CN116226646A

CN116226646A - Method, system, equipment and medium for predicting health state and residual life of bearing

Info

Publication number: CN116226646A
Application number: CN202310492059.6A
Authority: CN
Inventors: 陈朋超; 李华
Original assignee: National Petroleum And Natural Gas Pipeline Network Group Co ltd Science And Technology Research Institute Branch; China Oil and Gas Pipeline Network Corp
Current assignee: National Petroleum And Natural Gas Pipeline Network Group Co ltd Science And Technology Research Institute Branch; China Oil and Gas Pipeline Network Corp
Priority date: 2023-05-05
Filing date: 2023-05-05
Publication date: 2023-06-06
Anticipated expiration: 2043-05-05
Also published as: CN116226646B

Abstract

The invention discloses a method, a system, equipment and a medium for predicting the health state and the residual life of a bearing, and relates to the technical field of bearing health prediction, wherein the method comprises the following steps: acquiring a target vibration signal of a faulty target bearing after the faulty target bearing runs within a set time period; acquiring a target health degree trend feature set corresponding to the target bearing according to the target vibration signal, wherein the target health degree trend feature set comprises a plurality of preset health degree trend features and feature data corresponding to the health degree trend features; and predicting the health state and/or the residual life of the target bearing according to the target health trend characteristic set. The invention can accurately predict the health state and the residual life of the bearing, is not limited by the type of the vibration sensor, is simple and easy to operate, and improves the accuracy of predicting the health state and the residual life of the bearing.

Description

Method, system, equipment and medium for predicting health state and residual life of bearing

Technical Field

The invention relates to the technical field of bearing health prediction, in particular to a method, a system, equipment and a medium for predicting the health state and the residual life of a bearing.

Background

The bearing is used as a key component of rotary machinery, and is widely applied to various industrial fields such as electric power, petrochemical industry, aerospace and the like. Bearings are one of the most vulnerable components of rotary machines, and statistics show that in the existing failure cases of rotary machines, about 45 to 55% are caused by damage or failure of the bearings, and the health of the bearings determines the health of the mechanical equipment to some extent. Because the working environment of the bearing is complex and changeable, the residual service life condition of the bearing cannot be summarized, and the conditions of insufficient maintenance and excessive maintenance are easily caused by regular maintenance, so that the method has important significance in realizing the health state prediction of the bearing.

In the aspect of bearing health state prediction, a data driving method based on machine learning can overcome the problem of unknown model, and is a hot spot of current research. The traditional shallow machine learning method relies on the prior knowledge and signal processing technology of experts to a great extent, and along with the improvement of computer computing power, the strong learning ability of the method provides a new choice for the prediction of the health state of the bearing. The recurrent neural network (Recurrent Neural Network, RNN for short) is widely used in sequences with interdependence characteristics, such as the field of health status prediction, because of its special network structure, which can retain the status information implied at the previous time. The later-occurring Long Short-Term Memory (LSTM) is improved against the problems of gradient disappearance and gradient explosion which are easy to occur in RNN, so that research hotspots are formed in the field of health state prediction. The usual health state prediction procedure based on LSTM is roughly: and carrying out common time-frequency domain feature extraction based on vibration data, then fusing the features, taking the fused index as a health index of the bearing, and predicting by using a model built by LSTM, thereby realizing the prediction of the health state of the bearing. The process has the defects that the information in the vibration data can not be fully extracted only by extracting the time-frequency domain characteristics of the vibration data of the bearing, more characteristic information can be lost when the characteristics are fused, and finally the prediction accuracy of the health state of the bearing is low.

Disclosure of Invention

The technical problems to be solved by the invention are as follows: and the health state and the residual life of the bearing are predicted by only using a small amount of characteristic information in the vibration data of the bearing, so that the prediction accuracy is low. In order to solve the technical problem, the invention provides a method, a system, equipment and a medium for predicting the health state and the residual life of a bearing.

The technical scheme for solving the technical problems is as follows:

a method for predicting the health state and the residual life of a bearing comprises the following steps:

step S1, obtaining a target vibration signal of a target bearing with a fault state after running in a set time length;

step S2, acquiring a target health degree trend feature set corresponding to the target bearing according to the target vibration signal, wherein the target health degree trend feature set comprises a plurality of preset health degree trend features and feature data corresponding to the health degree trend features;

and step S3, predicting the health state and/or the residual life of the target bearing according to the target health degree trend characteristic set.

The beneficial effects of the invention are as follows: for the target bearing with the fault state being faulty, the method determines the target health degree trend feature set corresponding to the target bearing according to the target vibration signal of the target bearing, and finally, according to the feature data and the trained model corresponding to each preset health degree trend feature (the health degree trend feature is a feature index for judging the health state of the bearing) in the target health degree trend feature set, the health state and/or the residual life of the target bearing can be accurately predicted.

On the basis of the technical scheme, the invention can be improved as follows.

