CN112597705A - Multi-feature health factor fusion method based on SCVNN - Google Patents

Multi-feature health factor fusion method based on SCVNN Download PDF

Info

Publication number
CN112597705A
CN112597705A CN202011589552.2A CN202011589552A CN112597705A CN 112597705 A CN112597705 A CN 112597705A CN 202011589552 A CN202011589552 A CN 202011589552A CN 112597705 A CN112597705 A CN 112597705A
Authority
CN
China
Prior art keywords
feature
scvnn
layer
hidden layer
fusion method
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.)
Granted
Application number
CN202011589552.2A
Other languages
Chinese (zh)
Other versions
CN112597705B (en
Inventor
杨京礼
高天宇
姜守达
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.)
Harbin Institute of Technology
Original Assignee
Harbin Institute of Technology
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 Harbin Institute of Technology filed Critical Harbin Institute of Technology
Priority to CN202011589552.2A priority Critical patent/CN112597705B/en
Publication of CN112597705A publication Critical patent/CN112597705A/en
Application granted granted Critical
Publication of CN112597705B publication Critical patent/CN112597705B/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]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • 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
    • G06N3/045—Combinations of networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

A multi-feature health factor fusion method based on SCVNN relates to the technical field of fault prediction, aims at the problem that in the prior art, a model cannot reduce the influence of experience factors and remove redundant information, and adopts the idea of variational inference to carry out normal distribution modeling on an original signal, so that self-adaptive feature learning can be carried out on the signal to construct features representing the essence of the signal. Compared with the traditional feature fusion method, the model can reduce the influence of empirical factors and remove redundant information. The self-normalization idea is introduced into the SCVNN model, so that the activation value can be transmitted among layers of the network in a normalized state, the over-fitting phenomenon is avoided, and the characteristics containing rich information are obtained, so that the health state of the rotary machine is better represented.

