CN109992872B - Mechanical equipment residual life prediction method based on stacked separation convolution module - Google Patents

Mechanical equipment residual life prediction method based on stacked separation convolution module Download PDF

Info

Publication number
CN109992872B
CN109992872B CN201910235692.0A CN201910235692A CN109992872B CN 109992872 B CN109992872 B CN 109992872B CN 201910235692 A CN201910235692 A CN 201910235692A CN 109992872 B CN109992872 B CN 109992872B
Authority
CN
China
Prior art keywords
mechanical equipment
convolution
data
residual life
expression
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
Application number
CN201910235692.0A
Other languages
Chinese (zh)
Other versions
CN109992872A (en
Inventor
雷亚国
姜鑫伟
王彪
李乃鹏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian Jiaotong University
Original Assignee
Xian Jiaotong University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xian Jiaotong University filed Critical Xian Jiaotong University
Priority to CN201910235692.0A priority Critical patent/CN109992872B/en
Publication of CN109992872A publication Critical patent/CN109992872A/en
Application granted granted Critical
Publication of CN109992872B publication Critical patent/CN109992872B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G06F30/17—Mechanical parametric or variational design
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/04—Ageing analysis or optimisation against ageing
    • 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

Landscapes

  • Physics & Mathematics (AREA)
  • Geometry (AREA)
  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Optimization (AREA)
  • Mathematical Analysis (AREA)
  • Computer Hardware Design (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Mathematics (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

A method for predicting the residual life of mechanical equipment based on a stacked separation convolution module comprises the steps of firstly obtaining original vibration signals of the mechanical equipment under different working conditions, preprocessing the signals and establishing a residual life prediction model of the mechanical equipment based on the stacked separation convolution module; extracting high-level typical characteristics most relevant to the health condition of mechanical equipment in the original vibration signals by using a laminated separation convolution module; then inputting the high-level typical characteristics into a full-connection network to obtain a predicted value of the residual life of the mechanical equipment; constructing a mean square error objective function of the residual life prediction model, and iteratively updating parameters to be trained of the prediction model through an Adam optimization algorithm to obtain an optimal residual life prediction model; finally inputting the preprocessed mechanical equipment vibration signal to complete the prediction of the residual life of the mechanical equipment; the method has the advantages of higher prediction precision, better stability and stronger robustness.

