples is enhanced; by adopting the idea of model fusion, the accuracy and the generalization of the whole classification effect can be effectively improved by the model fusion.

Drawings

FIG. 1 is a flow chart of the data pre-processing of the present invention;

FIG. 2 is a flow chart of diagnostic model training and testing of the present invention;

FIG. 3 is a timing signal diagram of normal vibration signals, inner ring faults, outer ring faults and ball body faults extracted by the invention;

FIG. 4 is a demodulation spectrogram of normal vibration signal, inner ring fault, outer ring fault and ball body fault extracted by the present invention;

FIG. 5 is a power spectrum of normal vibration signal, inner ring fault, outer ring fault and ball fault extracted by the present invention;

FIG. 6 is a reverse frequency spectrum graph of normal vibration signals, inner ring faults, outer ring faults and ball body faults extracted by the present invention;

fig. 7 is a network architecture diagram of depth residual shrinkage.

Detailed Description

The technical solution of the present invention is further described below, but the scope of the claimed invention is not limited to the described.

As shown in FIGS. 1-7, the intelligent diagnosis method for the fan bearing fault based on the multi-source frequency spectrum characteristics comprises the following steps:

firstly, collecting vibration signals: collecting vibration signals of a fan rolling bearing in different states;

preprocessing data: preprocessing the vibration signal in the first step to obtain a diagnosis model training sample;

enhancing and standardizing data: performing data enhancement and standardization processing on a diagnostic model training sample to obtain a training set, a verification set and a test set;

fourthly, training: training a deep residual error shrinkage network by using a training set to obtain a fault diagnosis model, and adjusting the fault diagnosis model by using a verification set;

testing: and (4) testing the test set by using the fault diagnosis model in the step (IV), and performing weighted average after obtaining the diagnosis result to obtain the final fault diagnosis result.

In the step I, the vibration signals in different states comprise normal, outer ring fault, inner ring fault and ball fault states.

The step II comprises the following steps:

(2.1) carrying out equalization processing and wavelet packet filtering on the vibration signals;

and (2.2) respectively carrying out Hilbert transform, power spectrum analysis and cepstrum analysis on the vibration signals processed in the step (2.1) to obtain a demodulation spectrum sample, a power spectrum sample and a cepstrum sample.

In the step (2.1), the vibration signal after the equalization is removed is subjected to wavelet filtering processing by adopting sym6 wavelet in Symlets family, and noise signals are removed.

In the step (2.2), the rotating frequency and the higher harmonic component of the fan bearing are obtained through Hilbert transform; diagnosing faults of the bearing rolling body and the outer ring through power spectrum analysis; and (4) reflecting the modulation frequency of the fault characteristics through cepstrum analysis, thereby diagnosing the fault type.

In the second step, the diagnosis model training sample comprises a demodulation spectrum sample, a power spectrum sample and a cepstrum sample.

The third step is specifically divided into the following steps:

(3.1) grouping the diagnostic model training samples according to labels;

(3.2) traversing all training samples in the group, randomly splicing each training sample with other training samples in the group, and acquiring enhanced data;

(3.3) carrying out standardization processing on the enhanced data;

and (3.4) dividing the training samples into a training set, a verification set and a test set.

In the fourth step, the obtained fault diagnosis model comprises a demodulation spectrum fault diagnosis model, a power spectrum diagnosis model and a cepstrum diagnosis model.

Examples

As shown in fig. 1 and 2, the present invention comprises the steps of:

1. the method comprises the steps of collecting vibration signals of a fan rolling bearing in different states, wherein the vibration signals comprise normal, outer ring fault, inner ring fault and ball fault states.

177 normal samples, 140 outer ring faults, 45 inner ring faults and 141 ball faults are collected, and vibration signals are visualized as shown in fig. 3.

2. Carrying out data preprocessing on the acquired vibration signals to obtain a model training sample, wherein the specific flow is shown in figure 1:

(2.1) carrying out de-equalization processing and wavelet packet filtering on the signal x;

in order to facilitate subsequent signal processing, firstly, the signal is subjected to de-equalization processing to obtain a de-equalized x, and the processing algorithm is as follows:

after the equalization, the sym6 wavelet in Symlets family is adopted to perform wavelet filtering processing on the signal, and the noise signal is removed. The signal after the de-equalization and the wavelet packet filtering is the first stage signal x1 with a data scale of [1, 6000 ].

(2.1) demodulating by Hilbert transform to obtain demodulated spectrum data, and obtaining data after Hilbert demodulation of signals as shown in FIG. 4:

when the gear is broken or damaged and fails, the vibration signal is subjected to amplitude modulation along with the gear meshing process, and the amplitude modulation indicating the nth harmonic component passes through the gear once along with one rotation of the gear if the amplitude modulation fails, so that the function of taking the rotation frequency of the shaft as a cycle can be known. Extracting signal envelopes by using Hilbert transform, demodulating a fault frequency spectrum, and adopting a Hilbert transform principle:

what Hilbert demodulated is about the fan bearing frequency conversion and the higher harmonic components. The signal x1 undergoes hilbert demodulation to obtain signal demodulation spectrum samples.