Description

Multi-feature health factor fusion method based on SCVNN
Technical Field
The invention relates to the technical field of failure prediction, in particular to a multi-feature health factor fusion method based on SCVNN.
Background
As a core component of a mechanical system, a rotary machine is widely used in various fields including aerospace, military, communication, industrial control, and the like, and a fault of the rotary machine affects system functions and even causes a catastrophic accident. Therefore, in order to ensure continuous and effective operation of an industrial mechanical system, the development of the intelligent rotary machine prediction method is a trend in the field of mechanical health monitoring, and the establishment of effective health factors is a prerequisite for realizing accurate prediction of rotary machine faults.
Self-normalization Neural Networks (SNN) adopt Scaled Exponential Linear Units (SELU) as an activation function, so that an activation value can be guaranteed to be transmitted among all layers of the network in a normalized state, the value tends to a stable fixed point, once disturbance causes covariate offset, the value can be pulled back to the normalized state immediately, and the over-fitting phenomenon is avoided. Furthermore, the function has no dead zone, i.e. when the input is less than 0, the neurons can still be activated, and compared with a model using a Linear rectification Unit (ReLU) as an activation function, the SNN extracts richer features, thereby better characterizing the health status of the system.
By adding a sampling layer in the multi-layer perceptron, a Conditional Variable Neural Network (CVNN) constructs hidden variables with the same distribution as the original input. And then solving parameters of the neural network model by using the thought of variational inference, and resampling the hidden variables to obtain the characteristics output by the sampling layer. The CVNN performs normal distribution modeling on the original signal, can perform adaptive feature learning, and obtains the essential features of the original input. However, the prior art lacks a method capable of ensuring that the activation value is transmitted between the layers of the network in a normalized state, so as to avoid the overfitting phenomenon.
Disclosure of Invention
The purpose of the invention is: aiming at the problem that the prior art can not ensure that the activation value is transmitted between network layers in a normalized state and avoid the phenomenon of over-fitting, the multi-feature health factor fusion method based on the SCVNN is provided.
The technical scheme adopted by the invention to solve the technical problems is as follows:
a multi-feature health factor fusion method based on SCVNN comprises the following steps:
the method comprises the following steps: collecting an original vibration signal of a rotary machine;
step two: carrying out smoothing and denoising pretreatment on an original vibration signal of the rotary machine, then extracting time domain, frequency domain and time-frequency domain characteristics of the pretreated original vibration signal of the rotary machine, constructing an original characteristic set, and then carrying out normalization treatment on the signal in the original characteristic set;
step three: constructing a sensitive feature set after screening by using the normalized original feature set;
step four: inputting the sensitive feature set into an SCVNN model for feature fusion training, and inputting data of the test set into the trained model to obtain a health factor representing the health state of the rotary machine;
the SCVNN model comprises an input layer, a first hidden layer, a sampling layer, a second hidden layer and an output layer;
the SCVNN model specifically executes the following steps:
(1) an input layer: taking a sensitive feature set obtained by screening evaluation indexes as an input sample of an input layer
Figure BDA0002866648280000021
Corresponding sample label is
Figure BDA0002866648280000022
Where M is the number of samples, xmRepresenting samples of a sample set, cmRepresenting a tag in a set of tags;
(2) first hidden layer: labeling based on sample
Figure BDA0002866648280000023
And function fθ1For the sample
Figure BDA0002866648280000024
Performing compression transformation to obtain output of the first hidden layer
Figure BDA0002866648280000025
(3) Sampling layer: outputting the first hidden layer
Figure BDA0002866648280000026
Feature vector of
Figure BDA0002866648280000027
Mapping into a feature compression layer to obtain a mean value mumSum variance σ2mThen random sampling is carried out on normal distribution to obtain a noise vector epsilonmAnd finally according to the mean value mumVariance σ2mSum noise vector εmSampling in a random sampling operation layer to obtain an implicit variable zm；
(4) Second hidden layer: passing function
Figure BDA0002866648280000028
For hidden variable zmPerforming compression transformation to obtain feature vector
Figure BDA0002866648280000029
(5) An output layer: according to the feature vector
Figure BDA00028666482800000210
Obtaining a feature vector
Figure BDA00028666482800000211
And outputting the corresponding health factor.
Further, the feature vector
Figure BDA00028666482800000212
Expressed as:
Figure BDA00028666482800000213
wherein theta is1＝{W1,b1Is the network parameter of the first hidden layer, W1As a weight of the first hidden layer, b1For the first hidden layer bias to be applied,
Figure BDA00028666482800000214