Description

Mechanical equipment residual life prediction method based on stacked separation convolution module
Technical Field
The invention belongs to the technical field of residual life prediction of mechanical equipment, and particularly relates to a residual life prediction method of the mechanical equipment based on a stacked separation convolution module.
Background
With the increase of informatization level and industrial automation level, modern mechanical equipment is developing towards high performance, high precision and high efficiency. Because the equipment often works under more complicated working conditions, the failure occurrence rate of mechanical equipment is higher, and the equipment is difficult to operate safely and reliably. Therefore, the residual life of the mechanical equipment needs to be predicted, so that an effective early maintenance scheme is formulated, and the normal operation of the equipment is ensured. Due to the reasons of large number of monitoring points of the equipment, high sampling frequency, long data collection duration and the like, the mechanical health monitoring enters a big data era, and a serious challenge is brought to the residual life prediction of mechanical equipment. Therefore, a new method for predicting the residual life of the data-driven mechanical equipment needs to be invented to ensure the safe service of the mechanical equipment.
By introducing a deep learning theory, the method for predicting the residual service life of the mechanical equipment based on data driving can effectively mine degradation information of the mechanical equipment from monitoring data, and a good prediction result is obtained. However, the current method for predicting the remaining life of mechanical equipment based on data driving needs to manually extract features from vibration signals, and the process not only needs professional signal processing knowledge and rich expert experience, but also needs a lot of labor cost. Meanwhile, the monitoring data of the mechanical equipment is acquired through a plurality of sensors, and the data of different sensors contains the degradation information of the mechanical equipment with different degrees, so that the interaction mechanism among different component faults is reflected. However, the current method ignores the correlation between different sensor data, so that the most relevant information to the health condition of the mechanical equipment cannot be effectively extracted, and the accuracy of the residual life prediction result of the mechanical equipment under the background of big mechanical data is seriously influenced.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention aims to provide a method for predicting the residual life of mechanical equipment based on a stacked separation convolution module, which directly excavates information most relevant to the health condition of the mechanical equipment from a vibration signal of the mechanical equipment, automatically extracts high-level typical characteristics and predicts the residual life of the mechanical equipment by using the high-level typical characteristics; the method has the advantages of higher prediction precision, better stability and stronger robustness.
In order to achieve the purpose, the invention adopts the technical scheme that:
a method for predicting the residual life of mechanical equipment based on a stacked separation convolution module comprises the following steps:
1) obtaining original vibration signals of mechanical equipment under different working conditions
Figure BDA0002008111500000021
Wherein the content of the first and second substances,
Figure BDA0002008111500000022
m is the number of signal samples, N is the number of data points contained in each vibration signal sample, and C is the number of vibration sensors; meanwhile, constructing a prediction model of a laminated separation convolution module with a deep structure;
2) for original vibration signal
Figure BDA0002008111500000023
Carrying out pretreatment: firstly, to the vibration signal xiPerforming Z-score normalization processing, and then embedding x by a time window embedding strategyiEmbedding previous validity time information data into xiThe method comprises the following specific steps:
2.1) the vibration signal x is corrected by Z-scoreiCarrying out normalization processing to obtain monitoring data
Figure BDA0002008111500000028
The Z-score normalized operation expression is shown below:
Figure BDA0002008111500000024
in the formula, x is original data;
Figure BDA0002008111500000025
is the mean of the original data; sigma is the standard deviation of the original data; x is the number of*Data normalized for Z-score;
2.2) setting the time window size S, integrating data by time window embedding strategy
Figure BDA0002008111500000026
And S-1 vibration signal samples before the vibration signal sample to obtain data
Figure BDA0002008111500000027
Namely, it is
Figure BDA0002008111500000031
3) Stacking T separate convolution modules from the preprocessed original vibration signal
Figure BDA0002008111500000032
The method comprises the following steps of extracting high-level typical characteristics most relevant to the health condition of mechanical equipment:
3.1) input samples
Figure BDA0002008111500000033
Respectively using channel convolution kernel kcwAnd a point convolution kernel kpwAnd input samples
Figure BDA0002008111500000034
Performing convolution to obtain characteristics
Figure BDA0002008111500000035
3.2) features of the pairs
Figure BDA0002008111500000036
Performing average pooling to obtain average value of elements in non-overlapping pooling region, and obtaining pooled features
Figure BDA0002008111500000037
3.3) characterization of
Figure BDA0002008111500000038
Inputting the data into T stacked separate convolution modules with residual connection, and extracting high-level typical features most relevant to the health condition of mechanical equipment, wherein each separate convolution module is specifically realized in the following way:
(a) firstly, processing the input of a separation convolution module by using a pre-activation strategy, and sequentially passing through a batch normalization layer and a linear rectification function (Re L U) layer;
the batch normalization layer operation expression is as follows:
Figure BDA0002008111500000039
Figure BDA00020081115000000310
in the formula, xl-1Is the input of the separation convolution module; y isl-1The output is the output after batch normalization; mu.sB、
Figure BDA00020081115000000311
Are respectively input xl-1γ, β are reconstruction parameters that the normalization layer can learn;
(b) inputting the data subjected to the pre-activation operation in the step (a) into a separation convolution layer, and extracting deep expression characteristics of the data
Figure BDA00020081115000000312
The separation convolution includes convolution in the channel direction and convolution in the point direction, and the calculation expression is as follows:
Figure BDA0002008111500000041
Figure BDA0002008111500000042
in the formula (I), the compound is shown in the specification,
Figure BDA0002008111500000043
the data after the pre-activation operation is obtained;
Figure BDA0002008111500000044
is the output of the convolution in the channel direction;
Figure BDA0002008111500000045
is the output of the convolution in the direction of the point; k is a radical ofcw、
Figure BDA0002008111500000046
Represents the convolution kernel and bias in the channel direction; k is a radical ofpw、
Figure BDA0002008111500000047
Represents the convolution kernel and bias in the direction of the point; c represents the input channel of the c convolution operation; n represents the output channel of the nth convolution operation;
(c) performing steps (a) and (b) again, andone-step extraction of deep expression characteristics of data
Figure BDA0002008111500000048
Learning correlations between different sensor data;
(d) for the extracted deep expression characteristics
Figure BDA0002008111500000049
Performing average pooling to obtain average value of elements in non-overlapping pooling region, and obtaining pooled features
Figure BDA00020081115000000410
(e) Characteristics after pooling
Figure BDA00020081115000000411
Inputting the data into a characteristic calibration layer, and performing characteristic response recalibration through compression operation and self-calibration operation to obtain characteristics most relevant to the health state of the mechanical equipment;
the compression operation refers to global average pooling, and the computational expression is as follows:
Figure BDA00020081115000000412
in the formula, H represents the global length of the input of the compressed excitation layer;
Figure BDA00020081115000000413
an input representing a compressed excitation layer;
the self-calibration operation refers to estimating the information quantity of each channel by using an adaptive door mechanism, generating the weight of the corresponding channel, and calculating the expression as follows:
Figure BDA00020081115000000414
wherein, sigma (·) and (-) are respectively Sigmoid and Re L U activation functions;
Figure BDA00020081115000000415
Figure BDA00020081115000000416
wherein r is the ratio of dimensionality reduction, and C is the number of channels;
(f) finally, the separation convolution module adopts residual connection, and the output calculation expression is as follows:
xl＝xl-1+F(xl-1,Wl)
in the formula, xl-1Is the input of the l-th layer, xlIs the output of the l-th layer; f (-) is a residual function, and the expression is as follows:
F(xl-1):＝H(xl-1)-xl-1
in the formula, H (x)l-1) Is a desired mapping;
4) inputting the extracted high-level typical features into a full-connection network to obtain a service life predicted value of the mechanical equipment, wherein the specific steps are as follows;
4.1) performing global average pooling on the extracted typical features to ensure that the output feature maps only contain one element, and obtaining the pooled features
Figure BDA0002008111500000051
4.2) post-pooling feature
Figure BDA0002008111500000052
Lay flat into one-dimensional vector
Figure BDA0002008111500000053
Computing a remaining life of a mechanical device preRU L using a fully connected networki；
5) Based on Adam optimization algorithm, repeating steps 3) and 4), setting iteration times N, and iteratively updating parameters of the stacked separation convolution module and the fully-connected network to obtain an optimal residual life prediction model, namely a minimum mean square error objective function:
Figure BDA0002008111500000054
in the formula, yiThe actual residual life value of the mechanical equipment;
6) the preprocessed mechanical equipment vibration signal is transmitted
Figure BDA0002008111500000055
And inputting the residual life into an optimal residual life prediction model to predict the residual life of the mechanical equipment.
The invention has the beneficial effects that:
the method can directly extract high-level typical characteristics from the original vibration signals, and considers the correlation of different sensor data in typical characteristic learning, so that the degradation information most relevant to the health condition of mechanical equipment can be accurately mined. The method overcomes the defects that the traditional method excessively depends on expert experience and domain knowledge and can not effectively extract the information most relevant to the health condition of the mechanical equipment, and realizes accurate prediction of the residual life of the mechanical equipment under the condition of large mechanical data.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Fig. 2 is a diagram showing the basic components of the separation convolution module.
FIG. 3 is a diagram illustrating the predicted residual life of the rolling bearing under three operating conditions of the embodiment.
FIG. 4 is a graph showing a comparison of the predicted residual life of the rolling bearing of the embodiment.