(2.3) obtaining power spectrum data through a power spectrum analysis algorithm, wherein the data obtained through the power spectrum analysis is shown in figure 5:

when equipment has serious faults, larger excitation impact can be generated, the amplitude of a side frequency band generated by modulation is larger, the natural frequency (the natural frequency of a gear, the natural frequency of a bearing outer ring, the natural frequency of a box body and the like) of the equipment can be excited sometimes, the frequency spectrum moves backwards, the gravity center of the power spectrum is increased, and the phenomenon of 'holes' is reflected on energy. The faults of the bearing rolling body and the outer ring can be diagnosed by analyzing the signal power spectrum. The power spectrum principle is as follows:

wherein k is 0,1, …, N-1,

the signal x1 is subjected to power spectrum analysis to obtain a power spectrum sample.

(2.4) obtaining cepstrum data through secondary spectrum analysis, wherein the processed data is shown in fig. 6:

the cepstrum is also called quadratic spectrum analysis, and the essence of the cepstrum is to take the logarithm of the power spectrum and then perform spectrum analysis to obtain periodic components in the spectrum. For the modulation variable frequency bands of the gears and the rolling bearings in the gear box, the modulation frequency reflecting the fault characteristics can be analyzed by utilizing the cepstrum, so that the fault type can be diagnosed. The signal x1 is cepstrally analyzed to obtain cepstral samples.

(2.5) processing the preprocessed signal x1 by the steps of (2.2), (2.3) and (2.4) to obtain three sample sets of demodulation spectrum, power spectrum and cepstrum, namely C1, C2 and C3, wherein the scale of each sample is [1, 3000], and all samples are marked.

3. Training a deep residual shrinkage network to obtain a diagnosis model and a test, wherein the corresponding flow is as shown in FIG. 2:

3.1 data enhancement and normalization

(3.1.1) the training sample is enhanced, and the problem of small sample data size is solved:

data enhancement is an effective way to expand the size of data samples. Deep learning is a big data-based method, the larger the scale of data is, the higher the data quality is, the better the trained model can extract data features, and the generalization capability is stronger. Because the bearing fault sample is difficult to collect and the collection cost is high, the invention provides a data enhancement method for enhancing a training sample in a splicing mode, which comprises the following steps: first, each set of samples is grouped by label, and then all samples in the group are traversed, with each sample randomly stitched to other samples in the group. Table 1 shows the enhanced data of the demodulated spectrum samples:

table 1 data enhanced samples

Bearing condition	Number of samples	Sample size	Sample label
				Normal bearing	875	[1,6000]	[1,0,0,0]
Inner ring failure	690	[1,6000]	[0,1,0,0]
				Outer ring failure	215	[1,6000]	[0,0,1,0]
Failure of ball body	695	[1,6000]	[0,0,0,1]

(3.1.2) the data are standardized, and as can be seen from fig. 4-6, the fluctuation of the frequency spectrum characteristic data after signal processing is large, and the numerical range is wide. The data are standardized, so that the convergence speed of the deep neural network can be increased, and the training efficiency is improved.

3.2 model training and fusion

(3.2.1) dividing the sample into a training set and a testing set according to the ratio of 8: 2 into training and test sets.

(3.2.2) training a deep residual error shrinkage network by adopting a training set to obtain a corresponding fault diagnosis model;

noise signals are inevitably introduced in the signal acquisition process, so that the fault is weak, especially when the bearing fails in the early stage. The traditional signal noise reduction algorithm depends on the selection of the construction of a wavelet filtering function or a modal decomposition method and the selection of a threshold value in soft thresholding. However, different fault types and different fault stages of the bearing have different noise characteristics, and the noise influence is difficult to effectively reduce by adopting a noise suppression algorithm. The method adopts the deep residual shrinkage network training data, and the model can adaptively reduce the noise of signals and effectively extract the deep features of the data. The structure of the depth residual shrinking network is shown in fig. 7.

(3.2.3) respectively training the demodulation spectrum, the power spectrum and the cepstrum sample set according to the steps to obtain diagnosis models of model1, model2 and model 3.

3.3 model testing and Effect comparison

And the test sample passes through the three diagnosis models to obtain a diagnosis result, and then weighted average is carried out to obtain a final result. In the test process, the accuracy of cepstrum is low when the inner ring fault is diagnosed, but the overall accuracy of the model1 obtained by training a demodulation spectrum is high, and the model fusion effect cannot be well reflected by adopting a method of directly adding and averaging, in the method, the model weight of a model1 obtained by training the demodulation spectrum is increased, the model weight of a model3 obtained by training the cepstrum is reduced, so that the model diagnosis performance is improved, and the weight vectors for the models 1, 2 and 3 are [0.5,0.3 and 0.2 ].