is an activation function of the first hidden layer, and the activation function of the first hidden layer is a SELU.
Further, the mean value μmExpressed as:
Figure BDA00028666482800000215
wherein theta is2＝{W2,b2Is the sampling layer mean parameter, W2Is the mean weight, b2Is the mean offset of the sampled layers.
Further, the variance σ2mExpressed as:
Figure BDA0002866648280000031
wherein theta is3＝{W3,b3Is the sampling layer variance parameter, W3As a variance weight, b3Is the variance bias of the sampling layer.
Further, the hidden variable zmExpressed as:
zm＝μm+σ2m×εm。
further, the feature vector
Figure BDA0002866648280000032
Expressed as:
Figure BDA0002866648280000033
wherein theta is4＝{W4,b4Is the parameter to be optimized of the second hidden layer, W4As a weight of the second hidden layer, b4For second hidden layer bias, zmIn order to be a hidden variable, the method comprises the following steps of,
Figure BDA0002866648280000034
is a second hidden layer activation function, which is SELU.
Further, the SELU is represented as:
Figure BDA0002866648280000035
wherein λ is 1.0507009873554804 and α is 1.6732632423543772.
Further, in the third step, the normalized original feature set is used for screening by using a correlation index, a monotonicity index and a robustness index as evaluation criteria of the feature quantity, wherein the correlation index is used for measuring the linear correlation degree between the feature parameter sequence and the time sequence; the monotonicity index is used for reflecting the monotonous ascending or descending conversion degree of the characteristic parameter sequence; the robustness index is used for describing the capacity of the characteristic parameter sequence to contain abnormal factors including noise interference.
Further, the specific steps of screening by using the normalized raw feature set in the third step are as follows:
let E ═ E (1), E (2), E, (k) be assumed as the characteristic quantity sequence]The time sequence is T ═ T1,t2,...,tK]，e(tk) Represents the time tkAnd (3) dividing the characteristic parameter sequence into a stationary trend term and a random residue term by adopting a moving average method according to the characteristic value, wherein K represents the total time length:
e(tk)＝eT(tk)+eR(tk)
wherein e isT(tk) Represents the steady trend portion of the feature, and eR(tk) Then the random margin portion of the feature is represented,
the correlation between E and T is marked as Corr (E, T), the monotonicity index and the robustness index of E are respectively marked as Mon (E) and Rob (E), and the calculation processes of three evaluation indexes are as follows:
Figure BDA0002866648280000041
Figure BDA0002866648280000042
Figure BDA0002866648280000043
where K represents the total time length and δ (-) represents a simple unit step function, which is expressed as follows:
Figure BDA0002866648280000044
further, the loss function of the SCVNN model is:
Figure BDA0002866648280000045
wherein the content of the first and second substances,
Figure BDA0002866648280000046
is the output of the SCVNN model; y istIs the true tag value.
The invention has the beneficial effects that:
(1) the SCVNN model adopts the idea of variational inference to carry out normal distribution modeling on an original signal, can carry out self-adaptive feature learning from the signal and constructs the feature representing the essence of the signal. Compared with the traditional feature fusion method, the model can reduce the influence of empirical factors and remove redundant information.
(2) The self-normalization idea is introduced into the SCVNN model, so that the activation value can be transmitted among layers of the network in a normalized state, the over-fitting phenomenon is avoided, and the characteristics containing rich information are obtained, so that the health state of the rotary machine is better represented.
Drawings
FIG. 1 is a schematic diagram of an SCVNN model structure;
fig. 2 is a schematic diagram of a SCVNN-based multi-feature health factor fusion process.
Detailed Description
It should be noted that, in the present invention, the embodiments disclosed in the present application may be combined with each other without conflict.
The first embodiment is as follows: referring to fig. 1 and fig. 2, the present embodiment is specifically described, and the SCVNN-based multi-feature health factor fusion method in the present embodiment includes the following steps:
the method comprises the following steps: collecting an original vibration signal of a rotary machine;
step two: carrying out smoothing and denoising pretreatment on an original vibration signal of the rotary machine, then extracting time domain, frequency domain and time-frequency domain characteristics of the pretreated original vibration signal of the rotary machine, constructing an original characteristic set, and then carrying out normalization treatment on the signal in the original characteristic set;
step three: constructing a sensitive feature set after screening by using the normalized original feature set;
step four: inputting the sensitive feature set into an SCVNN model for feature fusion training, and inputting data of the test set into the trained model to obtain a health factor representing the health state of the rotary machine;
the SCVNN model comprises an input layer, a first hidden layer, a sampling layer, a second hidden layer and an output layer;
the SCVNN model specifically executes the following steps:
(1) an input layer: taking a sensitive feature set obtained by screening evaluation indexes as an input sample of an input layer
Figure BDA0002866648280000051
Corresponding sample label is