Further, the target vibration signal includes a plurality of signals to be decomposed;

the fault state of the target bearing is determined by:

for each signal to be decomposed, carrying out singular spectrum analysis on the signal to be decomposed to obtain a reconstructed vibration signal corresponding to the signal to be decomposed;

reconstructing the spectrum of the reconstructed vibration signal corresponding to the signal to be decomposed through a trained convolution self-encoder, and calculating the average value of the characteristic data output by the convolution self-encoder to obtain an average value sequence corresponding to the signal to be decomposed;

for each signal to be decomposed, determining a state value corresponding to the signal to be decomposed according to a mean value sequence corresponding to the signal to be decomposed and a preset index threshold value through a first formula, wherein the index threshold value comprises a mean value threshold value and a standard deviation threshold value, and the first formula is as follows:

；

wherein ,

representing the state value corresponding to the signal to be decomposed, < >>

Absolute value representing differential value of the mean value sequence corresponding to the signal to be decomposedValue pair (s)/(s) >

Representing the mean threshold,/->

Representing the standard deviation threshold;

if the state value which is larger than or equal to 0 exists in the state value corresponding to each signal to be decomposed, the fault state of the target bearing is faulty;

if the state value which is larger than or equal to 0 does not exist in the state value which corresponds to each signal to be decomposed, the fault state of the target bearing is fault-free.

The beneficial effects of adopting the further scheme are as follows: the average value threshold value and the standard deviation threshold value are calculated according to the running data of the bearing with the fault state being in a fault-free state, and the judgment of the fault state of the bearing based on the average value threshold value and the standard deviation threshold value is more adaptive than the judgment of the fault state of the bearing according to the threshold value which is simply set.

Further, the preset health trend features are obtained through the following steps:

acquiring a vibration signal of a sampling bearing after running under a preset first running condition, wherein the fault state of the sampling bearing is fault-free;

according to the vibration signal, determining a time domain statistical feature set, a frequency domain statistical feature set and a node energy feature set which correspond to the sampling bearing, wherein the time domain statistical feature set comprises feature data which correspond to each of a plurality of time domain statistical features, the frequency domain statistical feature set comprises feature data which correspond to each of a plurality of frequency domain statistical features, and the node energy feature set comprises feature data which correspond to each of a plurality of node energy features;

Reconstructing a time domain waveform of the vibration signal to obtain a time domain reconstruction feature set, wherein the time domain reconstruction feature set comprises feature data corresponding to each of a plurality of time domain reconstruction features;

reconstructing the frequency spectrum of the vibration signal to obtain a frequency spectrum reconstruction feature set, wherein the frequency spectrum reconstruction feature set comprises feature data corresponding to a plurality of frequency spectrum reconstruction features;

constructing a feature set to be verified according to the time domain statistical feature set, the frequency domain statistical feature set, the node energy feature set, the time domain reconstruction feature set and the frequency spectrum reconstruction feature set, wherein the feature set to be verified comprises a plurality of features to be verified, and each feature to be verified corresponds to one time domain statistical feature, one frequency domain statistical feature, one node energy feature, one time domain reconstruction feature or one frequency spectrum reconstruction feature;

for each feature to be verified, determining a feature evaluation index value corresponding to the feature to be verified, wherein the feature evaluation index value characterizes the influence degree between the corresponding feature to be verified and the degradation of the sampling bearing, and determining a target feature set according to the feature evaluation index value, wherein the target feature set comprises a plurality of target features, and each target feature corresponds to one feature to be verified;

And obtaining a plurality of health degree trend features corresponding to the sampling bearing according to the target feature set, and taking the plurality of health degree trend features corresponding to the sampling bearing as the preset plurality of health degree trend features.

The beneficial effects of adopting the further scheme are as follows: according to the vibration signals, a large amount of characteristic information (namely time domain statistical characteristics, frequency domain statistical characteristics, node energy characteristics, time domain reconstruction characteristics and frequency spectrum reconstruction characteristics) in the vibration data of the bearing is extracted, whether the health trend characteristics of the bearing are determined by utilizing the characteristics or not is determined according to the influence degree of each characteristic on the degradation of the bearing, the health state and the residual life of the bearing are determined based on the finally determined health trend characteristics, and the prediction accuracy of the health state and the residual life of the bearing can be improved.

Further, the determining, according to the vibration signal, the time domain statistical feature set, the frequency domain statistical feature set and the node energy feature set corresponding to the sampling bearing includes:

calculating time domain statistical characteristics of the vibration signals according to the vibration signals to obtain a time domain statistical characteristic set;

performing Fourier transform on the vibration signal to obtain a frequency spectrum of the vibration signal, and calculating frequency domain statistical characteristics of the vibration signal according to the frequency spectrum to obtain a frequency domain statistical characteristic set;

Carrying out signal decomposition on the vibration signal by a wavelet packet transformation method to obtain a node energy feature set;

reconstructing the time domain waveform of the vibration signal to obtain a time domain reconstruction feature set, including:

normalizing the time domain waveform of the vibration signal to obtain a target time domain waveform, and reconstructing the target time domain waveform through a trained convolution self-encoder to obtain a time domain reconstruction feature set, wherein time domain reconstruction features in the time domain reconstruction feature set are in one-to-one correspondence with time domain statistics features in the time domain statistics feature set;

the reconstructing the spectrum of the vibration signal to obtain a spectrum reconstruction feature set includes:

and carrying out normalization processing on the frequency spectrum of the vibration signal to obtain a target frequency spectrum, and reconstructing the target frequency spectrum through the convolution self-encoder to obtain a frequency domain reconstruction feature set, wherein the frequency domain reconstruction features in the frequency domain reconstruction feature set are in one-to-one correspondence with the time domain statistics features in the time domain statistics feature set.

The beneficial effects of adopting the further scheme are as follows: when the convolution self-encoder reconstructs the time waveform and the frequency spectrum, the local relevance between the data can be mined, the complex mode and the change of the data can be extracted, the high-level abstract feature of the data can be extracted, the influence of noise and abnormal values in the data on the feature value can be avoided, and the feature extracted by the convolution self-encoder has abstract property and robustness.

