WO2019061006A1

WO2019061006A1 - Bearing failure diagnosis method and device, readable storage medium, and electronic device

Info

Publication number: WO2019061006A1
Application number: PCT/CN2017/103304
Authority: WO
Inventors: Zhi QIU; Hualiang Hu; Lai Wei; Zikui MA
Original assignee: Schaeffler Technologies AG & Co. KG
Priority date: 2017-09-26
Filing date: 2017-09-26
Publication date: 2019-04-04
Also published as: CN111094927A

Abstract

Bearing failure diagnosis method and device, a readable storage medium, and an electronic device are provided. The bearing failure diagnosis method includes: acquiring a vibration acceleration signal of a bearing which is running (S101); performing wavelet decomposition to the vibration acceleration signal of the bearing to obtain a plurality of signals (S102); extracting feature parameters from the plurality of signals respectively and constructing feature vectors (S103); constructing a feature matrix based on the feature vectors (S104); feeding the feature matrix to a trained neural network model (S105); and comparing an output result of the trained neural network model with a predetermined target matrix to obtain a failure diagnosis result of the bearing (S106). Accuracy of bearing failure diagnosis may be improved.

Description

BEARING FAILURE DIAGNOSIS METHOD AND DEVICE, READABLE STORAGE MEDIUM, AND ELECTRONIC DEVICE

FIELD

The present disclosure generally relates to bearing failure detection field, and more particularly, to bearing failure diagnosis method and device, a readable storage medium, and an electronic device.

BACKGROUND

Bearing as a basic part of mechanical devices has been widely used in various equipment, such as a train. A working status of the bearing plays a vital role in the safety of train running. Therefore, it is quite important to diagnose bearing failures.

In existing techniques, to diagnose bearing failures, data analyzers generally make subjective judgement by observing an envelope spectrum waveform of a demodulated time domain vibration acceleration signal, which requires the data analyzers mastering related theory to perform analysis on data and thus is relatively subjective. There are also some bearing failure diagnosis methods based on neural network theory. However, the methods always result in low accuracy of bearing failure diagnosis.

SUMMARY

Embodiments of the present disclosure may improve accuracy of bearing failure diagnosis.

In an embodiment, a bearing failure diagnosis method is provided, including: acquiring a vibration acceleration signal of a bearing which is running； performing wavelet decomposition to the vibration acceleration signal of the bearing to obtain a plurality of signals； extracting feature parameters from the plurality of signals respectively and constructing feature vectors； constructing a feature matrix based on the feature vectors； feeding the feature matrix to a trained neural network model； and comparing an output result of the trained neural network model with a predetermined target matrix to obtain a failure diagnosis result of the bearing.

In some embodiments, the failure diagnosis result of the bearing may include at least one of the followings: whether a failure occurs in the bearing, a part of the bearing where a failure occurs, and a severity level of a failure.

In some embodiments, acquiring a vibration acceleration signal of a bearing may include: acquiring a time domain vibration acceleration signal of the bearing.

In some embodiments, performing wavelet decomposition to the vibration acceleration signal of the bearing to obtain a plurality of signals may include: selecting a mother wavelet function, wherein a waveform of the mother wavelet function is relevant to a waveform of the time domain vibration acceleration signal of the bearing； and performing wavelet decomposition to the mother wavelet function to obtain time domain signals in N frequency bands, where N＝2^j, and j is the number of layers of the wavelet decomposition.

In some embodiments, extracting feature parameters from the plurality of signals respectively and constructing feature vectors may include: demodulating the time domain signals in the N frequency bands and performing time-frequency transformation to obtain corresponding frequency domain signals in the N frequency bands； and extracting feature parameters from the corresponding frequency domain signals in the N frequency bands respectively, and constructing feature vectors.

In some embodiments, constructing a feature matrix based on the feature vectors may include: constructing a feature matrix based on the feature vectors, where a size of the feature matrix is M by N, and M is the number of the feature vectors.

In some embodiments, a size of the target matrix may be M by K, and K may be the number of parts to be diagnosed in the bearing.

In some embodiments, prior to performing wavelet decomposition to the vibration acceleration signal of the bearing, the method may further include: denoising the vibration acceleration signal of the bearing.

In some embodiments, denoising the vibration acceleration signal of the bearing may include: denoising the vibration acceleration signal of the bearing by using a wavelet shrinkage denoising method.

In some embodiments, the feature parameters may include at least one selected from: energy of each of the plurality of signals, a crest factor of each of the plurality of signals, standard deviation of each of the plurality of signals, a root mean square of each of the plurality of signals, and 90 percentile of each of the plurality of signals.

In some embodiments, prior to feeding the feature matrix to a trained neural network model, the method may further include: normalizing elements in the feature matrix.

In some embodiments, the trained neural network model may include an input layer, a hidden layer and an output layer, data of the input layer is the feature matrix, and data of the output layer is the predetermined target matrix.

