CN113158814B - Bearing health state monitoring method based on convolution self-encoder - Google Patents
Bearing health state monitoring method based on convolution self-encoder Download PDFInfo
- Publication number
- CN113158814B CN113158814B CN202110328161.3A CN202110328161A CN113158814B CN 113158814 B CN113158814 B CN 113158814B CN 202110328161 A CN202110328161 A CN 202110328161A CN 113158814 B CN113158814 B CN 113158814B
- Authority
- CN
- China
- Prior art keywords
- bearing
- health state
- encoder
- value
- digital vibration
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000012544 monitoring process Methods 0.000 title claims abstract description 37
- 238000000034 method Methods 0.000 title claims abstract description 32
- 238000012549 training Methods 0.000 claims abstract description 20
- 238000012216 screening Methods 0.000 claims abstract description 15
- 238000005070 sampling Methods 0.000 claims description 78
- 238000004364 calculation method Methods 0.000 claims description 19
- 230000006870 function Effects 0.000 claims description 18
- 238000000354 decomposition reaction Methods 0.000 claims description 15
- 238000004458 analytical method Methods 0.000 claims description 12
- 230000015556 catabolic process Effects 0.000 claims description 12
- 238000006731 degradation reaction Methods 0.000 claims description 12
- 238000007781 pre-processing Methods 0.000 claims description 10
- 230000003862 health status Effects 0.000 claims description 9
- 238000013135 deep learning Methods 0.000 claims description 8
- 238000000605 extraction Methods 0.000 claims description 6
- 238000012163 sequencing technique Methods 0.000 claims description 5
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 241000820057 Ithone Species 0.000 claims 1
- 238000010586 diagram Methods 0.000 description 3
- 238000004140 cleaning Methods 0.000 description 2
- 238000013144 data compression Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000011176 pooling Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000005802 health problem Effects 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000010187 selection method Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/08—Feature extraction
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
- G01M13/04—Bearings
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
- G01M13/04—Bearings
- G01M13/045—Acoustic or vibration analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E10/00—Energy generation through renewable energy sources
- Y02E10/70—Wind energy
- Y02E10/72—Wind turbines with rotation axis in wind direction
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Pure & Applied Mathematics (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Computational Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Acoustics & Sound (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Operations Research (AREA)
- Probability & Statistics with Applications (AREA)
- Signal Processing (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention provides a bearing health state monitoring method based on a convolution self-encoder, and belongs to the field of bearing fault prediction and health management. Firstly, acquiring a full-life-cycle digital vibration signal and a health state mark value of a brand-new bearing; respectively extracting the eigenmode component statistic characteristics of the digital vibration signal and the depth characteristics learned by a convolution self-encoder, splicing and screening the two characteristics, inputting the screened characteristics into a fully-connected regression network for regression training, and finally obtaining a health state curve graph of the bearing in the full life cycle; then, acquiring a current health state curve chart of the bearing to be monitored in the same model; and comparing the two graphs to obtain the health state monitoring result of the bearing to be monitored. According to the invention, the integrity of the bearing vibration signal characteristics is improved through the convolution self-encoder, redundant characteristics are removed by using a characteristic sorting and characteristic selecting method, and a more accurate bearing health state can be obtained.
Description
Technical Field
The invention belongs to the field of bearing fault prediction and health management, and particularly relates to a bearing health state monitoring method based on a convolution self-encoder.
Background
The bearing is an essential component in various rotary machines, whether the bearing is healthy or not directly affects the operation condition of the whole rotary machine system, and particularly in large-scale mechanical scenes, such as mechanical systems of airplanes, wind driven generators, elevators and the like, once the health problem of the bearing component occurs, great life and property loss can be caused. Therefore, in recent years, health monitoring of bearings and failure prediction thereof have been receiving more and more attention. For monitoring the health state of a bearing in a rotating machine, the health state cannot be obtained through direct measurement or visual estimation. Currently, the health status of the bearing is mostly obtained by analyzing the vibration signal of the bearing through a common monitoring method.
The bearing vibration signal is very complex, a large amount of characteristic information is contained in the one-dimensional vibration signal, the characteristic information needs to be obtained from the vibration signal through a certain analysis method, and the common bearing vibration signal analysis method mostly adopts a direct decomposition mode or a mode of decomposition after filtering. Empirical mode decomposition is a method commonly used in direct decomposition. A feature extraction method based on empirical mode decomposition is characterized in that feature extraction is carried out on the basis of empirical mode decomposition, firstly vibration signals are decomposed into a plurality of intrinsic mode components, then statistical feature analysis is carried out on each intrinsic mode component, the method only obtains vibration signal features based on expert experience, certain limitation exists in the aspect of feature integrity, a large number of redundant features exist in the obtained numerous features, and the redundant features have negative effects on monitoring of the bearing health state, so that the accuracy of monitoring of the bearing health state is influenced.