Further, for each feature to be verified, determining a feature evaluation index value corresponding to the feature to be verified includes:

for each feature to be verified, monotonicity calculation is carried out on the feature to be verified, and monotonicity index values corresponding to the feature to be verified are obtained;

for each feature to be verified, performing time correlation calculation on the feature to be verified to obtain a time correlation index value corresponding to the feature to be verified;

for each feature to be verified, determining a feature evaluation index value corresponding to the feature to be verified according to the monotonicity index value and the time correlation index value corresponding to the feature to be verified;

the determining a target feature set according to the feature evaluation index value comprises the following steps:

for each feature to be verified, if the feature evaluation index value corresponding to the feature to be verified is larger than a preset influence threshold, taking the feature to be verified as a target feature;

and determining a target feature set according to each target feature.

The beneficial effects of adopting the further scheme are as follows: by means of monotonicity and time correlation calculation, characteristics which cannot reflect the change of the health state of the bearing can be filtered, characteristic dimensions are reduced, the obtained target characteristics have the characteristics of strong time correlation and high monotonicity, the change of the health degree of the bearing can be reflected better, and the construction of a prediction model of the health state of the bearing with high prediction accuracy is facilitated.

Further, the obtaining, according to the target feature set, a plurality of health degree trend features and feature data corresponding to each health degree trend feature includes:

classifying each target feature by a clustering method to obtain a plurality of similar feature sets, wherein each similar feature set comprises a plurality of target features;

and for each similar feature set, carrying out feature dimension reduction fusion on feature data corresponding to each target feature in the similar feature set to obtain health degree trend features and feature data corresponding to the health degree trend features.

The beneficial effects of adopting the further scheme are as follows: by fusing the characteristics with similar change rules, the feature dimension can be further reduced, the trend characteristics which can better reflect the change of the health state can be obtained, the complexity and the calculated amount of the bearing health state prediction model are reduced, and the model prediction accuracy is improved.

Further, predicting the health state of the target bearing according to the target health trend feature set, including:

inputting the target health degree trend feature set into a trained bearing health state prediction model, and predicting the health state of the target bearing through the bearing health state prediction model;

Predicting the remaining life of the target bearing according to the target health trend feature set, including:

and inputting the target health degree trend feature set into a trained bearing residual life prediction model, and predicting the residual life of the target bearing through the bearing residual life prediction model.

The beneficial effects of adopting the further scheme are as follows: when the bearing fails, the working state and performance of the bearing can be changed, and under the condition, the health state and the residual life of the bearing can be predicted more accurately; meanwhile, after the fault state of the bearing is determined to be faulty, the service state and the service life change condition of the bearing can be better known by utilizing the trained health state and the residual service life corresponding to the predicted bearing according to the operation data of the bearing in the initial stage of the fault, so that relevant personnel can take corresponding measures conveniently, and a data basis is provided for enterprises to improve the production efficiency of the bearing and reduce the production cost of the bearing.

In order to solve the technical problem, the invention also provides a system for predicting the health state and the residual life of the bearing, which comprises:

the data acquisition module is used for acquiring a target vibration signal of the faulty target bearing after the faulty target bearing runs within a set time length;

The data extraction module is used for acquiring a target health degree trend feature set corresponding to the target bearing according to the target vibration signal, wherein the target health degree trend feature set comprises a plurality of preset health degree trend features and feature data corresponding to the health degree trend features;

and the prediction module is used for predicting the health state and/or the residual life of the target bearing according to the target health degree trend characteristic set.

In order to solve the technical problem, the invention also provides electronic equipment, which comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, wherein the processor realizes the method for predicting the health state and the residual life of the bearing when executing the computer program.

In order to solve the above technical problem, the present invention further provides a computer readable storage medium, on which a computer program is stored, which when executed by a processor, implements the method for predicting the health status and the remaining life of a bearing as described above.

Drawings

FIG. 1 is a flow chart of a method for predicting the health status and remaining life of a bearing according to the present invention;

FIG. 2 is a schematic diagram of the result of singular spectrum analysis of the target vibration signal in the present invention;

FIG. 3 is a schematic diagram of a vibration signal sampling arrangement according to the present invention;

FIG. 4 is a schematic diagram of the time domain statistics and corresponding feature data according to the present invention;

FIG. 5 is a schematic diagram of the frequency domain statistics and corresponding feature data according to the present invention;

FIG. 6 is a schematic diagram of the node energy characteristics and their corresponding characteristic data according to the present invention;

FIG. 7 is a schematic diagram of the time domain reconstruction feature and its corresponding feature data according to the present invention;

FIG. 8 is a schematic diagram of the frequency domain reconstruction feature and its corresponding feature data according to the present invention;

FIG. 9 is a cluster tree diagram generated when classifying the target features in the present invention;

FIG. 10 is a schematic diagram of health trend features and corresponding feature data according to the present invention;

FIG. 11 is a schematic illustration of determining bearing failure time points in the present invention;

FIG. 12 is a schematic diagram showing the effect of predicting the health and remaining life of a bearing according to the present invention.

Detailed Description

The principles and features of the present invention are described below with examples given for the purpose of illustration only and are not intended to limit the scope of the invention.