Figure BDA0002866648280000052
Where M is the number of samples, xmRepresenting samples of a sample set, cmRepresenting a tag in a set of tags;
(2) first hidden layer: labeling based on sample
Figure BDA0002866648280000053
And function
Figure BDA0002866648280000054
For the sample
Figure BDA0002866648280000055
Performing compression transformation to obtain output of the first hidden layer
Figure BDA0002866648280000056
(3) Sampling layer: outputting the first hidden layer
Figure BDA0002866648280000057
Feature vector of
Figure BDA0002866648280000058
Mapping into a feature compression layer to obtain a mean value mumSum variance σ2mThen random sampling is carried out on normal distribution to obtain a noise vector epsilonmAnd finally according to the mean value mumVariance σ2mSum noise vector εmSampling in a random sampling operation layer to obtain an implicit variable zm；
(4) Second hidden layer: passing function
Figure BDA0002866648280000059
For hidden variable zmPerforming compression transformation to obtain feature vector
Figure BDA00028666482800000510
(5) An output layer: according to the feature vector
Figure BDA00028666482800000511
Obtaining a feature vector
Figure BDA00028666482800000512
And outputting the corresponding health factor.
The invention provides a multi-feature health factor fusion method based on a Self-normalization Conditional variable Neural network (SCVNN) by combining the characteristics of the SNN and the CVNN, which realizes the fusion of multi-feature quantities through a supervised learning method, establishes feature parameters capable of representing the degradation state of a mechanical system and further realizes the tracking of the degradation state of the rotary mechanical performance and the prediction of the residual life under multi-features.
The SCVNN model structure provided by the present invention is shown in fig. 1, and is mainly divided into an input layer, a first hidden layer, a sampling layer, a second hidden layer, and an output layer. The model is described in detail as follows:
1. input layer
Taking a sensitive feature set obtained by screening evaluation indexes as an input sample of an input layer
Figure BDA0002866648280000061
Corresponding sample label is
Figure BDA0002866648280000062
Where the number of samples is M.
2. First hidden layer
Passing function
Figure BDA0002866648280000063
Performing a compression transform on the samples to obtain an output of the first hidden layer
Figure BDA0002866648280000064
The formula for the neuron is as follows:
Figure BDA0002866648280000065
wherein theta is1＝{W1,b1Is the network parameter of the first hidden layer, which needs to be determined optimally, W1Is the weight of the first hidden layer, b1Is an offset.
Figure BDA0002866648280000066
Is the activation function of the first hidden layer. Selecting a SELU as an activation function of the first hidden layer, wherein the expression is as follows:
Figure BDA0002866648280000067
wherein λ 1.0507009873554804, α 1.6732632423543772. The SELU activation function has the advantages of high convergence speed and output approximate zero center, and solves the problems of gradient disappearance and neuron necrosis.
To improve the generalization capability of the model, the Dropout algorithm is used for the first layer hidden layer. The Dropout algorithm discards neurons in one layer of the neural network with a certain probability, thereby preventing adaptation between neurons in the same layer. Therefore, the Dropout algorithm makes the characteristic expression of the neuron more independent, and can improve the expression capability of the network.
3. Sampling layer
The sampling layer is composed of three sub-layers including a feature compression layer, a random sampling operation layer and a calculation layer. The feature compression layer functions similarly to the hidden layer. Feature vector
Figure BDA0002866648280000068
Respectively mapping to 2 parts in the feature compression layer, and correspondingly calculating the mean value mumSum variance σ2mThe calculation formula is as follows:
Figure BDA0002866648280000069
Figure BDA00028666482800000610
wherein theta is2＝{W2,b2And theta3＝{W3,b3And is the parameter to be determined of the sampling layer. Then random sampling is carried out on N (0,1) to obtain a noise vector epsilonm. The latent variable z is accomplished in a random sampling operation layer in the following waymSampling:
zm＝μm+σ2m×εm (5)
in order to promote gradient calculation of the model, a re-parameterization skill is introduced in the sampling process to realize the average value mumAnd variance σ2mAnd (6) derivation. Obtaining a hidden variable z obeying a normal distributionm～N(μm,σ2m) Which is consistent with the probability distribution of the original sample. And, the generalization capability of the model is improved by adding the noise vector. Finally, calculating and determining an implicit variable z by adopting the idea of variation inference in a calculation layermThe model parameters of (1).
4. Second hidden layer
Similar to the first hidden layer, the second hidden layer passes through a function
Figure BDA0002866648280000071
Compressing and transforming the output of the sampling layer to obtain the characteristic vector
Figure BDA0002866648280000072
The formula for the neuron is as follows:
Figure BDA0002866648280000073
wherein theta is4＝{W4,b4Is the parameter to be optimized of the second hidden layer, W4Is the weight of the second hidden layer, b4Is an offset. Likewise, SELU is chosen as the second hidden layer activation function sf2And the Dropout algorithm is applied to the second hidden layer.