In some embodiments, comparing an output result of the trained neural network model with a predetermined target matrix to obtain a failure diagnosis result of the bearing may include: comparing the output result of the trained neural network model with the predetermined target matrix to obtain error values between the output result and the target matrix； and determining that a failure occurs in a part of the bearing which corresponds to the minimum error value among the obtained error values.

In an embodiment, a bearing failure diagnosis device is provided, including: an acquiring circuitry configured to acquire a vibration acceleration signal of a bearing which is running； a wavelet decomposing circuitry configured to perform wavelet decomposition to the vibration acceleration signal of the bearing to obtain a plurality of signals； a feature vector constructing circuitry configured to extract feature parameters from the plurality of signals respectively and construct feature vectors； a feature matrix constructing circuitry configured to construct a feature matrix based on the feature vectors； an inputting circuitry configured to feed the feature matrix to a trained neural network model； and a failure diagnosis result obtaining circuitry configured to compare an output result of the trained neural network model with a predetermined target matrix to obtain a failure diagnosis result of the bearing.

In some embodiments, the acquiring circuitry may be configured to acquire a time domain vibration acceleration signal of the bearing.

In some embodiments, the wavelet decomposing circuitry may be configured to: select a mother wavelet function, wherein a waveform of the mother wavelet function is relevant to a waveform of the time domain vibration acceleration signal of the bearing； and perform wavelet decomposition to the mother wavelet function to obtain time domain signals in N frequency bands, where N＝2^j, and j is the number of layers of the wavelet decomposition.

In some embodiments, the feature vector constructing circuitry may be configured to: demodulate the time domain signals in the N frequency bands and perform time-frequency transformation to obtain corresponding frequency domain signals in the N frequency bands； and extract feature parameters from the corresponding frequency domain signals in the N frequency bands respectively, and construct feature vectors.

In some embodiments, the feature matrix constructing circuitry may be configured to: construct a feature matrix based on the feature vectors, where a size of the feature matrix is M by N, and M is the number of the feature vectors.

In some embodiments, the bearing failure diagnosis device may further include a denoising circuitry configured to: denoise the vibration acceleration signal of the bearing before the wavelet decomposing circuitry performs wavelet decomposition to the vibration acceleration signal of the bearing.

In some embodiments, the denoising circuitry may be configured to: denoise the vibration acceleration signal of the bearing by using a wavelet shrinkage denoising method.

In some embodiments, the bearing failure diagnosis device may further include a normalizing circuitry configured to: normalize elements in the feature matrix before the inputting circuitry feeds the feature matrix to the trained neural network model.

In some embodiments, the failure diagnosis result obtaining circuitry may be configured to: compare the output result of the trained neural network model with the predetermined target matrix to obtain error values between the output result and the target matrix； and determine that a failure occurs in a part of the bearing which corresponds to the minimum error value among the obtained error values.

In an embodiment, a computer readable storage medium which has computer instructions stored therein is provided, wherein once the computer instructions are executed, any one of the above bearing failure diagnosis methods is performed.

In an embodiment, an electronic device including a memory and a processor is provided, wherein the memory has computer instructions stored therein, and once executing the computer instructions, the processor performs any one of the above bearing failure diagnosis methods.

Compared with the existing techniques, embodiments of the present disclosure may have following advantages. Wavelet decomposition is performed to the acquired vibration acceleration signal of the bearing to obtain a plurality of signals, feature parameters are extracted from the plurality of signals respectively and feature vectors are constructed, a feature matrix is constructed based on the feature vectors and fed to a trained neural network model which then processes the feature matrix to obtain an output result, and the output result is compared with a predetermined target matrix to obtain a failure diagnosis result of the bearing. The determination of the bearing failure does not depend on subjective judgement of a testing person, which may improve accuracy of bearing failure diagnosis.

Further, prior to performing wavelet decomposition to the vibration acceleration signal of the bearing, the vibration acceleration signal is denoised, so as to reduce interference to the vibration acceleration signal caused by other signals, which may further improve accuracy of bearing failure diagnosis.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flow chart of a bearing failure diagnosis method according to an embodiment；

FIG. 2 is a diagram of wavelet decomposition according to an embodiment； and

FIG. 3 is a structural diagram of a bearing failure diagnosis device according to an embodiment.

DETAILED DESCRIPTIONOF EMBODIMENTS

In the existing techniques, to diagnose bearing failures, data analyzers generally make subjective judgement by observing an envelope spectrum waveform of a demodulated time domain vibration acceleration signal, which requires the data analyzers mastering related theory to perform analysis on data and thus is relatively subjective. There are also some bearing failure diagnosis methods based on neural network theory. However, the methods always result in low accuracy of bearing failure diagnosis.