Example 1

As shown in fig. 1, the present embodiment provides a method for predicting a health state and a remaining life of a bearing, including:

step S1, obtaining a target vibration signal of a faulty target bearing running in a set time period, wherein the target vibration signal comprises a plurality of signals to be decomposed;

Optionally, the fault state is faulty and non-faulty, and the fault state of the target bearing is determined by:

；

wherein ,

representing the state value corresponding to the signal to be decomposed, < >>

Absolute value of differential value representing the mean sequence corresponding to the signal to be decomposed, < >>

Representing the mean threshold,/->

Representing the standard deviation threshold;

In this embodiment, the specific method for obtaining the reconstructed vibration signal is as follows: and decomposing the signal to be decomposed into 10 signal sequences (corresponding to the decomposition sequences in fig. 2), and performing signal reconstruction on the 2 nd to 8 th signal sequences to obtain a reconstructed vibration signal. According to the method, the trend signals and the noise signals in the target vibration signals are removed by carrying out singular spectrum analysis on the signals to be decomposed.

In this embodiment, the method for determining the average value threshold includes: acquiring operation data of a test bearing with a fault state being free of faults within a preset time threshold, acquiring an average value of absolute values of differential data of feature data corresponding to the health degree trend features in the operation data, and taking the average value as an average value threshold; the standard deviation threshold value determining method comprises the following steps: acquiring operation data of a test bearing with a fault state being free of faults within a preset time threshold, acquiring standard deviation of absolute values of differential data of feature data corresponding to the health degree trend features in the operation data, and taking the standard deviation as a standard deviation threshold. Wherein the time threshold may be 30 minutes.

The preset health trend features are obtained through the following steps:

The sampling bearing can be any bearing which has no fault and has the same material, the same structure and the same technical characteristics as the target bearing, the characteristics of the bearing without fault can be accurately reflected through the determined health trend characteristics of the bearing, and the health trend characteristics of the bearing without fault can be directly used as the health trend characteristics of the bearing without fault when judging the health state of the bearing without fault, so that the data processing efficiency can be improved.

In this application, feature data corresponding to the health trend features of the sampling bearing may be determined according to the target feature set.

In this embodiment, the first operation condition is an operation duration corresponding to a period from when the sampling bearing starts to operate to when the sampling bearing fails to operate, the sampling bearing adopts a UER204 rolling bearing, an acquisition frequency of a vibration signal of the sampling bearing is 25.6kHz, a sampling interval is 1min, and a single sampling duration is 1.28s, as shown in fig. 3.

The determining, according to the vibration signal, a time domain statistical feature set, a frequency domain statistical feature set and a node energy feature set corresponding to the sampling bearing includes:

normalizing the time domain waveform of the vibration signal to obtain a target time domain waveform, and reconstructing the target time domain waveform through the convolution self-encoder to obtain a time domain reconstruction feature set, wherein the time domain reconstruction features in the time domain reconstruction feature set are in one-to-one correspondence with the time domain statistics features in the time domain statistics feature set, and the number of the time domain reconstruction features in the time domain reconstruction feature set is equal to the number of the time domain statistics features in the time domain statistics feature set;

and carrying out normalization processing on the frequency spectrum of the vibration signal to obtain a target frequency spectrum, and reconstructing the target frequency spectrum through the convolution self-encoder to obtain a frequency domain reconstruction feature set, wherein the frequency domain reconstruction features in the frequency domain reconstruction feature set are in one-to-one correspondence with the time domain statistics features in the time domain statistics feature set, and the number of the frequency domain reconstruction features in the frequency domain reconstruction feature set is equal to the number of the time domain statistics features in the time domain statistics feature set.

In this embodiment, the time domain statistical features include a mean value, an absolute mean value, a variance, a standard deviation, a peak value, a root mean square, a square root amplitude, a kurtosis, a skewness factor, a peak index, a peak-to-peak value, an average energy, a margin index, a waveform index and a pulse index, as shown in fig. 4; the frequency domain statistics of the vibration signal, including center of gravity frequency, mean frequency, root mean square frequency, and spectral entropy, are calculated by performing a fast fourier transform on the vibration signal, as shown in fig. 5.

In this embodiment, the performing signal decomposition on the vibration signal by using a wavelet packet transformation method to obtain a node energy feature set includes:

Decomposing the vibration signal to the first by wavelet packet transformation method

Layer, get->

Individual wavelet packet nodes and individual ones of said wavelet packet nodesThe wavelet packet decomposition coefficients corresponding to the points comprise a plurality of wavelet packet decomposition coefficients corresponding to each wavelet packet node;

for each wavelet packet node, calculating the square sum of a plurality of wavelet packet decomposition coefficients corresponding to the wavelet packet node to obtain node energy characteristics;

and obtaining a node energy feature set according to the node energy features.

In this embodiment, when the signal decomposition is performed on the vibration signal by the wavelet packet transformation method, the vibration signal is decomposed into three by the wavelet base db2 according to the energy distribution and the signal complexity of the vibration signal (i.e.

) The layer can reduce the signal decomposition calculated amount, improve the calculation speed and well reserve the characteristics of different frequency bands. In this embodiment, the node energy feature set includes a node 1 energy feature, a node 2 energy feature, a node 3 energy feature, a node 4 energy feature, a node 5 energy feature, a node 6 energy feature, a node 7 energy feature, and a node 8 energy feature, as shown in fig. 6.

In this embodiment, the normalizing the time domain waveform of the vibration signal specifically includes: normalizing the time domain waveform of the vibration signal to a normalization function MaxAbsScaler

Is within the range of (2); the normalization processing is performed on the frequency spectrum of the vibration signal, specifically: normalizing the spectrum of the vibration signal to +.>

Within a range of (2).