Detailed Description
The invention is explained in further detail below with reference to the drawings.
Referring to fig. 1, a method for predicting the remaining life of mechanical equipment based on a stacked separation convolution module includes the following steps:
1) obtaining original vibration signals of mechanical equipment under different working conditions
Figure BDA0002008111500000061
Wherein the content of the first and second substances,
Figure BDA0002008111500000062
m is the number of signal samples, N is the number of data points contained in each vibration signal sample, and C is the number of vibration sensors; meanwhile, constructing a prediction model of a laminated separation convolution module with a deep structure;
2) for original vibration signal
Figure BDA0002008111500000063
Carrying out pretreatment: firstly, to the vibration signal xiPerforming Z-score normalization processing, and then embedding x by a time window embedding strategyiEmbedding previous validity time information data into xiThe method comprises the following specific steps:
2.1) the vibration signal x is corrected by Z-scoreiCarrying out normalization processing to obtain monitoring data
Figure BDA0002008111500000064
The Z-score normalized operation expression is shown below:
Figure BDA0002008111500000065
in the formula, x is original data;
Figure BDA0002008111500000066
is the mean of the original data; sigma is the standard deviation of the original data; x is the number of*Data normalized for Z-score;
2.2) setting the time window size S, integrating data by time window embedding strategy
Figure BDA0002008111500000071
And S-1 vibration signal samples before the vibration signal sample to obtain data
Figure BDA0002008111500000072
Namely, it is
Figure BDA0002008111500000073
3) Stacking T separate convolution modes as shown in FIG. 2Block from preprocessed original vibration signal
Figure BDA0002008111500000074
The method comprises the following steps of extracting high-level typical characteristics most relevant to the health condition of mechanical equipment:
3.1) input samples
Figure BDA0002008111500000075
Respectively using channel convolution kernel kcwAnd a point convolution kernel kpwAnd input samples
Figure BDA0002008111500000076
Performing convolution to obtain characteristics
Figure BDA0002008111500000077
The convolution operation includes convolution in the channel direction and convolution in the point direction, features
Figure BDA0002008111500000078
Can be calculated from the following convolution expression:
yc＝kcw*xc+bc
Figure BDA0002008111500000079
in the formula, xcThe vibration signal data is preprocessed; y iscIs the output of the convolution in the channel direction; z is a radical ofnIs the output of the convolution in the direction of the point; k is a radical ofcw、bcRepresents the convolution kernel and bias in the channel direction; k is a radical ofpw、bnRepresents the convolution kernel and bias in the direction of the point; c represents the input channel of the c convolution operation; n represents the output channel of the nth convolution operation;
3.2) features of the pairs
Figure BDA00020081115000000710
Carrying out average pooling to obtain the average value of elements in the non-overlapping pooling regionThe characteristics after the pond are that
Figure BDA00020081115000000711
The pooled features may be calculated by the following expression:
xAP＝pool(xcp,p,s)
in the formula, xcpRepresenting the features after convolution; x is the number ofAPRepresenting pooled features; pool (·) represents a downsampling function; p, s are respectively the size of the pooling and the sliding step length;
3.3) characterization of
Figure BDA00020081115000000712
Inputting the data into T stacked separate convolution modules with residual connection, and extracting high-level typical features most relevant to the health condition of mechanical equipment, wherein each separate convolution module is specifically realized in the following way:
(a) firstly, processing the input of a separation convolution module by using a pre-activation strategy, and sequentially passing through a batch normalization layer and a linear rectification function (Re L U) layer;
the batch normalization layer operation expression is as follows:
Figure BDA0002008111500000081
Figure BDA0002008111500000082
in the formula, xl-1Is the input of the separation convolution module; y isl-1The output is the output after batch normalization; mu.sB、
Figure BDA0002008111500000083
Are respectively input xl-1γ, β are reconstruction parameters that the normalization layer can learn;
(b) inputting the data subjected to the pre-activation operation in the step (a) into a separation convolution layer, and extracting deep expression characteristics of the data
Figure BDA0002008111500000084
The separation convolution includes convolution in the channel direction and convolution in the point direction, and the calculation expression is as follows:
Figure BDA0002008111500000085
Figure BDA0002008111500000086
in the formula (I), the compound is shown in the specification,
Figure BDA0002008111500000087
the data after the pre-activation operation is obtained;
Figure BDA0002008111500000088
is the output of the convolution in the channel direction;
Figure BDA0002008111500000089
is the output of the convolution in the direction of the point; k is a radical ofcw、
Figure BDA00020081115000000810
Represents the convolution kernel and bias in the channel direction; k is a radical ofpw、
Figure BDA00020081115000000811
Represents the convolution kernel and bias in the direction of the point; c represents the input channel of the c convolution operation; n represents the output channel of the nth convolution operation;
(c) performing the steps (a) and (b) again, and further extracting deep expression characteristics of the data
Figure BDA00020081115000000812
Learning correlations between different sensor data;
(d) for the extracted deep expression characteristics
Figure BDA0002008111500000091
Performing average pooling to obtain average value of elements in non-overlapping pooling region, and obtaining pooled features