Disclosure of Invention
The invention aims to overcome the defects of complex bearing vibration signals, incomplete feature extraction and redundant features in the health state monitoring of bearing parts in the conventional rotary mechanical equipment, and provides a bearing health state monitoring method based on a convolution self-encoder. According to the invention, the integrity of the bearing vibration signal characteristics is improved by adopting the convolution self-encoder, and the redundant characteristics are removed by using a characteristic sorting and characteristic selecting method, so that the more accurate bearing health state can be obtained.
The invention provides a bearing health state monitoring method based on a convolution self-encoder, which is characterized in that firstly, a vibration sensor is arranged on a brand-new bearing, and a digital vibration signal of the whole life cycle of the bearing and a corresponding health state mark value are obtained; two feature extraction operations are respectively carried out on the digital vibration signal: firstly, performing empirical mode decomposition on the digital vibration signal, selecting a plurality of intrinsic mode components and obtaining the statistic characteristics of each selected intrinsic mode component; secondly, inputting the digital vibration signal into a convolution self-encoder to carry out deep learning training, and obtaining the depth characteristic of the input signal after the training is finished; then splicing and screening the two characteristics, inputting the screened characteristics into a fully-connected regression network for regression training, and finally obtaining a health state curve graph of the bearing in the full life cycle; acquiring digital vibration signals of bearings to be monitored in the same model, acquiring a health state predicted value of the bearing by using a trained convolution self-encoder and a fully-connected regression network, and further generating a current health state curve graph of the bearing; and comparing the two graphs to obtain a health state monitoring result of the bearing to be monitored. The method comprises the following steps:
1) selecting a brand-new bearing, and acquiring a digital vibration signal and a corresponding health state mark value of the bearing in the whole life cycle; the method comprises the following specific steps:
1-1) selecting a brand new bearing;
1-2) installing a vibration sensor on a bearing seat of the bearing, acquiring vibration signals of the bearing in a full life cycle from the beginning to the scrapping by using the vibration sensor, and converting the acquired vibration signals into digital vibration signals;
1-3) carrying out preprocessing operation on the digital vibration signals, wherein the preprocessing operation comprises sampling the digital vibration signals obtained in the step 1-2), obtaining the digital vibration signals of all the preprocessed sampling points and calculating a health state marking value corresponding to each sampling point;
2) performing empirical mode decomposition on the digital vibration signals of all the sampling points after the pretreatment obtained in the step 1) to obtain corresponding intrinsic mode components, and then selecting the first N intrinsic mode components to perform statistic feature analysis to obtain statistic features corresponding to each selected intrinsic mode component;
3) inputting the digital vibration signals of all the sampling points preprocessed in the step 1) into a one-dimensional convolution self-encoder with random initialization parameters, carrying out deep learning training on the convolution self-encoder by reducing reconstruction errors, and obtaining the trained convolution self-encoder when the reconstruction errors are not reduced; the convolution self-encoder is composed of an encoder and a decoder;
inputting the digital vibration signals of all the sampling points preprocessed in the step 1) into a trained convolution self-encoder, and outputting depth characteristics corresponding to the input signals by the last layer of the encoder in the convolution self-encoder;
4) splicing all the characteristics obtained in the step 2) and the step 3) according to corresponding sampling points to obtain splicing characteristics of the digital vibration signals of the sampling points after pretreatment, and sequencing and screening the splicing characteristics in sequence by using monotonicity, trend and autocorrelation to obtain screened characteristics;
5) inputting the screened features of the step 4) into a full-connection regression network for regression training, and repeatedly reducing the value of the loss function by adopting a gradient descent method to train the full-connection regression network until the value of the loss function is not reduced any more, and finishing the training of the full-connection regression network; taking the predicted value of the health state index of the bearing at each sampling point output by the trained fully-connected regression network as the health state value of the bearing at the sampling point;
wherein, the calculation expression of the loss function is:
in the formula, labeliIndicating the health status of the ith sample point, predictioniThe predicted value of the health state of the ith sampling point output by the current fully-connected neural network is represented;
utilizing the health state value of the bearing at each sampling point to draw a two-dimensional curve of the change of the health state of the bearing along with time to obtain a health state curve graph of the bearing in the whole life cycle; dividing the full life cycle of the bearing into a stable stage, a degradation stage and a rapid degradation stage according to the change trend of the slope in the graph;
6) obtaining a bearing to be monitored with the same model as the model in the step 1), and obtaining a health state monitoring result of the bearing; the method comprises the following specific steps:
6-1) obtaining a new bearing with the same type as the bearing in the step 1) as a bearing to be monitored, and mounting a vibration sensor with the same type as the vibration sensor in the step 1-2) on a bearing seat of the bearing;
6-2) at any monitoring moment, acquiring a vibration signal of the bearing to be monitored from the beginning of operation to the monitoring moment by using the vibration sensor installed in the step 6-1), and converting the vibration signal into a digital vibration signal; preprocessing the digital vibration signal, and sampling the digital vibration signal to obtain digital vibration signals of all preprocessed sampling points;
6-3) carrying out empirical mode decomposition on the digital vibration signals of all the sampling points preprocessed in the step 6-2) to obtain the first N intrinsic mode components; performing statistic feature analysis on each intrinsic mode component to obtain statistic features corresponding to each intrinsic mode component;
6-4) inputting the digital vibration signals of all sampling points preprocessed in the step 6-2) into the trained convolution self-encoder in the step 3), and outputting depth characteristics corresponding to the input signals from the last layer of the encoder in the convolution self-encoder;
6-5) splicing all the characteristics obtained in the step 6-3) and the step 6-4) according to corresponding sampling points to obtain spliced characteristics, and sequencing and screening the spliced characteristics in sequence by using monotonicity, trend and autocorrelation to obtain screened characteristics;
6-6) inputting the characteristics screened in the step 6-5) into the fully-connected regression network trained in the step 5), and outputting the predicted value of the health state index of the bearing to be monitored at each sampling point as the health state value of the bearing at the sampling point by the network;
obtaining a health state curve graph of the bearing to be monitored from the beginning to the current moment by using the health state value of each sampling point;
6-7) comparing the health state curve graph obtained in the step 6-6) with the health state curve graph of the bearing obtained in the step 5) in the full life cycle to obtain the health state stage of the bearing to be monitored at the monitoring time, and finishing the monitoring.
The invention has the characteristics and beneficial effects that:
1) the deep learning method of the convolution self-encoder is adopted, so that the deep characteristic information in the bearing vibration signal can be extracted in a self-adaptive manner, and the integrity of the bearing vibration signal characteristic is improved.
2) Among the many characteristic quantities extracted from the bearing vibration signal, there are a large number of redundant characteristics that are liable to interfere with the monitoring of the bearing health. Monotonicity reflects the monotonous condition of data, trend reflects the degree of correlation between data and time, and autocorrelation reflects the fluctuation condition of data. Because the degradation condition of the bearing changes more and more strongly along with the time, the characteristics reflecting the health state of the bearing are good in monotonicity, trend and autocorrelation. The sorting selection method can effectively remove redundant features and improve the accuracy of the health state of the bearing.
3) According to the method, the inverse hyperbolic tangent function which is relatively in line with the bearing health state curve is selected as the label with supervision training, so that the obtained bearing health state curve is higher in accuracy and better in detection effect.
Drawings
FIG. 1 is a schematic diagram of a method for monitoring the health status of a bearing based on a convolutional self-encoder according to the present invention;
FIG. 2 is a graph of inverse hyperbolic tangent function according to an embodiment of the present invention;
FIG. 3 is a graph illustrating the original vibration signal and the degradation degree of the bearing according to the embodiment of the invention.
Detailed Description
The invention provides a bearing health state monitoring method based on a convolution self-encoder, which is further explained by combining the attached drawings and a specific implementation mode as follows:
the invention provides a bearing health state monitoring method based on a convolution self-encoder, the principle is shown in figure 1, the method obtains a bearing vibration signal through a vibration sensor of a bearing, and two characteristic extraction operations are respectively carried out on a one-dimensional vibration signal: firstly, carrying out empirical mode decomposition on a vibration signal to obtain a plurality of intrinsic mode components, selecting the first five intrinsic mode components to carry out next statistic analysis operation as the health state information of the bearing mostly exists in a high-frequency signal, and respectively obtaining the statistic characteristics of the mean value, the variance, the kurtosis, the root mean square, the energy and the like of each intrinsic mode component by the statistic analysis operation; secondly, inputting the vibration signal into a convolution self-encoder network for deep learning training, wherein the convolution self-encoder is an unsupervised deep learning method, and thus the deep characteristic of the bearing vibration signal is obtained in a self-adaptive mode. And then splicing the features obtained by the two operations, screening out proper features, inputting the screened features into a fully-connected regression network for regression training, and finally obtaining a health state curve graph of the bearing in the full life cycle. And then, acquiring digital vibration signals of the bearings to be monitored in the same type, namely outputting the predicted value of the health state of the bearings through the trained convolutional self-encoder network and the fully-connected regression network, further generating a current health state curve graph of the bearings, and obtaining the health state monitoring result of the bearings to be monitored through comparison of the two curve graphs.