In embodiments of the present disclosure, wavelet decomposition is performed to the acquired vibration acceleration signal of the bearing to obtain a plurality of signals, feature parameters are extracted from the plurality of signals respectively and feature vectors are constructed, a feature matrix is constructed based on the feature vectors and fed to a trained neural network model which then processes the feature matrix to obtain an output result, and the output result is compared with a predetermined target matrix to obtain a failure diagnosis result of the bearing. The determination of the bearing failure does not depend on subjective judgement of a testing person, which may improve accuracy of bearing failure diagnosis.

In order to clarify the object, solutions and advantages of embodiments of the present disclosure, embodiments of present disclosure will be described clearly in detail in conjunction with accompanying drawings.

In an embodiment, a bearing failure diagnosis method is provided. Referring to Figure 1, the method is described in detail below with the detailed steps.

In S101, a vibration acceleration signal of a bearing which is running is acquired.

In some embodiments, when the bearing is running, the vibration acceleration signal of the bearing is acquired from vibration signals of a mechanical system where the bearing is disposed.

In practice, the bearing is generally a component of a mechanical system. When the mechanical system is running, a preset acceleration sensor may be employed to acquire the vibration signals of the mechanical system in real time. The acceleration sensor may be disposed inside the mechanical system or independent from the mechanical system.

The vibration signals of the mechanical system acquired by the acceleration sensor include the vibration acceleration signal of the bearing, vibration acceleration signals of other components in the mechanical system, and some background noises. Therefore, in some embodiments, after the vibration signals of the mechanical system are acquired by the acceleration sensor, the vibration acceleration signal of the bearing is extracted.

In some embodiments, the vibration acceleration signal of the bearing may be directly acquired by a sensor coupled with the bearing. For example, a sensor is preset to be coupled with the bearing. When the bearing is running, the sensor is employed to acquire the vibration acceleration signal of the bearing in real time.

No matter the vibration acceleration signal of the bearing is extracted from the vibration signals of the mechanical system or directly acquired by the sensor coupled with the bearing, the obtained vibration acceleration signal of the bearing may include other noises. It could be understood that, the acquired vibration acceleration signal of the bearing in embodiments of the present disclosure is a signal acquired when the bearing is running.

Therefore, after the acquisition, the vibration acceleration signal of the bearing may be denoised, to reduce noises in the vibration acceleration signal of the bearing. In some embodiments, the vibration acceleration signal of the bearing may be denoised by using a wavelet shrinkage denoising method. In some embodiments, other denoising methods may be employed to denoise the vibration acceleration signal of the bearing, and are not described in detail here.

In S102, wavelet decomposition is performed to the vibration acceleration signal of the bearing.

In some embodiments, the vibration acceleration signal of the bearing may be a time domain vibration acceleration signal of the bearing or a frequency domain vibration acceleration signal of the bearing. If the vibration acceleration signal of the bearing is a time domain vibration acceleration signal of the bearing, wavelet decomposition is performed to the time domain vibration acceleration signal of the bearing. If the vibration acceleration signal of the bearing is a frequency domain vibration acceleration signal of the bearing, wavelet decomposition is performed to the frequency domain vibration acceleration signal of the bearing.

If wavelet decomposition is performed to the frequency domain vibration acceleration signal of the bearing, a mother wavelet function which has a waveform relevant to a waveform of the frequency domain vibration acceleration signal of the bearing may be selected, and wavelet decomposition is performed to the selected mother wavelet function. If wavelet decomposition is performed to the time domain vibration acceleration signal of the bearing, a mother wavelet function which has a waveform relevant to a waveform of the time domain vibration acceleration signal of the bearing may be selected, and wavelet decomposition is performed to the selected mother wavelet function.

In the below embodiment, wavelet decomposition performed to the time domain vibration acceleration signal of the bearing is described in detail.

In the embodiment, a mother wavelet function which has a waveform relevant to a waveform of the time domain vibration acceleration signal of the bearing is selected, and wavelet decomposition is performed to the selected mother wavelet function, to obtain time domain signals in N frequency bands, where N＝2^j, and j is the number of layers of the wavelet decomposition.

In some embodiments, the layers of wavelet decomposition may be determined based on practical requirements. In some embodiments, j＝3. Accordingly, after wavelet decomposition is performed to the selected mother wavelet function, time domain signals in eight frequency bands with a same band width are obtained.

FIG. 2 is a diagram of wavelet decomposition according to an embodiment. Referring to Figure 2, the number j of layers of wavelet decomposition is three.

Still referring to Figure 2, first, a signal S is decomposed to obtain signals S_A1 and S_D1. Second, the signal S_A1 is decomposed to obtain signals S_AA2 and S_DA2, and the signal S_D1 is decomposed to obtain signals S_AD2 and S_DD2. Third, the signal S_AA2 is decomposed to obtain signals S₃₀ and S₃₁, the signal S_DA2 is decomposed to obtain signals S₃₂ and S₃₃, the signal S_AD2 is decomposed to obtain signals S₃₄ and S₃₅, and the signal S_DD2 is decomposed to obtain signals S₃₆ and S₃₇.