Figure BDA0002008111500000092
(e) Characteristics after pooling
Figure BDA0002008111500000093
Inputting the data into a characteristic calibration layer, and performing characteristic response recalibration through compression operation and self-calibration operation to obtain characteristics most relevant to the health state of the mechanical equipment;
the compression operation refers to global average pooling, and the computational expression is as follows:
Figure BDA0002008111500000094
in the formula, H represents the global length of the input of the compressed excitation layer;
Figure BDA0002008111500000095
an input representing a compressed excitation layer;
the self-calibration operation refers to estimating the information quantity of each channel by using an adaptive door mechanism, generating the weight of the corresponding channel, and calculating the expression as follows:
Figure BDA0002008111500000096
wherein, sigma (·) and (-) are respectively Sigmoid and Re L U activation functions;
Figure BDA0002008111500000097
Figure BDA0002008111500000098
wherein r is the ratio of dimensionality reduction, and C is the number of channels;
(f) finally, the separation convolution module adopts residual connection, and the output calculation expression is as follows:
xl＝xl-1+F(xl-1,Wl)
in the formula, xl-1Is the input of the l-th layer, xlIs the output of the l-th layer; f (-) is a residual function, and the expression is as follows:
F(xl-1):＝H(xl-1)-xl-1
in the formula, H (x)l-1) Is a desired mapping;
4) inputting the extracted high-level typical features into a full-connection network to obtain a service life predicted value of the mechanical equipment, wherein the specific steps are as follows;
4.1) performing global average pooling on the extracted typical features to ensure that the output feature maps only contain one element, and obtaining the pooled features
Figure BDA0002008111500000101
4.2) post-pooling feature
Figure BDA0002008111500000102
Lay flat into one-dimensional vector
Figure BDA0002008111500000103
Computing a remaining life of a mechanical device preRU L using a fully connected networki；
5) Based on Adam optimization algorithm, repeating steps 3) and 4), setting iteration times N, and iteratively updating parameters of the stacked separation convolution module and the fully-connected network to obtain an optimal residual life prediction model, namely a minimum mean square error objective function:
Figure BDA0002008111500000104
in the formula, yiThe actual residual life value of the mechanical equipment;
6) the preprocessed mechanical equipment vibration signal is transmitted
Figure BDA0002008111500000105
Inputting the data into an optimal residual life prediction modelAnd predicting the residual life of the mechanical equipment.
Example (b): the effectiveness of the method is verified by taking a rolling bearing in mechanical equipment as a case and based on the experimental data of the accelerated life of the rolling bearing. The data set of the acceleration life experiment of the rolling bearing adopted in the embodiment comprises 3 subsets which respectively correspond to three different working conditions, namely 12kN/2100rpm, 11kN/2250rpm and 10kN/2400 rpm. Wherein, every operating mode contains 5 antifriction bearing's full life cycle vibration signal. As shown in table 1, when the method of the present invention is used to predict the remaining life of a rolling bearing, the first 4 rolling bearing data under each working condition are used as a training data set, and the last 1 rolling bearing data are used as a test data set.
The parameter settings of the prediction model of the stacked separation convolution module are as follows: the number T of the separation convolution modules is 3; the time window S is 5; the size of the convolution kernel in the channel direction is 8, and the number of the convolution kernels is 16; the size and the step length of the pooling area are set to be 4; the dimensionality reduction ratio is set to 16; 3 layers of separation convolution modules are adopted for stacking; the number of small batch trains is 128; the number of iterations is chosen to be 100. The method is used for predicting the residual life of the rolling bearing test data set under 3 working conditions, the prediction result is shown in figure 3, and as can be seen from figure 3, although the deviation between the real life and the predicted life of the rolling bearing at the early stage is large, the predicted life of the rolling bearing gradually approaches to the real life along with the time, which shows that the method can effectively predict the residual life of the rolling bearing. In order to further verify the superiority of the method, the method is compared with a residual life prediction method (DBN) based on a deep confidence network and a residual life prediction Method (MCNN) based on a multi-scale convolutional neural network, evaluation is carried out on the three methods by using a scoring function and a root mean square error prediction performance index, and the result is shown in FIG. 4. As can be seen from FIG. 4, in the residual life prediction of three test bearings, two prediction performance index values of the method are smaller than those of the other two prediction methods, which shows that the residual life prediction precision of the method is higher, the stability is better and the robustness is stronger.
TABLE 1
Figure BDA0002008111500000111
According to the method, the laminated separation convolution module is utilized, high-level typical characteristics can be directly extracted from an original vibration signal, the correlation of different sensor data in typical characteristic learning is considered, then the degradation information most relevant to the health condition of mechanical equipment can be accurately mined, the precision of predicting the residual life of the mechanical equipment is effectively improved, and more excellent prediction performance is obtained.