The convolutional self-encoder is a neural network which takes input information as a learning target and performs characterization learning on the input information, and comprises an encoder and a decoder. In this embodiment, the encoder includes three one-dimensional convolutional layers and three pooling layers, one of which is threeThe dimension convolution layers are respectively a first convolution layer, a second convolution layer and a third convolution layer, the three pooling layers are respectively a first pooling layer, a second pooling layer and a third pooling layer, the first convolution layer, the first pooling layer, the second convolution layer, the second pooling layer, the third convolution layer and the third pooling layer are sequentially connected, and the parameters of each dimension convolution layer comprise

、/>

、/>

、/>

And

the parameters of each layer of the pooling layer comprise +. >

、/>

and />

, wherein ,/>

Representing input dimension +.>

Representing output dimension +.>

Representing convolution kernel size, +.>

Indicates step size, & gt>

Representing the fill size; the decoder comprises five layers of one-dimensional deconvolution layers, namely a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, a fourth deconvolution layer and a fifth deconvolution layer, wherein the first deconvolution layer, the second deconvolution layer, the third deconvolution layer, the fourth deconvolution layer and the fifth deconvolution layer are sequentially connected, and the parameters of each one-dimensional deconvolution layer comprise>

、/>

、/>

、/>

and />

。

In order to improve the reconstruction effect of the convolution self-encoder on signals and reduce the data dimension output by the encoder and improve the calculation speed, the method sets the numerical value of each parameter of the third convolution layer as follows:

，

，/>

，/>

，/>

the method comprises the steps of carrying out a first treatment on the surface of the The values of the parameters of the third pooling layer are set as follows: />

、/>

and />

. In this embodiment, the parameters in the convolutional encoder are set as shown in table 1.

TABLE 1 parameter settings for each component in convolutional self-encoder

When reconstructing the target time domain waveform by the convolution self-encoder, determining feature data corresponding to each time domain reconstruction feature by calculating output data of the encoder, wherein the time domain reconstruction feature comprises a mean value, an absolute average value, a variance, a standard deviation, a peak value, a root mean square, a square root amplitude, a kurtosis, a skewness factor, a peak value index, a peak-to-peak value, an average energy, a margin index, a waveform index and a pulse index, as shown in fig. 7; when the convolution self-encoder is used for reconstructing the target frequency spectrum, the output data of the encoder is calculated to determine feature data corresponding to each frequency domain reconstruction feature, wherein the frequency domain reconstruction feature comprises a mean value, an absolute average value, a variance, a standard deviation, a peak value, a root mean square, a square root amplitude, a kurtosis, a skewness factor, a peak value index, a peak-to-peak value, average energy, a margin index, a waveform index and a pulse index, as shown in fig. 8.

In this method, for each feature to be verified in the feature set to be verified, the feature data corresponding to the feature to be verified includes a plurality of feature data, each feature data is obtained by collecting the vibration signal according to the first operating condition, and the feature data corresponding to the feature to be verified is obtained according to the sequence of sampling time points corresponding to the feature data.

In order to remove the characteristics which are irrelevant to the bearing degradation process or can not well reflect the bearing degradation process from the characteristic set to be verified, the method completes the screening of target characteristics based on Monotonicity and time Correlation coreaction. Wherein, for each feature to be verified, determining a feature evaluation index value corresponding to the feature to be verified includes:

and determining a target feature set according to each target feature.

In this embodiment, for each feature to be verified, a monotonicity index value corresponding to the feature to be verified is determined by a second formula, where the second formula is:

；

wherein ,

representing the feature to be verified->

Corresponding monotonicity index value,/->

Representing the +.f in the feature set to be verified>

Target feature->

Representing the feature to be verified->

The number of difference values in the difference sequence of the corresponding normalized data is positive, < >>

Representing the feature to be verified->

Number of negative differential values in differential sequences of corresponding normalized data, < >>

Representing the feature to be verified->

The total number of differential values in the differential sequence of the corresponding normalized data; said feature to be verified->

The corresponding normalized data is applied to the feature to be verified by means of a normalization function MinMaxScale>

The corresponding feature data is normalized to determine, in particular, the feature to be verified is ++ >

Corresponding characteristic data are normalized to +.>

Within a range of (2).

In this embodiment, for each feature to be verified, a time correlation index value corresponding to the feature to be verified is determined by a third formula, where the third formula is:

；

wherein ,

representing the feature to be verified->

Corresponding time dependency index value,/>

Representing the feature to be verified->

Differential sequence and time sequence of corresponding normalized data +.>

Covariance of the time sequence +.>

According to the first operating condition (i.e. the time sequence +.>

Setting the operation time period from the start of the operation of the sampling bearing to the failure of the operation)>

Representing the feature to be verified->

Variance of the differential sequence of the corresponding normalized data, +.>

Representing the time sequence->

Is a variance of (c).

In this embodiment, for each feature to be verified, a feature evaluation index value corresponding to the feature to be verified is determined by a fourth formula, where the fourth formula is:

；

wherein ,

representing the feature to be verified->

Corresponding characteristic evaluation index value,/>

The value of (2) characterizes the extent to which the corresponding feature to be verified affects the bearing degradation process,/->

The larger the value of (c) indicates that the corresponding feature to be verified has a greater impact on the bearing degradation process. The feature to be verified with the corresponding feature evaluation index value greater than 1.4 can fully reflect the bearing degradation process, and in this embodiment, the feature to be verified with the corresponding feature evaluation index value greater than 1.4 is taken as a target feature, and the target feature and the corresponding monotonicity index value, time correlation index value and feature evaluation index value are shown in table 2.