From above, the signal S is subjected to three-layer wavelet decomposition and reconstruction to obtain third-layer signals which include time domain signals in eight frequency bands with the same band width, i.e., S₃₀, S₃₁, S₃₂, S₃₃, S₃₄, S₃₅, S₃₆ and S₃₇.

In some embodiments, for the selection of the mother wavelet function, a mother wavelet function which has a waveform relevant to a waveform of the vibration acceleration signal of the bearing may be selected. Relevant waveform means that the waveform of the mother wavelet function is similar with the waveform of the vibration acceleration signal of the bearing, and the similarity reaches a predetermined threshold.

For example, a mother wavelet function which has a waveform similar with a waveform of the time domain vibration acceleration signal of the bearing may be selected according to practical experiences. As the waveform of the selected mother wavelet function is relevant to the waveform of the time domain vibration acceleration signal of the bearing, after wavelet decomposition is performed to the selected mother wavelet function, signals relevant to the time domain vibration acceleration signal of the bearing, among the time domain signals in the N frequency bands with the same band width, may be outlined, while signals irrelevant to the time domain vibration acceleration signal of the bearing may be weakened.

In some embodiments, the time domain signals in the N frequency bands are demodulated. If the time domain vibration acceleration signal of the bearing includes an impact frequency component caused by a bearing failure, the impact frequency component may be extracted from a high-frequency carrier signal after the demodulation of the time domain signals in the N frequency bands, as characteristics of the time domain signals in each frequency band are more distinct after the demodulation.

It is known in practice that, after time-frequency transformation is performed to the time domain signals to obtain frequency domain signals, characteristics of waveforms in the frequency domain may be more distinct than that of waveforms in the time domain. Therefore, after the demodulation of the time domain signals in the N frequency bands, time-frequency transformation is performed to the demodulated time domain signals in the N frequency bands, to transform the time domain signals in the N frequency bands into corresponding frequency domain signals in the N frequency bands.

In some embodiments, the time-frequency transformation may employ a Fast Fourier Transformation (FFT) method, to transform the time domain signals in the N frequency bands into the corresponding frequency domain signals in the N frequency bands.

In S103, feature parameters are extracted from the plurality of signals respectively and feature vectors are constructed.

In some embodiments, the extracted feature parameters may include at least one selected from: energy of each of the plurality of signals, a crest factor of each of the plurality of signals, standard deviation of each of the plurality of signals, a root mean square of each of the plurality of signals, and 90 percentile of each of the plurality of signals.

In some embodiments, after wavelet decomposition is performed to the mother wavelet function to obtain frequency domain signals in N frequency bands, same feature parameters may be extracted from the frequency domain signals in each of the N frequency bands to obtain the feature vectors, where the number of the feature vectors is the same as the number of the feature parameters extracted from the frequency domain signals in each frequency band, and the number of elements in each feature vector is N.

For example, five feature parameters are extracted from the frequency domain signals in each frequency band, and N is 8. Accordingly, the number of the feature vectors is 5, and the number of elements in each feature vector is 8.

In some embodiments, the feature parameters extracted from the frequency domain signals in the N frequency bands may include at least one selected from: energy of the frequency domain signals in each frequency band, a crest factor of the frequency domain signals in each frequency band, standard deviation of the frequency domain signals in each frequency band, a root mean square of the frequency domain signals in each frequency band, and 90 percentile of the frequency domain signals in each frequency band.

In S104, a feature matrix is constructed based on the feature vectors.

In some embodiments, the feature matrix may be constructed after the feature vectors are obtained in S103.

In some embodiments, after wavelet decomposition is performed to the mother wavelet function to obtain frequency domain signals in the N frequency bands, same feature parameters may be extracted from the frequency domain signals in each of the N frequency bands to obtain the feature vectors. Accordingly, a size of the constructed feature matrix is M by N, and M is the number of the feature vectors.

For example, three-layer wavelet decomposition is performed to the time domain vibration acceleration signal of the bearing, where j＝3 and N＝8. For the frequency domain signals in each frequency band, five feature parameters are extracted, and thus five feature vectors are constructed. In this way, a size of the constructed feature matrix is 5 by 8, that is, the constructed feature matrix includes five rows and eight columns.

S103 and S104 are described in detail below by taking N＝8 as an example.

It is known in practice that, according to the wavelet packet reconstruction theory, the signal S can be perfectly reconstructed by the decomposed signals. Based on the law of conservation of energy, energy in time domain is equal to energy in frequency domain. Therefore, the following Equation can be obtained,

FS＝FS₃₀+ FS₃₁+ FS₃₂+ FS₃₃+ FS₃₄+ FS₃₅+ FS₃₆+ FS₃₇ (1) ,

where FS is the frequency domain vibration acceleration signal corresponding to the time domain vibration acceleration signal, and FS₃₀ to FS₃₇ are the frequency domain signals corresponding toS₃₀～S₃₇.