Claims (1)

1. A method for predicting the residual life of mechanical equipment based on a stacked separation convolution module is characterized by comprising the following steps:
1) obtaining original vibration signals of mechanical equipment under different working conditions
Figure FDA0002008111490000011
Wherein the content of the first and second substances,
Figure FDA0002008111490000012
m is the number of signal samples, N is the number of data points contained in each vibration signal sample, and C is the number of vibration sensors; meanwhile, constructing a prediction model of a laminated separation convolution module with a deep structure;
2) for original vibration signal
Figure FDA0002008111490000013
Carrying out pretreatment: firstly, to the vibration signal xiPerforming Z-score normalization processing, and then embedding x by a time window embedding strategyiEmbedding previous validity time information data into xiThe method comprises the following specific steps:
2.1) the vibration signal x is corrected by Z-scoreiCarrying out normalization processing to obtain monitoring data
Figure FDA0002008111490000014
The Z-score normalized operation expression is shown below:
Figure FDA0002008111490000015
in the formula, x is original data;
Figure FDA0002008111490000016
is the mean of the original data; sigma is the standard deviation of the original data; x is the number of*Data normalized for Z-score;
2.2) setting the time window size S, integrating data by time window embedding strategy
Figure FDA0002008111490000017
And S-1 vibration signal samples before the vibration signal sample to obtain data
Figure FDA0002008111490000018
Namely, it is
Figure FDA0002008111490000019
3) Stacking T separate convolution modules from the preprocessed original vibration signal
Figure FDA00020081114900000110
The method comprises the following steps of extracting high-level typical characteristics most relevant to the health condition of mechanical equipment:
3.1) input samples
Figure FDA00020081114900000111
Respectively using channel convolution kernel kcwAnd a point convolution kernel kpwAnd input samples
Figure FDA00020081114900000112
Performing convolution to obtain characteristics
Figure FDA00020081114900000113
3.2) features of the pairs
Figure FDA0002008111490000021
Performing average pooling to obtain average value of elements in non-overlapping pooling region, and obtaining pooled features
Figure FDA0002008111490000022
3.3) characterization of
Figure FDA0002008111490000023
Inputting the data into T stacked separate convolution modules with residual connection, and extracting high-level typical features most relevant to the health condition of mechanical equipment, wherein each separate convolution module is specifically realized in the following way:
(a) firstly, processing the input of a separation convolution module by using a pre-activation strategy, and sequentially passing through a batch normalization layer and a linear rectification function (Re L U) layer;
the batch normalization layer operation expression is as follows:
Figure FDA0002008111490000024
Figure FDA0002008111490000025
in the formula, xl-1Is the input of the separation convolution module; y isl-1The output is the output after batch normalization; mu.sB、
Figure FDA0002008111490000026
Are respectively input xl-1γ, β are reconstruction parameters that the normalization layer can learn;
(b) subjecting the mixture of step (a) to preactivationInputting the operated data into the separated convolution layer, and extracting the deep expression characteristics of the data
Figure FDA0002008111490000027
The separation convolution includes convolution in the channel direction and convolution in the point direction, and the calculation expression is as follows:
Figure FDA0002008111490000028
Figure FDA0002008111490000029
in the formula (I), the compound is shown in the specification,
Figure FDA00020081114900000210
the data after the pre-activation operation is obtained;
Figure FDA00020081114900000211
is the output of the convolution in the channel direction;
Figure FDA00020081114900000212
is the output of the convolution in the direction of the point; k is a radical ofcw、
Figure FDA00020081114900000213
Represents the convolution kernel and bias in the channel direction; k is a radical ofpw、Represents the convolution kernel and bias in the direction of the point; c represents the input channel of the c convolution operation; n represents the output channel of the nth convolution operation;
(c) performing the steps (a) and (b) again, and further extracting deep expression characteristics of the data
Figure FDA0002008111490000031
Learning different sensingCorrelation between device data;
(d) for the extracted deep expression characteristics
Figure FDA0002008111490000032
Performing average pooling to obtain average value of elements in non-overlapping pooling region, and obtaining pooled features
Figure FDA0002008111490000033
(e) Characteristics after pooling
Figure FDA0002008111490000034
Inputting the data into a characteristic calibration layer, and performing characteristic response recalibration through compression operation and self-calibration operation to obtain characteristics most relevant to the health state of the mechanical equipment;
the compression operation refers to global average pooling, and the computational expression is as follows:
Figure FDA0002008111490000035
in the formula, H represents the global length of the input of the compressed excitation layer;
Figure FDA0002008111490000036
an input representing a compressed excitation layer;
the self-calibration operation refers to estimating the information quantity of each channel by using an adaptive door mechanism, generating the weight of the corresponding channel, and calculating the expression as follows:
Figure FDA0002008111490000037
wherein, sigma (·) and (-) are respectively Sigmoid and Re L U activation functions;
Figure FDA0002008111490000038
Figure FDA0002008111490000039
wherein r is the ratio of dimensionality reduction, and C is the number of channels;
(f) finally, the separation convolution module adopts residual connection, and the output calculation expression is as follows:
xl＝xl-1+F(xl-1,Wl)
in the formula, xl-1Is the input of the l-th layer, xlIs the output of the l-th layer; f (-) is a residual function, and the expression is as follows:
F(xl-1):＝H(xl-1)-xl-1
in the formula, H (x)l-1) Is a desired mapping;
4) inputting the extracted high-level typical features into a full-connection network to obtain a service life predicted value of the mechanical equipment, wherein the specific steps are as follows;
4.1) performing global average pooling on the extracted typical features to ensure that the output feature maps only contain one element, and obtaining the pooled features
Figure FDA0002008111490000041
4.2) post-pooling feature
Figure FDA0002008111490000042
Lay flat into one-dimensional vector
Figure FDA0002008111490000043
Computing a remaining life of a mechanical device preRU L using a fully connected networki；
5) Based on Adam optimization algorithm, repeating steps 3) and 4), setting iteration times N, and iteratively updating parameters of the stacked separation convolution module and the fully-connected network to obtain an optimal residual life prediction model, namely a minimum mean square error objective function:
Figure FDA0002008111490000044
in the formula，yiThe actual residual life value of the mechanical equipment;
6) the preprocessed mechanical equipment vibration signal is transmitted
Figure FDA0002008111490000045
And inputting the residual life into an optimal residual life prediction model to predict the residual life of the mechanical equipment.
CN201910235692.0A 2019-03-27 2019-03-27 Mechanical equipment residual life prediction method based on stacked separation convolution module Active CN109992872B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910235692.0A CN109992872B (en) 2019-03-27 2019-03-27 Mechanical equipment residual life prediction method based on stacked separation convolution module