5. Output layer
And outputting the multi-feature fusion health factor based on SCVNN by the output layer.
Fig. 2 shows a processing procedure of the SCVNN-based multi-feature health factor fusion method, which includes the following steps:
(1) collecting vibration signals of a rotating machine as data input;
(2) carrying out smooth denoising pretreatment on the original vibration data, extracting characteristics such as a time domain, a frequency domain, a time-frequency domain and the like to construct an original characteristic set, and then normalizing the characteristics;
(3) and (3) screening to obtain characteristics closely tracked with the degradation process of the rotary machine and constructing a sensitive characteristic set by using the correlation, monotonicity and robustness indexes as evaluation standards of various characteristic quantities.
(4) And inputting the sensitive feature set into an SCVNN model, performing feature fusion training, and inputting the data of the test set into the trained model to obtain a health factor representing the health state of the rotary machine.
The rotary machine performance degradation state tracking and residual life prediction method selects the rotary machine of a certain mechanical system as a research object, utilizes the vibration sensor to acquire the vibration data of the rotary machine, extracts the characteristics of the vibration data, and then inputs the vibration data into the SCVNN model to construct health factors, thereby realizing the performance degradation state tracking and the residual life prediction of the rotary machine.
For the fault prediction example of the rotating machine, the method comprises the following specific steps:
(1-1) collecting vibration data of the rotary machine by using a vibration sensor;
(2-1) performing smooth denoising pretreatment on the original vibration data by adopting a wavelet filter;
and (2-2) extracting time domain, frequency domain and time-frequency domain characteristics of the vibration data, constructing an original characteristic set and carrying out normalization processing, so as to represent comprehensive degradation state information of the rotary machine, wherein the characteristics in the original characteristic set are shown in table 1.
TABLE 1 time-Domain, frequency-Domain and time-frequency-Domain features
Figure BDA0002866648280000074
Figure BDA0002866648280000081
Where x (N), N1, 2, N represents the original vibration signal sequence acquired from the rotating machine of a mechanical system. N represents the signal sequence length. p (x (n)) represents the probability of occurrence of each data and satisfies
Figure BDA0002866648280000082
And (3-1) adopting correlation, monotonicity and robustness indexes as evaluation criteria of various characteristic quantities. In the evaluation criterion, the correlation index can measure the linear correlation degree between the characteristic parameter sequence and the time sequence; the monotonicity index can reflect the monotonous ascending or descending conversion degree of the characteristic parameter sequence; the robustness index may describe the ability of the characteristic parameter sequence to accommodate anomalous factors including noise interference.
Let E ═ E (1), E (2), E, (k) be assumed as the characteristic quantity sequence]The time sequence is T ═ T1,t2,...,tK]，e(tk) Represents the time tkWherein K represents the total time length. Firstly, a characteristic parameter sequence is divided into a stationary trend term and a random residue term by adopting a moving average method:
e(tk)＝eT(tk)+eR(tk) (7)
wherein e isT(tk) Represents the steady trend portion of the feature, and eR(tk) The random margin portion of the feature is represented.
The correlation between E and T is marked as Corr (E, T), the monotonicity index and the robustness index of E are respectively marked as Mon (E) and Rob (E), and formulas (8) to (11) describe the calculation processes of three evaluation indexes:
Figure BDA0002866648280000091
Figure BDA0002866648280000092
Figure BDA0002866648280000093
where K represents the total time length and δ (-) represents a simple unit step function, which is expressed as follows:
Figure BDA0002866648280000094
and (3-2) screening to obtain 10 characteristics closely tracked with the degradation process of the rotary machine so as to construct a sensitive characteristic set.
(4-1) construction of training set Using remaining service-life samples of rotating machine
Figure BDA0002866648280000095
Wherein x is selectedt∈RN*1N sensitive features at time t, yt∈[0,1]Is a label associated with the percentage of degradation of the mechanical rotating component at time t. For example, assuming the rotating machine has a fault time of 2800s and the current checkpoint of 1400s, then tag yt0.5. Therefore, the SCVNN model is trained by minimizing a loss function
Figure BDA0002866648280000096
Wherein the content of the first and second substances,
Figure BDA0002866648280000097
is the output of the SCVNN model; y istIs the true tag value.
(4-2) directly inputting the sensitive characteristics of the test set into the trained SCVNN model to obtain a health factor representing the health state of the rotating machine.
It should be noted that the detailed description is only for explaining and explaining the technical solution of the present invention, and the scope of protection of the claims is not limited thereby. It is intended that all such modifications and variations be included within the scope of the invention as defined in the following claims and the description.