Assume that energy of the frequency domain signals in the i^th frequency band is E_3i,

where i＝0, 1, 2, …, 7, k＝1, 2, …, n, and x_ik is a wavelet packet coefficient of the third-layer signal S_3i after wavelet decomposition. Therefore, the total energy of the frequency domain signals in the N frequency bands is

Relative wavelet packet coefficient energy of the frequency domain signals in each frequency band in the third-layer is

After the energy of the frequency domain signals in each frequency band is obtained, the feature vector based on the energy is constructed as R＝ [R₃₀, R₃₁, …, R₃₇] .

It is known in practice that, the crest factor of the frequency domain signal reflects a mechanical vibration level of the frequency domain signal.

In some embodiments, the crest factor of the frequency domain signals in the i^th frequency band is

where peak_3i is a crest value of the frequency domain signals in the i^th frequency band, and RMS_3i is a root mean square of the frequency domain signals in the i^th frequency band.

After the crest factor of the frequency domain signals in each frequency band is obtained, the feature vector based on the crest factor is constructed as C＝ [C₃₀, C₃₁, …, C₃₇] .

It is known in practice that, the standard deviation reflects a dispersion degree of data in an array. When a failure occurs in the bearing, an impact phenomenon will happen. A failure signal in each frequency band has a remarkable feature after demodulation, thus the standard deviation can indicate a bearing failure.

In some embodiments, the standard deviation of the frequency domain signals in the i^th frequency band is

where L is the number of values of the frequency domain signals in the i^th frequency band, x_l is the l^th value of the frequency domain signals in the i^th frequency band, r is an average value of values of the frequency domain signals in the i^th frequency band, and

After the standard deviation of the frequency domain signals in each frequency band is obtained, the feature vector based on the standard deviation is constructed as σ＝ [σ₃₀, σ₃₁, …, σ₃₇] .

It is known in practice that, the root mean square reflects a total energy level of a signal, thus it can serve as one feature parameter in bearing failure diagnosis.

In some embodiments, the root mean square of the frequency domain signals in the i^th frequency band is

where L is the number of values of the frequency domain signals in the i^th frequency band, and x_l is the l^th value of the frequency domain signals in the i^th frequency band.

After the root mean square of the frequency domain signals in each frequency band is obtained, the feature vector based on the root mean square is constructed as RMS＝ [RMS₃₀, RMS₃₁, …, RMS₃₇] .

In some embodiments, for the frequency domain signals in the i^th frequency band, 90 percentile P_3i of the maximum value in the i^th frequency band is selected. After the 90 percentile of the frequency domain signals in each frequency band is obtained, the feature vector based on the 90 percentile is constructed as P＝ [P₃₀, P₃₁, …, P₃₇] .

It could be understood that, the feature parameters extracted from the frequency domain signals in the N frequency bands are not limited to the types of parameters provided in the above embodiments. Other feature parameters that can reflect characteristics of the frequency signals may be extracted, and are not described in detail here.

In some embodiments, according to practical requirements, one or more of the feature vectors generated in S103 is selected to construct the feature matrix. For example, N＝8, and the feature vectors based on the energy and the feature vectors based on the crest factor are selected from the feature vectors generated in S103 to construct the feature matrix. Accordingly, a size of the feature matrix is 2 by 8. For another example, N＝8, and all the feature vectors generated in S103 are selected to construct the feature matrix. Accordingly, a size of the feature matrix is 5 by 8.

In S105, the feature matrix is fed to a trained neural network model.

In some embodiments, a neural network may be pre-generated and trained. In practice, a plurality of groups of data of known failures to be detected or diagnosed may be collected and a generated neural network model is trained. During the training of the generated neural network model, weighting values of all layers may be continuously adjusted based on an input known feature matrix, until an error of the neural network model reaches a predetermined minimum error.

In some embodiments, to accelerate a converge speed of the neural network model, normalization is performed to the feature matrix before the feature matrix is fed to the neural network model. During the normalization, an element with the maximum value is selected from the feature matrix, and other elements in the feature matrix are divided by the selected element respectively to realize the normalization of the feature matrix.

In S106, an output result of the trained neural network model is compared with a predetermined target matrix to obtain a failure diagnosis result of the bearing.

S105 and S106 are described in detail below.

In some embodiments, the target matrix is predetermined. A size of the target matrix is M by K, and K is the number of parts to be diagnosed in the bearing.

It is known in practice that, a bearing generally consists of Outer Ring (OR) , Inner Ring (IR) , roller, and cage. In practice, a training procedure of a neural network model may include data preparation and training of the neural network model. In the data preparation, the acquired vibration acceleration signal, a failure signal feature matrix corresponding to the OR of the bearing, a failure signal feature matrix corresponding to the IR of the bearing, a failure signal feature matrix corresponding to the roller of the bearing, and a failure signal feature matrix corresponding to the cage of the bearing serve as the input of the neural network model, and the neural network model is trained by taking a predetermined target matrix as an output. In the training of the neural network model, weighting values and thresholds are continuously adjusted according to a predetermined minimum error, until the predetermined minimum error is achieved.