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910235692.0A CN109992872B (en) 2019-03-27 2019-03-27 Mechanical equipment residual life prediction method based on stacked separation convolution module

Publications (2)

Publication Number Publication Date
CN109992872A CN109992872A (en) 2019-07-09
CN109992872B true CN109992872B (en) 2020-07-28

Family

ID=67131710

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910235692.0A Active CN109992872B (en) 2019-03-27 2019-03-27 Mechanical equipment residual life prediction method based on stacked separation convolution module

Country Status (1)

Country Link
CN (1) CN109992872B (en)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110807257A (en) * 2019-11-04 2020-02-18 中国人民解放军国防科技大学 Method for predicting residual life of aircraft engine
CN111783362B (en) * 2020-07-09 2022-07-26 哈尔滨工程大学 Method and system for determining residual service life of electric gate valve
CN112881518B (en) * 2021-01-08 2022-07-12 东冶及策河北能源技术有限公司 Method for predicting residual life of dynamic filter compensator
CN113686577A (en) * 2021-08-17 2021-11-23 山东科技大学 Bearing fault diagnosis method based on rapid nonlinear sparse spectrum
CN115048873B (en) * 2022-08-12 2022-11-01 太原科技大学 Residual service life prediction system for aircraft engine
CN116579505B (en) * 2023-07-12 2023-10-13 中国科学院空间应用工程与技术中心 Electromechanical equipment cross-domain residual life prediction method and system without full life cycle sample
CN117010116A (en) * 2023-09-06 2023-11-07 长兴昇阳科技有限公司 Bearing residual service life assessment method based on elastic neuron colloid