The invention provides a bearing health state monitoring method based on a convolution self-encoder, which comprises the following steps:
1) and selecting a brand-new bearing, and acquiring a digital vibration signal and a corresponding health state mark value of the bearing in a full life cycle. The method comprises the following specific steps:
1-1) selecting a brand new bearing which can be of any type;
1-2) installing a horizontal vibration sensor (or a vertical vibration sensor) on a bearing seat, acquiring a vibration signal of the bearing in the horizontal direction by using the horizontal vibration sensor (or acquiring the vibration signal of the bearing in the vertical direction by using the vertical vibration sensor), wherein the vibration signal needs to acquire the data of the full life cycle of the bearing from the beginning of operation to the end of scrapping, and converting the acquired vibration signal into a digital vibration signal;
1-3) carrying out preprocessing operation on the digital vibration signals, wherein the preprocessing operation comprises sampling the digital vibration signals obtained in the step 1-2), obtaining the digital vibration signals of all the preprocessed sampling points and calculating to obtain the health state marking value corresponding to each sampling point.
The preprocessing operation in the invention comprises data cleaning, data compression and the like. In this embodiment, data cleaning is performed, i.e., data with amplitude exceeding 20g is removed. Data compression means that when the data amount is huge, sampling selection can be performed again from original data, for example, 1 data can be sampled every 5 data, so that the purposes of compressing data and reducing the operation amount are achieved.
And (4) recording the total number of the sampling points as n, calculating the health state mark value corresponding to each sampling point by adopting an inverse hyperbolic tangent function, namely taking the time corresponding to each sampling point as the input of the inverse hyperbolic tangent function, and outputting the health state mark value corresponding to the sampling point. The expression of the inverse hyperbolic tangent function is:
wherein t isiIndicates the time corresponding to the ith sampling point (i is a natural number 0, 1, 2 …; tiThe unit is second, namely 0 is used for starting the bearing to run, 10 seconds is used for running 10 seconds, and … is used for running 20 seconds, t represents the total service life duration of the bearing (the time elapsed from the start of the bearing to scrapping), yiIs tiAnd marking the health state of the bearing at the moment.
Fig. 2 is an image of the inverse hyperbolic tangent function of the present embodiment, in which the horizontal axis represents the independent variable and the vertical axis represents the dependent variable, i.e., the function value.
2) Carrying out empirical mode decomposition on the preprocessed digital vibration signals of all sampling points obtained in the step 1) to obtain a plurality of intrinsic mode components; because the fault information of the bearing is mostly contained in the high-frequency components, the first N (the first 5) intrinsic mode components are taken to perform the subsequent statistic feature analysis operation, and the statistic feature corresponding to each intrinsic mode component is obtained; the statistic feature analysis operation of each eigenmode component adopts the following calculation mode:
let the number of sampling points be n, xiRepresenting the corresponding value of the ith sampling point in any intrinsic mode component in the intrinsic mode component, wherein each intrinsic mode component comprises the corresponding values of n sampling points; then:
the mean value of each eigenmode component is calculated as:
the variance of each eigenmode component is calculated as:
the kurtosis calculation formula of each eigenmode component is as follows:
the root mean square calculation formula of each eigenmode component is as follows:
the energy calculation formula of each eigenmode component is as follows:
3) inputting the digital vibration signals of all the sampling points preprocessed in the step 1) into a one-dimensional convolution self-encoder with random initialization parameters, carrying out deep learning training on the convolution self-encoder for multiple times by reducing reconstruction errors, and obtaining the trained convolution self-encoder when the reconstruction errors are not reduced.
The convolutional self-encoder is composed of an encoder and a decoder, the encoder is composed of a convolutional layer and a pooling layer, and the decoder is composed of an anti-convolutional layer and an anti-pooling layer.
In the present embodiment, the variation of the data dimension in the encoder is 1, 25, 50, 75, 100, the convolution kernel size is 16, 8, 4, 2, respectively, and the step size is 2, 1, 2, respectively. The variation of data dimension in the decoder is 100, 75, 50, 25, 1, the deconvolution kernel size is 4, 8, 16, 25, 31, respectively, and the step size is 2, 3, respectively.
Inputting the digital vibration signals of all the sampling points preprocessed in the step 1) into a trained convolution self-encoder, and outputting the depth characteristics corresponding to the input signals by the last layer of the encoder in the convolution self-encoder.
4) Splicing all the characteristics obtained in the step 2) and the step 3) according to corresponding vibration signal sampling points to obtain splicing characteristics of the digital vibration signals of the sampling points after pretreatment, and sequencing and screening the splicing characteristics in sequence by using monotonicity, trend and autocorrelation to obtain screened characteristics;
wherein,
the monotonicity calculation formula of each feature in the splicing features is as follows:
wherein z represents any of the stitching features, ziThe value of the ith sample point representing the feature, d/dz ═ zi+1-ziA differential value representing a health status index; no. ofd/dz>0 and No. ofd/dz<0 represents a positive differential value and a negative differential value, respectively. The larger Mon (z) represents the better monotonicity of the feature.