Claims (10)

1. A multi-feature health factor fusion method based on SCVNN is characterized by comprising the following steps:
the method comprises the following steps: collecting an original vibration signal of a rotary machine;
step two: carrying out smoothing and denoising pretreatment on an original vibration signal of the rotary machine, then extracting time domain, frequency domain and time-frequency domain characteristics of the pretreated original vibration signal of the rotary machine, constructing an original characteristic set, and then carrying out normalization treatment on the signal in the original characteristic set;
step three: constructing a sensitive feature set after screening by using the normalized original feature set;
step four: inputting the sensitive feature set into an SCVNN model for feature fusion training, and inputting data of the test set into the trained model to obtain a health factor representing the health state of the rotary machine;
the SCVNN model comprises an input layer, a first hidden layer, a sampling layer, a second hidden layer and an output layer;
the SCVNN model specifically executes the following steps:
(1) an input layer: taking a sensitive feature set obtained by screening evaluation indexes as an input sample of an input layer
Figure FDA0002866648270000011
Corresponding sample label is
Figure FDA0002866648270000012
Where M is the number of samples, xmRepresenting samples of a sample set, cmRepresenting a tag in a set of tags;
(2) first hiddenComprising a layer: labeling based on sample
Figure FDA0002866648270000013
And function
Figure FDA0002866648270000014
For the sample
Figure FDA0002866648270000015
Performing compression transformation to obtain output of the first hidden layer
Figure FDA0002866648270000016
(3) Sampling layer: outputting the first hidden layer
Figure FDA0002866648270000017
Feature vector of
Figure FDA0002866648270000018
Mapping into a feature compression layer to obtain a mean value mumSum variance σ2mThen random sampling is carried out on normal distribution to obtain a noise vector epsilonmAnd finally according to the mean value mumVariance σ2mSum noise vector εmSampling in a random sampling operation layer to obtain an implicit variable zm；
(4) Second hidden layer: passing function
Figure FDA0002866648270000019
For hidden variable zmPerforming compression transformation to obtain feature vector
Figure FDA00028666482700000110
(5) An output layer: according to the feature vector
Figure FDA00028666482700000111
Obtaining a feature vector
Figure FDA00028666482700000112
And outputting the corresponding health factor.
2. The SCVNN-based multi-feature health factor fusion method according to claim 1, wherein the feature vectors
Figure FDA00028666482700000113
Expressed as:
Figure FDA00028666482700000114
wherein theta is1＝{W1,b1Is the network parameter of the first hidden layer, W1As a weight of the first hidden layer, b1For the first hidden layer bias to be applied,
Figure FDA0002866648270000021
is an activation function of the first hidden layer, and the activation function of the first hidden layer is a SELU.
3. The SCVNN-based multi-feature health factor fusion method according to claim 2, wherein the mean μmExpressed as:
Figure FDA0002866648270000022
wherein theta is2＝{W2,b2Is the sampling layer mean parameter, W2Is the mean weight, b2Is the mean offset of the sampled layers.
4. The SCVNN-based multi-feature health factor fusion method according to claim 3, whereinAt the variance σ2mExpressed as:
Figure FDA0002866648270000023
wherein theta is3＝{W3,b3Is the sampling layer variance parameter, W3As a variance weight, b3Is the variance bias of the sampling layer.
5. The SCVNN-based multi-feature health factor fusion method according to claim 4, wherein the hidden variable zmExpressed as:
zm＝μm+σ2m×εm。
6. the SCVNN-based multi-feature health factor fusion method according to claim 5, wherein the feature vectors
Figure FDA0002866648270000024
Expressed as:
Figure FDA0002866648270000025
wherein theta is4＝{W4,b4Is the parameter to be optimized of the second hidden layer, W4As a weight of the second hidden layer, b4For second hidden layer bias, zmIn order to be a hidden variable, the method comprises the following steps of,
Figure FDA0002866648270000026
is a second hidden layer activation function, which is SELU.
7. The SCVNN-based multi-feature health factor fusion method according to claim 2 or 6, wherein the SELU is expressed as:
Figure FDA0002866648270000027
wherein λ is 1.0507009873554804 and α is 1.6732632423543772.
8. The SCVNN-based multi-feature health factor fusion method according to claim 1, wherein in the third step, the normalized original feature set is used for screening by using a correlation index, a monotonicity index and a robustness index as evaluation criteria of feature quantities, wherein the correlation index is used for measuring the linear correlation degree between a feature parameter sequence and a time sequence; the monotonicity index is used for reflecting the monotonous ascending or descending conversion degree of the characteristic parameter sequence; the robustness index is used for describing the capacity of the characteristic parameter sequence to contain abnormal factors including noise interference.
9. The SCVNN-based multi-feature health factor fusion method according to claim 8, wherein the screening with the normalized raw feature set in the third step comprises:
let E ═ E (1), E (2), E, (k) be assumed as the characteristic quantity sequence]The time sequence is T ═ T1,t2,...,tK]，e(tk) Represents the time tkAnd (3) dividing the characteristic parameter sequence into a stationary trend term and a random residue term by adopting a moving average method according to the characteristic value, wherein K represents the total time length:
e(tk)＝eT(tk)+eR(tk)
wherein e isT(tk) Represents the steady trend portion of the feature, and eR(tk) Then the random margin portion of the feature is represented,
the correlation between E and T is marked as Corr (E, T), the monotonicity index and the robustness index of E are respectively marked as Mon (E) and Rob (E), and the calculation processes of three evaluation indexes are as follows:
Figure FDA0002866648270000031
Figure FDA0002866648270000032
Figure FDA0002866648270000033
where K represents the total time length and δ (-) represents a simple unit step function, which is expressed as follows:
Figure FDA0002866648270000034
10. the SCVNN-based multi-feature health factor fusion method of claim 9, wherein the loss function of the SCVNN model is:
Figure FDA0002866648270000035
wherein the content of the first and second substances,
Figure FDA0002866648270000036
is the output of the SCVNN model; y istIs the true tag value.
CN202011589552.2A 2020-12-28 2020-12-28 Multi-feature health factor fusion method based on SCVNN Active CN112597705B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011589552.2A CN112597705B (en) 2020-12-28 2020-12-28 Multi-feature health factor fusion method based on SCVNN