The above test sample inputs model1, model2, model3 result in results result1, result2, result3, and the final output result is:

result＝0.5*result1+0.3*result2+0.2*result3

the method only adopts demodulation spectrum, power spectrum and cepstrum to greatly improve the precision, has good generalization, and has the algorithm effect comparison table shown in table 2.

TABLE 2 comparison of the effects of the algorithm

Compared with the prior art, the invention has the following effects:

1. the invention adopts a plurality of signal processing algorithms and simultaneously extracts the multi-source frequency domain characteristics of the vibration signal: the current algorithms for diagnosing the bearing fault based on machine learning or deep learning basically only adopt the frequency domain characteristics after Fourier transform or wavelet transform, and cannot well reflect different fault types; by extracting the multi-source frequency domain characteristics, the algorithm extracts the difference of different fault frequency domain characteristics, and the accuracy of the whole method is improved;

2. according to the invention, a sample enhancement algorithm is adopted, and the data splicing is carried out on the samples after signal preprocessing, so that the characteristic expression capability of the label samples is enhanced; sample diversity is effectively enhanced, the convergence speed of the model is accelerated, and the precision of the model is improved;

3. the invention adopts a depth residual error shrinkage network, the network can automatically learn the noise characteristics of the sample and inhibit the interference of noise to fault diagnosis, and meanwhile, the network adopts a one-dimensional depth residual error module, so that the feature extraction is more efficient compared with the traditional convolution network;

4. the invention adopts the idea of model fusion, and the model fusion can effectively improve the accuracy and the generalization of the whole classification effect: the corresponding fault diagnosis model is trained for each signal frequency domain characteristic, each signal to be diagnosed obtains three diagnosis results through the model respectively, then the final diagnosis result is obtained through weighted averaging, the problem that the inner-loop fault diagnosis effect is poor due to cepstrum can be solved, and the overall accuracy and the generalization of the model are improved.

Claims (8)

1. A fan bearing fault intelligent diagnosis method based on multi-source frequency spectrum characteristics is characterized by comprising the following steps: the method comprises the following steps:

firstly, collecting vibration signals: collecting vibration signals of a fan rolling bearing in different states;

preprocessing data: preprocessing the vibration signal in the first step to obtain a diagnosis model training sample;

enhancing and standardizing data: performing data enhancement and standardization processing on a diagnostic model training sample to obtain a training set, a verification set and a test set;

fourthly, training: training a deep residual shrinkage network by using a training set to obtain a fault diagnosis model, and adjusting the fault diagnosis model by using a verification set;

testing: and (4) testing the test set by using the fault diagnosis model in the step (IV), and performing weighted average after obtaining the diagnosis result to obtain the final fault diagnosis result.

2. The multi-source spectrum feature-based fan bearing fault intelligent diagnosis method of claim 1, characterized in that: in the step I, the vibration signals in different states comprise normal, outer ring fault, inner ring fault and ball fault states.

3. The multi-source spectrum feature-based fan bearing fault intelligent diagnosis method of claim 1, characterized in that: the step II comprises the following steps:

(2.1) carrying out equalization processing and wavelet packet filtering on the vibration signals;

and (2.2) respectively carrying out Hilbert transform, power spectrum analysis and cepstrum analysis on the vibration signals processed in the step (2.1) to obtain a demodulation spectrum sample, a power spectrum sample and a cepstrum sample.

4. The multi-source spectrum feature-based fan bearing fault intelligent diagnosis method of claim 3, characterized in that: in the step (2.1), the vibration signal after the equalization is removed is subjected to wavelet filtering processing by adopting sym6 wavelet in Symlets family, and noise signals are removed.

5. The multi-source spectrum feature-based fan bearing fault intelligent diagnosis method of claim 3, characterized in that: in the step (2.2), the rotating frequency and the higher harmonic component of the fan bearing are obtained through Hilbert transform; diagnosing faults of a bearing rolling body and faults of an outer ring through power spectrum analysis; and (4) reflecting the modulation frequency of the fault characteristics through cepstrum analysis, thereby diagnosing the fault type.

6. The multi-source spectrum feature-based fan bearing fault intelligent diagnosis method of claim 1, characterized in that: in the second step, the diagnosis model training sample comprises a demodulation spectrum sample, a power spectrum sample and a cepstrum sample.

7. A fan bearing fault intelligent diagnosis method based on multi-source frequency spectrum characteristics as claimed in claim 1, characterized in that: the third step is specifically divided into the following steps:

(3.1) grouping the diagnostic model training samples according to labels;

(3.2) traversing all training samples in the group, randomly splicing each training sample with other training samples in the group, and acquiring enhanced data;

(3.3) carrying out standardization processing on the enhanced data;

and (3.4) dividing the training samples into a training set, a verification set and a test set.

8. The multi-source spectrum feature-based fan bearing fault intelligent diagnosis method of claim 1, characterized in that: in the fourth step, the obtained fault diagnosis model comprises a demodulation spectrum fault diagnosis model, a power spectrum diagnosis model and a cepstrum diagnosis model.

Images