The trend calculation formula is:
in the formula, tiIndicating the time corresponding to the ith value in each featureAt that moment (i.e. the moment corresponding to the ith sampling point),
representing the mean of all time instants to which the features correspond. The larger the Tre (z, t) value, the better the trend.
The autocorrelation calculation formula is as follows:
in the formula, a smaller auto (z) value represents a smaller fluctuation in the characteristic z.
In this embodiment, first, monotonicity of all features is calculated, all features are sorted according to the monotonicity from large to small, 80% of the top ranked features are screened to enter trend screening, then, trend of all the features entering the trend screening is calculated, the features are sorted according to the trend from large to small, 60% of the top ranked features are screened to enter autocorrelation screening, autocorrelation of all the features entering the autocorrelation screening is calculated, the features are sorted according to the autocorrelation from small to large, and the top 40% of the features are taken out as the finally screened features. Thus, feature sorting and feature screening are completed.
5) Inputting the screened features of the step 4) into a fully connected regression network for regression training; in this embodiment, the data dimension of the fully connected regression network is changed as follows: 40. 128, 64, 32, 16, 4, 1, the network outputs a predicted value of the health status indicator corresponding to each sampling point. Repeatedly reducing the value of the loss function by adopting a gradient descent method to train the fully-connected regression network until the value of the loss function is not reduced any more, and finishing the training of the fully-connected regression network; and taking the predicted value of the health state index of the bearing at each sampling point output by the trained fully-connected regression network as the health state value of the bearing at the sampling point.
Wherein, the calculation formula of the loss function is as follows:
in the formula, labeliIndicating the health status of the ith sample point, predictioniAnd the predicted value of the health state of the ith sampling point which represents the output of the current fully-connected neural network.
After the health state value of the bearing at each sampling point (namely the full life cycle) is obtained, a two-dimensional curve of the change of the health state of the bearing along with time is drawn, so that a health state curve chart of the bearing in the full life cycle is obtained.
And obtaining a corresponding slope by calculating a first-order guide of the healthy state curve, wherein when the slope is close to zero or less than 10, the bearing is considered to be in a stable stage, when the slope is larger and less than 100, the bearing is considered to be in a degradation stage, and when the slope is continuously larger and exceeds 100, the bearing is considered to be in a rapid degradation stage.
In fig. 3, fig. 3a is a diagram illustrating a health state of a bearing according to an embodiment of the present invention, and fig. 3b is a diagram illustrating an original vibration signal of the bearing according to an embodiment of the present invention. The health state of the bearing changes at 1 and 2, so that the health state of the bearing can be roughly divided into three stages, namely a steady stage, a degradation stage and a rapid degradation stage according to the state of the bearing. Thus, the bearing health status of the full life cycle is reflected.
6) Obtaining a bearing to be monitored with the same model as the model in the step 1), and obtaining a health state monitoring result of the bearing; the method comprises the following specific steps:
6-1) obtaining a new bearing with the same type as the bearing in the step 1) as a bearing to be monitored, and mounting a vibration sensor with the same type as the vibration sensor in the step 1-2) on a bearing seat of the bearing;
6-2) at any monitoring moment, acquiring a vibration signal of the bearing to be monitored from the beginning of operation to the monitoring moment by using the vibration sensor installed in the step 6-1), and converting the vibration signal into a digital vibration signal; and (3) preprocessing the digital vibration signal (the digital vibration signal comprises all data of the bearing from the beginning to the current moment), and sampling the digital vibration signal to obtain a preprocessed digital vibration signal of each sampling point.
6-3) carrying out empirical mode decomposition on the digital vibration signals of all sampling points preprocessed in the step 6-2) to obtain the first N intrinsic mode components, wherein the types of the N intrinsic mode components can be different from that in the step 2); performing subsequent statistic feature analysis operation on each eigenmode component to obtain statistic features corresponding to each eigenmode component (wherein the selected statistic features are consistent with those in training);
6-4) inputting the digital vibration signals of all the sampling points preprocessed in the step 6-2) into the trained convolution self-encoder in the step 3), and outputting the depth characteristics corresponding to the input signals from the last layer of the encoder in the convolution self-encoder.
6-5) splicing all the characteristics obtained in the step 6-3) and the step 6-4) according to corresponding vibration signal sampling points to obtain splicing characteristics, and repeating the step 4) to obtain a plurality of screened characteristics corresponding to the digital signals in the step 6-2). (wherein, the step only needs to carry out characteristic screening according to the same proportion as the step 4), and the number of the screened characteristics can be different from the result of the step 4)
6-6) inputting the characteristics screened in the step 6-5) into the fully-connected regression network trained in the step 5), and outputting the predicted value of the health state index of the bearing to be monitored at each sampling point as the health state value of the bearing at the sampling point by the network;
and (3) obtaining the health state curve chart of the bearing to be monitored, which is selected in the step 6-1), from the beginning to the current moment by using the health state value of each sampling point.