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011589552.2A CN112597705B (en) 2020-12-28 2020-12-28 Multi-feature health factor fusion method based on SCVNN

Publications (2)

Publication Number Publication Date
CN112597705A true CN112597705A (en) 2021-04-02
CN112597705B CN112597705B (en) 2022-05-24

Family

ID=75203378

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011589552.2A Active CN112597705B (en) 2020-12-28 2020-12-28 Multi-feature health factor fusion method based on SCVNN

Country Status (1)

Country Link
CN (1) CN112597705B (en)

Citations (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105160204A (en) * 2015-10-28 2015-12-16 河海大学 Carbon emission price combination prediction method
CN106092574A (en) * 2016-05-30 2016-11-09 西安工业大学 The Method for Bearing Fault Diagnosis selected with sensitive features is decomposed based on improving EMD
CN106682688A (en) * 2016-12-16 2017-05-17 华南理工大学 Pile-up noise reduction own coding network bearing fault diagnosis method based on particle swarm optimization
CN109000930A (en) * 2018-06-04 2018-12-14 哈尔滨工业大学 A kind of turbogenerator performance degradation assessment method based on stacking denoising self-encoding encoder
CN109976308A (en) * 2019-03-29 2019-07-05 南昌航空大学 A kind of extracting method of the fault signature based on Laplce's score value and AP cluster
CN110057584A (en) * 2019-04-01 2019-07-26 西安交通大学 A kind of locomotive traction motor bearing degradation monitoring method
US20190285517A1 (en) * 2017-10-25 2019-09-19 Nanjing Univ. Of Aeronautics And Astronautics Method for evaluating health status of mechanical equipment
CN111245375A (en) * 2020-01-19 2020-06-05 西安空间无线电技术研究所 Power amplifier digital predistortion method of complex value full-connection recurrent neural network model
CN111397902A (en) * 2020-03-22 2020-07-10 华南理工大学 Rolling bearing fault diagnosis method based on feature alignment convolutional neural network
CN111462729A (en) * 2020-03-31 2020-07-28 因诺微科技(天津)有限公司 Fast language identification method based on phoneme log-likelihood ratio and sparse representation
CN111680661A (en) * 2020-06-19 2020-09-18 哈尔滨工业大学 Mechanical rotating part performance degradation tracking method based on multi-feature fusion
CN111680446A (en) * 2020-01-11 2020-09-18 哈尔滨理工大学 Rolling bearing residual life prediction method based on improved multi-granularity cascade forest
CN112119464A (en) * 2018-02-17 2020-12-22 瑞泽恩制药公司 GAN-CNN for prediction of MHC peptide binding

Patent Citations (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105160204A (en) * 2015-10-28 2015-12-16 河海大学 Carbon emission price combination prediction method
CN106092574A (en) * 2016-05-30 2016-11-09 西安工业大学 The Method for Bearing Fault Diagnosis selected with sensitive features is decomposed based on improving EMD
CN106682688A (en) * 2016-12-16 2017-05-17 华南理工大学 Pile-up noise reduction own coding network bearing fault diagnosis method based on particle swarm optimization
US20190285517A1 (en) * 2017-10-25 2019-09-19 Nanjing Univ. Of Aeronautics And Astronautics Method for evaluating health status of mechanical equipment
CN112119464A (en) * 2018-02-17 2020-12-22 瑞泽恩制药公司 GAN-CNN for prediction of MHC peptide binding
CN109000930A (en) * 2018-06-04 2018-12-14 哈尔滨工业大学 A kind of turbogenerator performance degradation assessment method based on stacking denoising self-encoding encoder
CN109976308A (en) * 2019-03-29 2019-07-05 南昌航空大学 A kind of extracting method of the fault signature based on Laplce's score value and AP cluster
CN110057584A (en) * 2019-04-01 2019-07-26 西安交通大学 A kind of locomotive traction motor bearing degradation monitoring method
CN111680446A (en) * 2020-01-11 2020-09-18 哈尔滨理工大学 Rolling bearing residual life prediction method based on improved multi-granularity cascade forest
CN111245375A (en) * 2020-01-19 2020-06-05 西安空间无线电技术研究所 Power amplifier digital predistortion method of complex value full-connection recurrent neural network model
CN111397902A (en) * 2020-03-22 2020-07-10 华南理工大学 Rolling bearing fault diagnosis method based on feature alignment convolutional neural network
CN111462729A (en) * 2020-03-31 2020-07-28 因诺微科技(天津)有限公司 Fast language identification method based on phoneme log-likelihood ratio and sparse representation
CN111680661A (en) * 2020-06-19 2020-09-18 哈尔滨工业大学 Mechanical rotating part performance degradation tracking method based on multi-feature fusion

Non-Patent Citations (8)

* Cited by examiner, † Cited by third party
Title
YANG JINGLI: "An Efficient Method for Monitoring Degradation and Predicting the Remaining Useful Life of Mechanical Rotating Components", 《IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT》 *
侯晓东 等: "基于多指标相似性的T/R组件剩余寿命融合预测方法", 《国防科技大学学报》 *
刘胜兰 等: "基于自适应顺序的滚动轴承最优特征选取与寿命预测", 《舰船科学技术》 *
吴春志 等: "深度学习在旋转机械设备故障诊断中的应用研究综述", 《噪声与振动控制》 *
王新 等: "基于变分模态分解和极限学习机轴承寿命预测", 《制造业自动化》 *
王昱 等: "基于条件变分自编码器的齿轮箱故障诊断", 《装备环境工程》 *
邱爽 等: "自归一化卷积神经网络的人脸识别方法", 《云南大学学报(自然科学版)》 *
陈仁祥 等: "融合无量纲指标与信息熵的不同转速下旋转机械故障诊断", 《振动与冲击》 *

Also Published As

Publication number Publication date
CN112597705B (en) 2022-05-24

Similar Documents

Publication Publication Date Title
Wang et al. ReLTanh: An activation function with vanishing gradient resistance for SAE-based DNNs and its application to rotating machinery fault diagnosis
Li et al. Deep learning-based remaining useful life estimation of bearings using multi-scale feature extraction
Liang et al. Intelligent fault diagnosis of rolling bearing based on wavelet transform and improved ResNet under noisy labels and environment
Li et al. Domain generalization in rotating machinery fault diagnostics using deep neural networks
CN109033719B (en) Wind turbine bearing fault diagnosis method
CN114964778A (en) Bearing fault diagnosis method based on wavelet time-frequency graph and deep learning
CN111060316A (en) Rolling bearing state monitoring method and system based on convolutional neural network model
Zhao et al. A novel deep fuzzy clustering neural network model and its application in rolling bearing fault recognition
Saleh et al. Predicting systolic blood pressure in real-time using streaming data and deep learning
CN114580288A (en) Intelligent fault diagnosis method based on DSECJAN
Zou et al. Rolling bearing transfer fault diagnosis method based on adversarial variational autoencoder network
Yao et al. RUL prediction method for rolling bearing using convolutional denoising autoencoder and bidirectional LSTM
Liu et al. Helical fault diagnosis model based on data-driven incremental mergence
CN112597705B (en) Multi-feature health factor fusion method based on SCVNN
Han et al. Bearing fault diagnosis method based on Gramian angular field and ensemble deep learning
Shi et al. DecouplingNet: A stable knowledge distillation decoupling net for fault detection of rotating machines under varying speeds
Zhang et al. Intelligent fault diagnosis using image representation of multi-domain features
Zhao et al. Bearing fault diagnosis based on mel frequency cepstrum coefficient and deformable space-frequency attention network
Zhao et al. Unbalanced fault diagnosis of rolling bearings using transfer adaptive boosting with squeeze-and-excitation attention convolutional neural network
Tahan et al. Development of fully convolutional neural networks based on discretization in time series classification
Xiao et al. Health assessment for piston pump using LSTM neural network
CN114861740B (en) Self-adaptive mechanical fault diagnosis method and system based on multi-head attention mechanism
Wu et al. Multiscale reduction clustering of vibration signals for unsupervised diagnosis of machine faults
CN116405139A (en) Spectrum prediction model and method based on Informar
Zeng et al. Remaining useful life prediction for rotating machinery based on dynamic graph and spatial–temporal network

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