In some embodiments, data needed in the training of the neural network model and the determination of the minimum error may take a failure diagnosis result as reference, such as whether a failure occurs, an object of the failure diagnosis (i.e., a part of the bearing which needs to be diagnosed) , and a severity level of a failure in different working conditions.

In some embodiments, k is 4, and the target matrix may be set as

In the target matrix, a first row is a target vector representing a failure occurs in the OR of the bearing, a second row is a target vector representing a failure occurs in the IR of the bearing, a third row is a target vector representing a failure occurs in the roller of the bearing, and a fourth row is a target vector representing a failure occurs in the cage of the bearing.

As the feature matrix serves as the input of the neural network model, and the size of the feature matrix is M by N, the target matrix may be reconstructed. Each column in the target matrix may be expanded to be an independent M×K matrix, so as to obtain K target matrices which correspond to K parts of the bearing respectively. For each M×K matrix, elements in different rows are the same.

In some embodiments, K is 4 and M is 5. The four columns in the target matrix are expanded to be four independent 5×4 target matrices which are the target matrices corresponding to the OR of the bearing, the target matrix corresponding to the IR of the bearing, the target matrix corresponding to the roller of the bearing, and the target matrix corresponding to the cage of the bearing, where the target matrix corresponding to the bearing in a normal status is

the target matrix corresponding to the OR of the bearing is

the target matrix corresponding to the IR of the bearing is

the target matrix corresponding to the roller of the bearing is

and the target matrix corresponding to the cage of the bearing is

In the above embodiments, the target matrix corresponding to the OR of the bearing, the target matrix corresponding to the IR of the bearing, the target matrix corresponding to the roller of the bearing, and the target matrix corresponding to the cage of the bearing are generated by expanding the predetermined target matrix.

In some embodiments, the trained neural network model may include an input layer, a hidden layer and an output layer. Data of the input layer is the feature matrix, and data of the output layer is the predetermined target matrix. To facilitate obtaining a diagnosis result, the data of the output layer may include the target matrix corresponding to the OR of the bearing, the target matrix corresponding to the IR of the bearing, the target matrix corresponding to the roller of the bearing, and the target matrix corresponding to the cage of the bearing that are generated by expanding the predetermined target matrix.

In some embodiments, the output result of the trained neural network model may be compared with the predetermined target matrix to obtain error values between the output result and the target matrix； and it is determined that a failure occurs in a part of the bearing which corresponds to the minimum error value among the obtained error values.

In some embodiments, comparing the output result of the trained neural network model with the predetermined target matrix may include: comparing the output result of the trained neural network model with the plurality of target matrices generated by expansion, that is, comparing the output result of the trained neural network model with the target matrix corresponding to the OR of the bearing, the target matrix corresponding to the IR of the bearing, the target matrix corresponding to the roller of the bearing, and the target matrix corresponding to the cage of the bearing respectively, to obtain the minimum error value with the target matrices generated by expansion, and to further determine that a failure occurs in which part of the bearing.

For example, the output result of the trained neural network model is a 5×4 matrix, and elements in the first column are within a range from 0.9 to 1. It can be determined that an error between the output result of the trained neural network model and the target matrix corresponding to the OR of the bearing is minimum, which further leads to a conclusion that a failure occurs in the OR of the bearing.

From above, wavelet decomposition is performed to the acquired time domain vibration acceleration signal of the bearing, where both high-frequency part and low-frequency part are considered. The feature parameters are extracted from the frequency domain signals in the N frequency bands respectively and the feature vectors are constructed, the feature matrix is constructed based on the feature vectors and fed to the neural network model, and the failure diagnosis result of the bearing is obtained based on the output result of the neural network model. The determination of the bearing failure does not depend on subjective judgement of a testing person, which may improve accuracy of bearing failure diagnosis.

FIG. 3 is a structural diagram of a bearing failure diagnosis device according to an embodiment. Referring to Figure 3, the bearing failure diagnosis device 30 includes: an acquiring circuitry 301 configured to acquire a vibration acceleration signal of a bearing which is running； a wavelet decomposing circuitry 302 configured to perform wavelet decomposition to the vibration acceleration signal of the bearing to obtain a plurality of signals； a feature vector constructing circuitry 303 configured to extract feature parameters from the plurality of signals respectively and construct feature vectors； a feature matrix constructing circuitry 304 configured to construct a feature matrix based on the feature vectors； an inputting circuitry 305 configured to feed the feature matrix to a trained neural network model； and a failure diagnosis result obtaining circuitry 306 configured to compare an output result of the trained neural network model with a predetermined target matrix to obtain a failure diagnosis result of the bearing.

In some embodiments, the acquiring circuitry 301 may be configured to acquire a time domain vibration acceleration signal of the bearing.

In some embodiments, except the acquiring circuitry 301, other circuitries in the bearing failure diagnosis device 30 may be set to be a portion of an on-line system or an off-line system according to practical requirements.

In some embodiments, the wavelet decomposing circuitry 302 may be configured to: select a mother wavelet function, wherein a waveform of the mother wavelet function is relevant to a waveform of the time domain vibration acceleration signal of the bearing； and perform wavelet decomposition to the mother wavelet function to obtain time domain signals in N frequency bands, where N＝2^j, and j is the number of layers of the wavelet decomposition.

In some embodiments, the feature vector constructing circuitry 303 may be configured to: demodulate the time domain signals in the N frequency bands and perform time-frequency transformation to obtain corresponding frequency domain signals in the N frequency bands； and extract feature parameters from the corresponding frequency domain signals in the N frequency bands respectively, and construct feature vectors.

In some embodiments, the feature matrix constructing circuitry 304 may be configured to: construct a feature matrix based on the feature vectors, where a size of the feature matrix is M by N, and M is the number of the feature vectors.

In some embodiments, the bearing failure diagnosis device 30 may further include a denoising circuitry (not shown in Figure 3) configured to: denoise the vibration acceleration signal of the bearing before the wavelet decomposing circuitry performs wavelet decomposition to the vibration acceleration signal of the bearing.

In some embodiments, the bearing failure diagnosis device 30 may further include a normalizing circuitry (not shown in Figure 3) configured to: normalize elements in the feature matrix before the inputting circuitry feeds the feature matrix to the trained neural network model.

In some embodiments, the failure diagnosis result obtaining circuitry 306 may be configured to: compare the output result of the trained neural network model with the predetermined target matrix to obtain error values between the output result and the target matrix； and determine that a failure occurs in a part of the bearing which corresponds to the minimum error value among the obtained error values.

In an embodiment, a computer readable storage medium which has computer instructions stored therein is provided, wherein once the computer instructions are executed, the bearing failure diagnosis method provided in the above embodiments is performed.

In an embodiment, an electronic device including a memory and a processor is provided, wherein the memory has computer instructions stored therein, and once executing the computer instructions, the processor performs the bearing failure diagnosis method provided in the above embodiments.

Those skilled in the art can understand that all of or a portion of the processes in the method provided in the above embodiments can be implemented by related hardware with instruction of computer program. The computer program may be stored in a readable storage medium, such as a Read-Only Memory (ROM) , a Random Access Memory (RAM) , a magnetic disk or an optical disk.

Although the present disclosure has been disclosed above with reference to preferred embodiments thereof, it should be understood that the disclosure is presented by way of example only, and not limitation. Those skilled in the art can modify and vary the embodiments without departing from the spirit and scope of the present disclosure.

Claims

A bearing failure diagnosis method, comprising:

acquiring a vibration acceleration signal of a bearing which is running；

performing wavelet decomposition to the vibration acceleration signal of the bearing to obtain a plurality of signals；

extracting feature parameters from the plurality of signals respectively and constructing feature vectors；

constructing a feature matrix based on the feature vectors；

feeding the feature matrix to a trained neural network model； and

comparing an output result of the trained neural network model with a predetermined target matrix to obtain a failure diagnosis result of the bearing.
The method according to claim 1, wherein the failure diagnosis result of the bearing comprises at least one of the followings: whether a failure occurs in the bearing, a part of the bearing where a failure occurs, and a severity level of a failure.
The method according to claim 1 or 2, wherein acquiring a vibration acceleration signal of a bearing comprises:

acquiring a time domain vibration acceleration signal of the bearing.
The method according to any one of claims 1 to 3, wherein performing wavelet decomposition to the vibration acceleration signal of the bearing to obtain a plurality of signals comprises:

selecting a mother wavelet function, wherein a waveform of the mother wavelet function is relevant to a waveform of the time domain vibration acceleration signal of the bearing； and

performing wavelet decomposition to the mother wavelet function to obtain time domain signals in N frequency bands, where N＝2^j, and j is the number of layers of the wavelet decomposition.
The method according to claim 4, wherein extracting feature parameters from the plurality of signals respectively and constructing feature vectors comprises:

demodulating the time domain signals in the N frequency bands and performing time-frequency transformation to obtain corresponding frequency domain signals in the N frequency bands； and

extracting feature parameters from the corresponding frequency domain signals in the N frequency bands respectively, and constructing feature vectors.
The method according to any one of claims 1 to 5, wherein constructing a feature matrix based on the feature vectors comprises:

constructing a feature matrix based on the feature vectors, where a size of the feature matrix is M by N, and M is the number of the feature vectors.
The method according to any one of claims 1 to 6, wherein a size of the target matrix is M by K, and K is the number of parts to be diagnosed in the bearing.
The method according to any one of claims 1 to 7, wherein prior to performing wavelet decomposition to the vibration acceleration signal of the bearing, the method further comprises:

denoising the vibration acceleration signal of the bearing.
The method according to claim 8, wherein denoising the vibration acceleration signal of the bearing comprises:

denoising the vibration acceleration signal of the bearing by using a wavelet shrinkage denoising method.
The method according to any one of claims 1 to 9, wherein the feature parameters comprise at least one selected from:

energy of each of the plurality of signals, a crest factor of each of the plurality of signals, standard deviation of each of the plurality of signals, a root mean square of each of the plurality of signals, and 90 percentile of each of the plurality of signals.
The method according to any one of claims 1 to 10, wherein prior to feeding the feature matrix to a trained neural network model, the method further comprises:

normalizing elements in the feature matrix.
The method according to any one of claims 1 to 11, wherein the trained neural network model comprises an input layer, a hidden layer and an output layer, data of the input layer is the feature matrix, and data of the output layer is the predetermined target matrix.
The method according to any one of claims 1 to 12, wherein comparing an output result of the trained neural network model with a predetermined target matrix to obtain a failure diagnosis result of the bearing comprises:

comparing the output result of the trained neural network model with the predetermined target matrix to obtain error values between the output result and the target matrix； and

determining that a failure occurs in a part of the bearing which corresponds to the minimum error value among the obtained error values.
A bearing failure diagnosis device, comprising:

an acquiring circuitry configured to acquire a vibration acceleration signal of a bearing which is running；

a wavelet decomposing circuitry configured to perform wavelet decomposition to the vibration acceleration signal of the bearing to obtain a plurality of signals；

a feature vector constructing circuitry configured to extract feature parameters from the plurality of signals respectively and construct feature vectors；

a feature matrix constructing circuitry configured to construct a feature matrix based on the feature vectors；

an inputting circuitry configured to feed the feature matrix to a trained neural network model； and

a failure diagnosis result obtaining circuitry configured to compare an output result of the trained neural network model with a predetermined target matrix to obtain a failure diagnosis result of the bearing.
The device according to claim 14, wherein the failure diagnosis result of the bearing comprises at least one of the followings: whether a failure occurs in the bearing, a part of the bearing where a failure occurs, and a severity level of a failure.
The device according to claim 14 or 15, wherein the acquiring circuitry is configured to acquire a time domain vibration acceleration signal of the bearing.
The device according to any one of claims 14 to 16, wherein the wavelet decomposing circuitry is configured to:

select a mother wavelet function, wherein a waveform of the mother wavelet function is relevant to a waveform of the time domain vibration acceleration signal of the bearing； and

perform wavelet decomposition to the mother wavelet function to obtain time domain signals in N frequency bands, where N＝2^j, and j is the number of layers of the wavelet decomposition.
The device according to claim 17, wherein the feature vector constructing circuitry is configured to:

demodulate the time domain signals in the N frequency bands and perform time-frequency transformation to obtain corresponding frequency domain signals in the N frequency bands； and

extract feature parameters from the corresponding frequency domain signals in the N frequency bands respectively, and construct feature vectors.
The device according to any one of claims 14 to 18, wherein the feature matrix constructing circuitry is configured to:

construct a feature matrix based on the feature vectors, where a size of the feature matrix is M by N, and M is the number of the feature vectors.
The device according to any one of claims 14 to 19, wherein a size of the target matrix is M by K, and K is the number of parts to be diagnosed in the bearing.
The device according to any one of claims 14 to 20, further comprising a denoising circuitry configured to:

denoise the vibration acceleration signal of the bearing before the wavelet decomposing circuitry performs wavelet decomposition to the vibration acceleration signal of the bearing.
The device according to claim 21, wherein the denoising circuitry is configured to:

denoise the vibration acceleration signal of the bearing by using a wavelet shrinkage denoising method.
The device according to any one of claims 14 to 22, wherein the feature parameters comprise at least one selected from:

energy of each of the plurality of signals, a crest factor of each of the plurality of signals, standard deviation of each of the plurality of signals, a root mean square of each of the plurality of signals, and 90 percentile of each of the plurality of signals.
The device according to any one of claims 14 to 23, further comprising a normalizing circuitry configured to:

normalize elements in the feature matrix before the inputting circuitry feeds the feature matrix to the trained neural network model.
The method according to any one of claims 14 to 24, wherein the trained neural network model comprises an input layer, a hidden layer and an output layer, data of the input layer is the feature matrix, and data of the output layer is the predetermined target matrix.
The method according to any one of claims 14 to 25, wherein the failure diagnosis result obtaining circuitry is configured to:

compare the output result of the trained neural network model with the predetermined target matrix to obtain error values between the output result and the target matrix； and

determine that a failure occurs in a part of the bearing which corresponds to the minimum error value among the obtained error values.
A computer readable storage medium which has computer instructions stored therein, wherein once the computer instructions are executed, the method according to any one of claims 1 to 13 is performed.
An electronic device comprising a memory and a processor, wherein the memory has computer instructions stored therein, and once executing the computer instructions, the processor performs the method according to any one of claims 1 to 13.