6-7) comparing the health state curve graph obtained in the step 6-6) with the health state curve graph of the bearing obtained in the step 5) in the full life cycle, so that the health state stage of the bearing to be monitored at the monitoring moment can be obtained, if the bearing is in a stable stage and a degradation stage, the bearing does not need to be maintained, and if the bearing is in a rapid degradation stage, the bearing is about to be damaged, the bearing needs to be replaced by a new bearing, and thus the monitoring is completed.
Claims (5)
1. A bearing health state monitoring method based on a convolution self-encoder is characterized in that firstly, a vibration sensor is arranged on a brand-new bearing, and a digital vibration signal of the bearing in a full life cycle and a corresponding health state mark value are obtained; two feature extraction operations are respectively carried out on the digital vibration signal: firstly, performing empirical mode decomposition on the digital vibration signal, selecting a plurality of intrinsic mode components and obtaining the statistic characteristics of each selected intrinsic mode component; secondly, inputting the digital vibration signal into a convolution self-encoder to carry out deep learning training, and obtaining the depth characteristic of the input signal after the training is finished; then splicing and screening the two characteristics, inputting the screened characteristics into a fully-connected regression network for regression training, and finally obtaining a health state curve graph of the bearing in the full life cycle; acquiring digital vibration signals of bearings to be monitored in the same model, acquiring a health state predicted value of the bearing by using a trained convolution self-encoder and a fully-connected regression network, and further generating a current health state curve graph of the bearing; obtaining a health state monitoring result of the bearing to be monitored by comparing the two graphs; the method comprises the following steps:
1) selecting a brand-new bearing, and acquiring a digital vibration signal and a corresponding health state mark value of the bearing in the whole life cycle; the method comprises the following specific steps:
1-1) selecting a brand new bearing;
1-2) installing a vibration sensor on a bearing seat of the bearing, acquiring vibration signals of the bearing in a full life cycle from the beginning of operation to scrapping by using the vibration sensor, and converting the acquired vibration signals into digital vibration signals;
1-3) carrying out preprocessing operation on the digital vibration signals, wherein the preprocessing operation comprises sampling the digital vibration signals obtained in the step 1-2), obtaining the digital vibration signals of all the preprocessed sampling points and calculating a health state marking value corresponding to each sampling point;
2) performing empirical mode decomposition on the digital vibration signals of all the sampling points after the pretreatment obtained in the step 1) to obtain corresponding intrinsic mode components, and then selecting the first N intrinsic mode components to perform statistic feature analysis to obtain statistic features corresponding to each selected intrinsic mode component;
3) inputting the digital vibration signals of all the sampling points preprocessed in the step 1) into a one-dimensional convolution self-encoder with random initialization parameters, carrying out deep learning training on the convolution self-encoder by reducing reconstruction errors, and obtaining the trained convolution self-encoder when the reconstruction errors are not reduced; the convolution self-encoder is composed of an encoder and a decoder;
inputting the digital vibration signals of all the sampling points preprocessed in the step 1) into a trained convolution self-encoder, and outputting depth characteristics corresponding to the input signals by the last layer of the encoder in the convolution self-encoder;
4) splicing all the characteristics obtained in the step 2) and the step 3) according to corresponding sampling points to obtain splicing characteristics of the digital vibration signals of the sampling points after pretreatment, and sequencing and screening the splicing characteristics in sequence by using monotonicity, trend and autocorrelation to obtain screened characteristics;
5) inputting the screened features in the step 4) into a full-connection regression network for regression training, and repeatedly reducing the value of the loss function by adopting a gradient descent method to train the full-connection regression network until the value of the loss function is not reduced any more, and finishing the training of the full-connection regression network; taking the predicted value of the health state index of the bearing at each sampling point output by the trained fully-connected regression network as the health state value of the bearing at the sampling point;
wherein, the calculation expression of the loss function is:
in the formula, labeliIndicating the health status of the ith sample point, predictioniIndicates the currentThe predicted value of the health state of the ith sampling point output by the fully-connected neural network;
utilizing the health state value of the bearing at each sampling point to draw a two-dimensional curve of the change of the health state of the bearing along with time to obtain a health state curve graph of the bearing in the whole life cycle; dividing the whole life cycle of the bearing into a stable stage, a degradation stage and a rapid degradation stage according to the change trend of the slope in the health state curve graph;