Family Cites Families (8)

* Cited by examiner, † Cited by third party
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
CN103793752B (en) * 2013-09-13 2018-02-06 中国人民解放军第二炮兵工程大学 A kind of equipment failure number Forecasting Methodology based on modeling of degenerating
CN106769048B (en) * 2017-01-17 2019-05-24 苏州大学 Adaptive depth confidence network Method for Bearing Fault Diagnosis based on Nesterov momentum method
CN107451760B (en) * 2017-09-04 2020-06-26 西安交通大学 Rolling bearing fault diagnosis method based on time window slip limited Boltzmann machine
US10732618B2 (en) * 2017-09-15 2020-08-04 General Electric Company Machine health monitoring, failure detection and prediction using non-parametric data
CN108106830B (en) * 2017-12-13 2019-10-29 武汉科技大学 A kind of Variable Speed Rotating Machinery method for diagnosing faults based on time-frequency spectrum segmentation
CN109376401B (en) * 2018-09-29 2022-12-09 西安交通大学 Self-adaptive multi-source information optimization and fusion mechanical residual life prediction method
CN109460618B (en) * 2018-11-13 2023-02-10 华中科技大学 Rolling bearing residual life online prediction method and system

Also Published As

Publication number Publication date
CN109992872A (en) 2019-07-09

Similar Documents

Publication Publication Date Title
CN109992872B (en) Mechanical equipment residual life prediction method based on stacked separation convolution module
CN112149316B (en) Aero-engine residual life prediction method based on improved CNN model
CN112629863B (en) Bearing fault diagnosis method for dynamic joint distribution alignment network under variable working conditions
CN110779746B (en) Diagnosis method for improving composite fault of deep sparse self-encoder network rotating machinery
CN110108456A (en) A kind of rotating machinery health evaluating method of depth convolutional neural networks
CN112396109A (en) Motor bearing fault diagnosis method based on recursion graph and multilayer convolution neural network
CN109360120B (en) Intelligent diagnosis method for working condition of electric submersible pump well based on convolutional neural network
Roy et al. A stacked autoencoder neural network based automated feature extraction method for anomaly detection in on-line condition monitoring
CN113673346B (en) Motor vibration data processing and state identification method based on multiscale SE-Resnet
CN114218872B (en) DBN-LSTM semi-supervised joint model-based residual service life prediction method
CN111783362A (en) Method and system for determining residual service life of electric gate valve
CN113869563A (en) Method for predicting remaining life of aviation turbofan engine based on fault feature migration
CN115791174B (en) Rolling bearing abnormality diagnosis method, system, electronic equipment and storage medium
CN115290326A (en) Rolling bearing fault intelligent diagnosis method
CN114492642A (en) Mechanical fault online diagnosis method for multi-scale element depth residual shrinkage network
CN114021620A (en) Electrical submersible pump fault diagnosis method based on BP neural network feature extraction
CN115143128B (en) Fault diagnosis method and system for small-sized submersible electric pump
CN116244596A (en) Industrial time sequence data anomaly detection method based on TCN and attention mechanism
CN113052060B (en) Bearing residual life prediction method and device based on data enhancement and electronic equipment
CN112560252A (en) Prediction method for residual life of aircraft engine
CN113920124A (en) Brain neuron iterative segmentation method based on segmentation and error guidance
CN113159114A (en) High-dimensional data dimension reduction cross validation analysis method based on application in NIR data
CN117493980B (en) Bearing fault diagnosis method integrating feature extraction and sequencing causal discovery
CN115309736B (en) Time sequence data anomaly detection method based on self-supervision learning multi-head attention network
CN116756484A (en) Multi-signal fusion fault detection method, device and medium for cigarette 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