TABLE 2 target characteristics and associated index values corresponding thereto

The obtaining, according to the target feature set, a plurality of health degree trend features and feature data corresponding to the health degree trend features includes:

In this embodiment, classifying each of the target features by using a clustering method to obtain a plurality of similar feature sets includes:

for each of the target features, a window length of

Performing smoothing processing on the feature data corresponding to the target feature to obtain target feature data corresponding to the target feature; wherein (1)>

May have a value of 20;

for each target feature, calculating the similarity between target feature data corresponding to the target feature and target feature data corresponding to each target feature except the target feature, and determining whether the target feature and each target feature except the target feature are similar features according to the size relation between the similarity and a preset clustering threshold;

And dividing each target feature determined to be the similar feature into the same set to obtain a plurality of similar feature sets.

In this embodiment, classification of the target features is completed by adopting a hierarchical clustering method, specifically, a euclidean distance method is used to calculate the euclidean distance between the target feature data corresponding to two target features, the euclidean distance is used as the similarity between the target feature data corresponding to two target features, two or more target features with the corresponding similarity smaller than the clustering threshold value are classified as similar features, wherein the value of the clustering threshold value may be 0.5, a cluster tree diagram generated when the target features are classified by adopting the hierarchical clustering method is shown in fig. 9, an abscissa represents the target features, and an ordinate represents the euclidean distance between the target feature data corresponding to the target features, for example, the euclidean distance between the target feature data corresponding to the energy features of the node 8 and the target feature data corresponding to the energy features of the node 7 is 0.254, and the euclidean distance between the target feature data corresponding to the square root in the frequency domain reconstructed features after the mean value and the absolute mean value in the frequency domain reconstructed features are combined. And classifying each target feature by a hierarchical clustering method to obtain three similar feature sets.

Principal component analysis (Principal Component Analysis, PCA for short) can be used to reduce the dimensionality of the data set while preserving features in the data set that contribute most to the variance; singular value decomposition (Singular Value Decomposition, abbreviated as SVD) and PCA both belong to matrix decomposition algorithms, and when the PCA is used for carrying out data dimension reduction operation, the SVD is utilized to complete the calculation of a complex matrix in the PCA, so that the operation efficiency of the PCA can be improved. In this embodiment, for each of the similar feature sets, feature dimension reduction fusion is performed on feature data corresponding to each target feature in the similar feature set to obtain a health degree trend feature and feature data corresponding to the health degree trend feature, including:

for each similar feature set, constructing a similar data matrix corresponding to the similar feature set according to feature data corresponding to each target feature in the similar feature set

；

For each of the homogeneous feature sets, a homogeneous data matrix corresponding to the homogeneous feature set

Singular value decomposition is carried out to obtain a singular value matrix corresponding to the similar feature set>

The singular value matrix->

Including a plurality of singular values;

for the followingEach homogeneous feature set is used for carrying out singular value matrix corresponding to the homogeneous feature set

Divided into->

Columns, get multiple submatrices->

According to the respective said submatrix +.>

Determining a plurality of dimension reduction feature data corresponding to the same feature set

； wherein ,/>

Is a positive integer>

，/>

，/>

；

For each of the similar feature sets, according to the multiple dimension-reducing feature data corresponding to the similar feature sets

And carrying out principal component analysis, and determining health degree trend characteristics and characteristic data corresponding to the health degree trend characteristics. In this embodiment, for each of the homogeneous feature sets +.>

The corresponding dimension reduction feature data is used as feature data corresponding to health degree trend features determined according to the similar feature sets, and the feature data corresponding to the health degree trend features keeps each of the similar feature setsAccording to the feature information of the target feature, three health degree trend features and corresponding feature data thereof are obtained according to the three similar feature sets, as shown in fig. 10, and the feature data corresponding to each health degree trend feature at different moments is represented by the amplitude values in fig. 10.

In the present embodiment of the present invention,

the criterion is used for determining a fault state corresponding to the target bearing as well as a fault time point of the target bearing with the fault state, wherein the determining method of the fault time point of the target bearing with the fault state comprises the following steps:

Acquiring a target vibration signal after the fault state is that the target bearing with the fault operates within a set time period, wherein the target vibration signal comprises a plurality of signals to be decomposed;

acquiring a state value corresponding to each signal to be decomposed, wherein each state value corresponds to a signal acquisition time point;

generating a state value sequence corresponding to the target bearing according to the sequence of the signal acquisition time points;

and taking a signal acquisition time point corresponding to a first state value which is more than or equal to 0 in the state value sequence as a fault time point of the bearing.

As shown in FIG. 11, the horizontal lines of dots correspond to the values on the vertical axis as

The first ordinate value in FIG. 11 is equal to

The signal acquisition time point corresponding to the point(s) is the fault time point of the target bearing with the fault state. />

Wherein predicting the health state of the target bearing according to the target health trend feature set comprises:

In order to realize the prediction of the health state of the target bearing, the embodiment establishes the bearing health state prediction model through the LSTM, and the LSTM is a special recurrent neural network, and the special forgetting gate structure enables the recurrent neural network to avoid the problem that gradient vanishes or gradient explosions easily occur in the recurrent neural network, so that the method has good effect in time sequence data prediction.

In this embodiment, the bearing health state prediction model includes two LSTM layers and a Dense layer, where the two LSTM layers are a first LSTM layer and a second LSTM layer, the number of units of the first LSTM layer and the second LSTM layer is 32 and 8, and, in order to ensure accuracy of a prediction result, a Dropout layer is further disposed between the two LSTM layers, and a parameter-Dropout ratio of the Dropout layer is set to 0.1; the connection relation among all the constituent units of the bearing health state prediction model is as follows: the first LSTM layer, the Dropout layer, the second LSTM layer and the Dense layer are sequentially connected.

In this embodiment, in order to train the bearing health state prediction model, the window length of the data input to the bearing health state prediction model (i.e., the feature data corresponding to each health degree trend feature in the target health degree trend feature set corresponding to the sampling bearing) is selected to be 20, iterative training is performed by the Adam optimizer until the loss function reaches a preset value, so as to obtain a trained bearing health state prediction model (i.e., a pre-built bearing health state prediction model), the number of data used in one iteration during model training is 20, the training frequency is 100, and the loss function adopts a mean square error MSE.

In order to achieve prediction of the remaining life of the target bearing, the embodiment uses an exponential model to build a bearing remaining life prediction model, and fits the health degree of the target bearing (i.e. the health state of the target bearing) through the bearing remaining life prediction model, so as to complete prediction of the remaining life of the target bearing, wherein the bearing remaining life prediction model is as follows:

；

wherein ,

fitting results (corresponding to the fitting curve in fig. 12) representing fitting of the health degree of the target bearing by the bearing remaining life prediction model, +. >

For the health of the target bearing, +.>

、/>

、/>

Constant coefficients of the bearing residual life prediction model are obtained; in this embodiment, <' > a->

，/>

，/>

The setting of each constant coefficient can be adjusted according to the prediction accuracy of the bearing residual life prediction model.

In the present embodiment of the present invention,

the corresponding time point when the value of (2) is 0 is theThe effect of predicting the remaining life of the target bearing with respect to the degree of health and the remaining life of the bearing in which the partial failure state is a failure is shown in fig. 12, and in fig. 12, the predicted value is a predicted value of the remaining life of the bearing in which the failure state is a failure by the present method, and the actual value is an actual value of the remaining life of the bearing in which the failure state is a failure.

Example two

Based on the same principle as the method for predicting the health state and the remaining life of the bearing in the first embodiment, the present embodiment provides a system for predicting the health state and the remaining life of the bearing, including:

Optionally, the target vibration signal includes a plurality of signals to be decomposed; the system also includes a fault condition prediction module that includes:

the signal decomposition unit is used for carrying out singular spectrum analysis on the signals to be decomposed for each signal to be decomposed to obtain reconstructed vibration signals corresponding to the signals to be decomposed;

the average value sequence determining unit is used for reconstructing the frequency spectrum of the reconstructed vibration signal corresponding to each signal to be decomposed through the trained convolution self-encoder, calculating the average value of the characteristic data output by the convolution self-encoder and obtaining an average value sequence corresponding to the signal to be decomposed;

the state value determining unit is configured to determine, for each signal to be decomposed, a state value corresponding to the signal to be decomposed according to a mean value sequence corresponding to the signal to be decomposed and a preset index threshold, by using a first formula, where the index threshold includes a mean value threshold and a standard deviation threshold, and the first formula is:

；

wherein ,

representing the state value corresponding to the signal to be decomposed, < >>

Representing the mean threshold,/->

Representing the standard deviation threshold;

the fault state determining unit is used for determining the fault state of the target bearing, and if the state value which is greater than or equal to 0 exists in the state value corresponding to each signal to be decomposed, the fault state of the target bearing is faulty; if the state value which is larger than or equal to 0 does not exist in the state value which corresponds to each signal to be decomposed, the fault state of the target bearing is fault-free.

Wherein, the system also includes a health trend feature determination module, the health trend feature determination module includes:

the sampling signal acquisition unit is used for acquiring a vibration signal of the sampling bearing after the sampling bearing operates under a preset first operating condition, and the fault state of the sampling bearing is fault-free;

the first correlation feature determining unit is used for determining a time domain statistical feature set, a frequency domain statistical feature set and a node energy feature set corresponding to the sampling bearing according to the vibration signal, wherein the time domain statistical feature set comprises feature data corresponding to each of a plurality of time domain statistical features, the frequency domain statistical feature set comprises feature data corresponding to each of a plurality of frequency domain statistical features, and the node energy feature set comprises feature data corresponding to each of a plurality of node energy features;

The second correlation characteristic determining unit is used for reconstructing the time domain waveform of the vibration signal to obtain a time domain reconstruction characteristic set, wherein the time domain reconstruction characteristic set comprises characteristic data corresponding to a plurality of time domain reconstruction characteristics;

the third related characteristic determining unit is used for reconstructing the frequency spectrum of the vibration signal to obtain a frequency spectrum reconstruction characteristic set, wherein the frequency spectrum reconstruction characteristic set comprises characteristic data corresponding to a plurality of frequency spectrum reconstruction characteristics;

the feature to be verified determining unit is configured to construct a feature set to be verified according to the time domain statistical feature set, the frequency domain statistical feature set, the node energy feature set, the time domain reconstruction feature set and the frequency spectrum reconstruction feature set, where the feature set to be verified includes a plurality of features to be verified, and each feature to be verified corresponds to one of the time domain statistical feature, the frequency domain statistical feature, the node energy feature, the time domain reconstruction feature or the frequency spectrum reconstruction feature;

a target feature determining unit, configured to determine, for each feature to be verified, a feature evaluation index value corresponding to the feature to be verified, where the feature evaluation index value characterizes an influence degree between the corresponding feature to be verified and the degradation of the sampling bearing, and determine, according to the feature evaluation index value, a target feature set, where the target feature set includes a plurality of target features, and each target feature corresponds to one feature to be verified;

And the health degree trend feature determining unit is used for obtaining a plurality of health degree trend features corresponding to the sampling bearing according to the target feature set, and taking the plurality of health degree trend features corresponding to the sampling bearing as the preset plurality of health degree trend features.

The first correlation characteristic determining unit is configured to determine, according to the vibration signal, a time domain statistical characteristic set, a frequency domain statistical characteristic set, and a node energy characteristic set corresponding to the sampling bearing, where the first correlation characteristic determining unit is specifically configured to:

and carrying out signal decomposition on the vibration signal by a wavelet packet transformation method to obtain a node energy feature set.

The second correlation characteristic determining unit is configured to reconstruct a time domain waveform of the vibration signal, and is specifically configured to:

normalizing the time domain waveform of the vibration signal to obtain a target time domain waveform, and reconstructing the target time domain waveform through the convolution self-encoder to obtain a time domain reconstruction feature set, wherein time domain reconstruction features in the time domain reconstruction feature set are in one-to-one correspondence with time domain statistics features in the time domain statistics feature set.

The third correlation characteristic determining unit is configured to reconstruct a spectrum of the vibration signal, and is specifically configured to:

Wherein the target feature determination unit includes:

a monotonicity index value determining subunit, configured to perform monotonicity calculation on the feature to be verified for each feature to be verified, to obtain a monotonicity index value corresponding to the feature to be verified;

the time correlation index value determining subunit is used for carrying out time correlation calculation on the feature to be verified for each feature to be verified to obtain a time correlation index value corresponding to the feature to be verified;

a feature evaluation index value determining subunit, configured to determine, for each feature to be verified, a feature evaluation index value corresponding to the feature to be verified according to a monotonicity index value and a time correlation index value corresponding to the feature to be verified;

A target feature set determining subunit, configured to determine a target feature set according to the feature evaluation index value;

the target feature set determination subunit is specifically configured to:

and determining a target feature set according to each target feature.

Wherein the health trend feature determination unit includes:

the similar feature dividing subunit is used for classifying each target feature through a clustering method to obtain a plurality of similar feature sets, wherein each similar feature set comprises a plurality of target features;

and the health degree trend feature determining subunit is used for carrying out feature dimension reduction fusion on feature data corresponding to each target feature in the similar feature sets for each similar feature set to obtain health degree trend features and feature data corresponding to the health degree trend features.

Optionally, the prediction module includes:

the health state prediction unit is used for inputting the target health degree trend feature set into a trained bearing health state prediction model, and predicting the health state of the target bearing through the bearing health state prediction model;

The residual life prediction unit is used for inputting the target health degree trend feature set into a trained bearing residual life prediction model, and predicting the residual life of the target bearing through the bearing residual life prediction model.

Example III

In order to solve the above technical problem, the present embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, where the processor implements the method for predicting the health status and the remaining life of a bearing according to the first embodiment when executing the computer program.

Example IV

To solve the above-mentioned problems, the present embodiment provides a computer readable storage medium, on which a computer program is stored, which when executed by a processor, implements the method for predicting the health status and the remaining life of a bearing according to the first embodiment.

In the description of the present invention, it should be understood that the terms "first," "second," and the like are used for descriptive purposes only and are not to be construed as indicating or implying a relative importance or number of technical features indicated. Thus, a feature defining "a first" or "a second" may explicitly or implicitly include at least one such feature. In the description of the present invention, the meaning of "plurality" means at least two, for example, two, three, etc., unless specifically defined otherwise.

In the description of the present specification, a description referring to the terms "one embodiment," "some embodiments," "examples," "particular examples," "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms are not necessarily directed to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, the different embodiments or examples described in this specification and the features of the different embodiments or examples may be combined and combined by those skilled in the art without contradiction.

While embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and not to be construed as limiting the invention, and that variations, modifications, alternatives and variations may be made to the above embodiments by one of ordinary skill in the art within the scope of the invention.

Claims

1. A method for predicting the health status and remaining life of a bearing, comprising:

2. The method of claim 1, wherein the target vibration signal comprises a plurality of signals to be decomposed;

the fault state of the target bearing is determined by:

；

wherein ,

representing the state value corresponding to the signal to be decomposed, < >>

Representing the mean threshold,/->

Representing the standard deviation threshold;

3. The method of claim 1, wherein the predetermined plurality of health trend features are obtained by:

4. A method according to claim 3, wherein determining the time domain statistical feature set, the frequency domain statistical feature set and the node energy feature set corresponding to the sampling bearing according to the vibration signal comprises:

5. A method according to claim 3, wherein said determining, for each of the features to be verified, a feature evaluation index value corresponding to the feature to be verified comprises:

and determining a target feature set according to each target feature.

6. The method according to claim 3, wherein the obtaining, from the target feature set, a plurality of health trend features and feature data corresponding to each of the health trend features includes:

7. The method according to any one of claims 1 to 6, wherein predicting the health state of the target bearing from the target health trend feature set comprises:

8. A system for predicting the health and remaining life of a bearing, comprising:

9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method of predicting bearing health and remaining life as claimed in any one of claims 1 to 7 when the computer program is executed.

10. A computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, which when executed by a processor, implements the method for predicting the health and remaining life of a bearing according to any one of claims 1 to 7.