6) obtaining a bearing to be monitored with the same model as the model in the step 1), and obtaining a health state monitoring result of the bearing; the method comprises the following specific steps:
6-1) obtaining a new bearing with the same type as the bearing in the step 1) as a bearing to be monitored, and mounting a vibration sensor with the same type as the vibration sensor in the step 1-2) on a bearing seat of the bearing;
6-2) at any monitoring moment, acquiring a vibration signal of the bearing to be monitored from the beginning of operation to the monitoring moment by using the vibration sensor installed in the step 6-1), and converting the vibration signal into a digital vibration signal; preprocessing the digital vibration signal, and sampling the digital vibration signal to obtain digital vibration signals of all preprocessed sampling points;
6-3) carrying out empirical mode decomposition on the digital vibration signals of all the sampling points preprocessed in the step 6-2) to obtain the first N intrinsic mode components; performing statistic feature analysis on each intrinsic mode component to obtain statistic features corresponding to each intrinsic mode component;
6-4) inputting the digital vibration signals of all sampling points preprocessed in the step 6-2) into the trained convolution self-encoder in the step 3), and outputting depth characteristics corresponding to the input signals from the last layer of the encoder in the convolution self-encoder;
6-5) splicing all the characteristics obtained in the step 6-3) and the step 6-4) according to corresponding sampling points to obtain spliced characteristics, and sequencing and screening the spliced characteristics in sequence by using monotonicity, trend and autocorrelation to obtain screened characteristics;
6-6) inputting the characteristics screened in the step 6-5) into the fully-connected regression network trained in the step 5), and outputting the predicted value of the health state index of the bearing to be monitored at each sampling point as the health state value of the bearing at the sampling point by the network;
obtaining a health state curve graph of the bearing to be monitored from the beginning to the current moment by using the health state value of each sampling point;
6-7) comparing the health state curve graph obtained in the step 6-6) with the health state curve graph of the bearing obtained in the step 5) in the full life cycle to obtain the health state stage of the bearing to be monitored at the monitoring time, and finishing the monitoring.
2. The method of claim 1, wherein the vibration sensor is any one of a horizontal vibration sensor or a vertical vibration sensor.
3. The method according to claim 1, wherein the health status label value corresponding to each sampling point in the steps 1-3) is calculated as follows:
the total number of sampling points is recorded as n, a health state marking value corresponding to each sampling point is calculated by adopting an inverse hyperbolic tangent function, and the expression of the inverse hyperbolic tangent function is as follows:
wherein t isiThe time corresponding to the ith sampling point is shown, t represents the total service life duration of the bearing, and yiIs tiAnd marking the health state of the bearing at the moment.
4. The method according to claim 1, wherein the statistical features corresponding to each eigenmode component in the steps 2) and 6-3) are calculated as follows:
let the number of sampling points be n, xiRepresenting the corresponding value of the ith sampling point in any intrinsic mode component in the intrinsic mode component, then:
the mean calculation expression for each eigenmode component is:
the variance calculation expression for each eigenmode component is:
the kurtosis calculation expression for each eigenmode component is:
the root mean square calculation expression of each eigenmode component is:
the energy calculation expression for each eigenmode component is:
5. the method as claimed in claim 1, wherein the trend, monotonicity and autocorrelation calculation method in the steps 4) and 6-5) is as follows:
the monotonicity calculation expression of each feature in the splicing features is as follows:
wherein z represents any of the stitching features, ziThe ith one showing the characteristicsValues of sampling points, d/dz ═ zi+1-ziA differential value representing a state of health index; no. of d/dz>0 and No. of d/dz<0 represents a positive differential value and a negative differential value, respectively;
the trend calculation expression is:
in the formula, tiIndicating the time instant to which the ith value in each feature corresponds,
representing the mean value of all the moments corresponding to the features;
the autocorrelation calculation expression is:
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110328161.3A CN113158814B (en) | 2021-03-26 | 2021-03-26 | Bearing health state monitoring method based on convolution self-encoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110328161.3A CN113158814B (en) | 2021-03-26 | 2021-03-26 | Bearing health state monitoring method based on convolution self-encoder |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113158814A CN113158814A (en) | 2021-07-23 |
| CN113158814B true CN113158814B (en) | 2022-06-03 |
Family
ID=76885624
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110328161.3A Active CN113158814B (en) | 2021-03-26 | 2021-03-26 | Bearing health state monitoring method based on convolution self-encoder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113158814B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113872024A (en) * | 2021-12-01 | 2021-12-31 | 中国工程物理研究院电子工程研究所 | Intelligent fault diagnosis method for multi-source physical monitoring quantity of optical fiber laser system |
| CN114970598B (en) * | 2022-02-24 | 2024-04-30 | 清华大学 | Mechanical health state monitoring method and device |
| DE102022113888A1 (en) * | 2022-06-01 | 2023-12-07 | Rheinisch-Westfälische Technische Hochschule (RWTH) Aachen, Körperschaft des öffentlichen Rechts | Method and system for monitoring a plain bearing |
| CN115185313B (en) * | 2022-08-05 | 2024-08-16 | 五凌电力有限公司 | Trend tracking and early warning method and device for bearing bush temperature of hydroelectric generating set |
| CN115310490B (en) * | 2022-08-17 | 2024-03-29 | 中国核动力研究设计院 | Rotary equipment fault analysis method based on multi-domain feature and sensitive feature selection |
| CN115993247B (en) * | 2022-12-08 | 2023-09-01 | 中国矿业大学 | Drilling machine spindle bearing health state assessment method based on time sequence decomposition and order preserving regression |
| CN116226646B (en) * | 2023-05-05 | 2023-07-21 | 国家石油天然气管网集团有限公司 | Method, system, equipment and medium for predicting health state and residual life of bearing |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109947088A (en) * | 2019-04-17 | 2019-06-28 | 北京天泽智云科技有限公司 | Equipment fault early-warning system based on model lifecycle management |
| CN111289250A (en) * | 2020-02-24 | 2020-06-16 | 湖南大学 | Method for predicting residual service life of rolling bearing of servo motor |
| CN112036042A (en) * | 2020-09-02 | 2020-12-04 | 哈尔滨工程大学 | Power equipment abnormality detection method and system based on variational modal decomposition |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7260501B2 (en) * | 2004-04-21 | 2007-08-21 | University Of Connecticut | Intelligent model-based diagnostics for system monitoring, diagnosis and maintenance |
-
2021
- 2021-03-26 CN CN202110328161.3A patent/CN113158814B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109947088A (en) * | 2019-04-17 | 2019-06-28 | 北京天泽智云科技有限公司 | Equipment fault early-warning system based on model lifecycle management |
| CN111289250A (en) * | 2020-02-24 | 2020-06-16 | 湖南大学 | Method for predicting residual service life of rolling bearing of servo motor |
| CN112036042A (en) * | 2020-09-02 | 2020-12-04 | 哈尔滨工程大学 | Power equipment abnormality detection method and system based on variational modal decomposition |
Non-Patent Citations (3)
| Title |
|---|
| 基于Internet的轴承运行状态远程监测系统设计;朱耀楚等;《电子科技》;20180515(第05期);全文 * |
| 基于卷积自编码器的旋转机械故障特征提取方法研究;申彦斌;《南方农机》;20200215(第03期);全文 * |
| 基于多模态堆叠自动编码器的感应电机故障诊断;蒋爱国等;《电子测量与仪器学报》;20180815(第08期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113158814A (en) | 2021-07-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113158814B (en) | Bearing health state monitoring method based on convolution self-encoder | |
| CN111325095B (en) | Intelligent detection method and system for equipment health state based on acoustic wave signals | |
| CN111914883B (en) | Spindle bearing state evaluation method and device based on deep fusion network | |
| Han et al. | Intelligent fault diagnosis method for rotating machinery via dictionary learning and sparse representation-based classification | |
| CN112067916A (en) | Time series data intelligent fault diagnosis method based on deep learning | |
| CN110596506A (en) | Converter fault diagnosis method based on time convolution network | |
| CN111650453A (en) | Power equipment diagnosis method and system based on windowing characteristic Hilbert imaging | |
| CN110059845B (en) | Metering device clock error trend prediction method based on time sequence evolution gene model | |
| CN113505655A (en) | Bearing fault intelligent diagnosis method for digital twin system | |
| CN112036042A (en) | Power equipment abnormality detection method and system based on variational modal decomposition | |
| CN116593157A (en) | Complex working condition gear fault diagnosis method based on matching element learning under small sample | |
| CN112949402A (en) | Fault diagnosis method for planetary gear box under minimum fault sample size | |
| CN116977708B (en) | Bearing intelligent diagnosis method and system based on self-adaptive aggregation visual view | |
| CN112541510A (en) | Intelligent fault diagnosis method based on multi-channel time series data | |
| CN115034137A (en) | RVM and degradation model-based two-stage hybrid prediction method for residual life of bearing | |
| CN114216682A (en) | Service life prediction method and device of rolling bearing based on TCN and BLS | |
| CN114169377A (en) | G-MSCNN-based fault diagnosis method for rolling bearing in noisy environment | |
| CN114819315A (en) | Bearing degradation trend prediction method based on multi-parameter fusion health factor and time convolution neural network | |
| CN113705405B (en) | Nuclear pipeline fault diagnosis method | |
| CN112560674B (en) | Method and system for detecting sound signal quality | |
| CN112016800B (en) | Feature selection method and system based on effectiveness index | |
| CN111539381B (en) | Construction method of wind turbine bearing fault classification diagnosis model | |
| CN112613494A (en) | Power line monitoring abnormity identification method and system based on deep countermeasure network | |
| CN116776284A (en) | Fault diagnosis method for electromechanical device, computer device, and storage medium | |
| CN114397521A (en) | Fault diagnosis method and system for electronic equipment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |