CN111562110B - Fault diagnosis model based on convolutional neural network and cross-component fault diagnosis method - Google Patents

Fault diagnosis model based on convolutional neural network and cross-component fault diagnosis method Download PDF

Info

Publication number
CN111562110B
CN111562110B CN202010449011.3A CN202010449011A CN111562110B CN 111562110 B CN111562110 B CN 111562110B CN 202010449011 A CN202010449011 A CN 202010449011A CN 111562110 B CN111562110 B CN 111562110B
Authority
CN
China
Prior art keywords
component
fault diagnosis
optimal
neural network
diagnosis model
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
CN202010449011.3A
Other languages
Chinese (zh)
Other versions
CN111562110A (en
Inventor
李恒
尹晨
王禹林
何健梁
向倍辰
房启成
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.)
Nanjing University of Science and Technology
Original Assignee
Nanjing University of Science and 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 Nanjing University of Science and Technology filed Critical Nanjing University of Science and Technology
Priority to CN202010449011.3A priority Critical patent/CN111562110B/en
Publication of CN111562110A publication Critical patent/CN111562110A/en
Application granted granted Critical
Publication of CN111562110B publication Critical patent/CN111562110B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • 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
    • 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

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

Abstract

The invention provides a fault diagnosis model based on a convolutional neural network and a cross-component fault diagnosis method. Firstly, respectively acquiring operation data of a mechanical component A in normal and fault states, constructing a series of diagnosis models for fault diagnosis of the mechanical component A by using an orthogonal optimization design method and a one-dimensional convolutional neural network, and further analyzing the models to obtain an optimal fault diagnosis model of the component A based on orthogonal optimization experiment results; then, performing cross-component fault diagnosis modeling based on a network model reuse transfer learning method, transferring the weight of the optimal fault diagnosis model of the component A in a layer-by-layer adaptive mode, and constructing a series of transfer diagnosis models for fault diagnosis of the component B; and finally, retraining all the migration diagnosis models by using the data of the mechanical component B, and obtaining the optimal fault diagnosis model of the mechanical component B by comparing the accuracy of the migration diagnosis models, thereby realizing cross-component fault diagnosis based on the orthogonal optimization convolutional neural network and the migration learning.

Description

Fault diagnosis model based on convolutional neural network and cross-component fault diagnosis method
Technical Field
The invention belongs to the technical field of fault diagnosis of mechanical parts, and particularly relates to a fault diagnosis model based on a convolutional neural network and a cross-part fault diagnosis method.
Background
High-end equipment such as large-scale fans, high-grade numerical control machines, aerospace equipment, modern agricultural machinery equipment and the like are all composed of high-precision structural functional components, and the fault diagnosis of the important structural functional components of the high-end equipment is of great significance in order to guarantee the normal operation of the equipment. In recent years, data-driven intelligent methods represented by deep learning have been widely used for failure diagnosis of important structural functional components, and have achieved remarkable diagnostic effects. However, the experimental data of some structural functional components are lost due to objective reasons such as high experimental cost and long fault period, which hinders the implementation and application of the deep learning method in the structural functional components.
In recent years, a large amount of research is carried out on the diagnosis of rolling bearings, ball screws and the like under stable working conditions and variable working conditions by scholars at home and abroad: CN110210381A discloses a domain separation adaptive one-dimensional convolutional neural network intelligent fault diagnosis method, which solves the problem of diagnosis precision reduction caused by training data and test data from different domains, but does not describe how to construct an optimal diagnosis model based on a one-dimensional convolutional neural network. CN110188822A discloses a domain-impedance adaptive one-dimensional convolutional neural network intelligent fault diagnosis method, which comprises a construction stage and a learning stage, wherein the construction stage only discusses the construction of a classification loss function, and does not relate to how to select the optimal one-dimensional convolutional neural network structure parameters such as the number of layers of a convolutional network, the number of neurons in each layer and the like. CN105046322A discloses a lead screw fault breaking method, which adopts a deep neural network and a Softmax regression classifier, but does not perform optimization analysis on key influence factors influencing the performance of the neural network. CN109146055A proposes an improved particle swarm optimization method based on an orthogonalization experiment and an artificial neural network, but the orthogonal experiment design method is only used for setting a training data set and a testing data set of the neural network, and does not involve optimization of a network structure. CN109902393A discloses a bearing fault diagnosis method under variable working conditions based on deep layer and transfer learning, which adopts JGSA to carry out adaptive processing on a source domain characteristic sample and a target sample; CN110543860A discloses a mechanical fault diagnosis method and system based on TJM transfer learning, which solves the problem of low fault diagnosis efficiency among different working conditions of the same type of rotating machinery; CN110751207A discloses a fault diagnosis method for anti-migration learning based on a deep convolution domain, which utilizes labeled samples under historical conditions to perform high-precision fault diagnosis on a current sample to be tested of a rotating machine, but is limited to migration fault diagnosis of functional components of the same kind of important structures.
In summary, the current method for diagnosing the fault of the mechanical functional component mainly has the following two limitations: firstly, how to efficiently construct an optimal fault diagnosis model based on convolutional neural network science is lack of systematic explanation, and secondly, most of the existing migration is based on model migration of the same type of components under different working conditions and different operation stages, and migration fault diagnosis among important structural functional components is not realized.
Disclosure of Invention
The invention aims to provide a fault diagnosis model based on a convolutional neural network and a cross-component fault diagnosis method, systematically expounds general steps for efficiently and scientifically constructing an optimal fault diagnosis model based on an orthogonal design method, and further discloses a cross-component fault diagnosis modeling step by using a network model reuse transfer learning method so as to realize transfer fault diagnosis between different types of mechanical functional components.
The technical scheme of the invention is as follows:
a fault diagnosis model based on a convolutional neural network is established by the following steps:
step 1, acquiring operation data of a component in a stable operation process in a normal state and a fault state respectively, and establishing a fault diagnosis model of the component based on a one-dimensional convolutional neural network;
step 2, listing influence factors which obviously influence the performance of the one-dimensional convolutional neural network, determining the typical level of each influence factor in the component fault modeling process, and establishing a corresponding orthogonal experiment parameter table;
step 3, selecting an orthogonal matrix meeting analysis requirements based on an orthogonal optimization theory, obtaining a factor horizontal combination needing to be researched and analyzed in the component fault modeling process, and constructing a one-dimensional convolutional neural network fault diagnosis model corresponding to the factor horizontal combination one by one;
and 4, training and testing the diagnostic models by using the collected component operation data, obtaining the average correct rate of each diagnostic model of the component after repeated experiments, and calculating by adopting a range analysis method and an variance analysis method to obtain an optimal factor level combination so as to obtain an optimal diagnostic model of the component.
Preferably, in step 1, the operation data during the stable operation of the component may be one or more of force, current, vibration, temperature or acoustic emission data.
Preferably, in step 2, the influencing factors that significantly influence the performance of the one-dimensional convolutional neural network include the number of neurons in a convolutional layer, the dimension of an input signal, the size of a one-dimensional convolutional kernel, the learning rate, and the number of batch processes.
Preferably, the process of step 4 is as follows:
step 4.1, with the accuracy as an index, respectively calculating the range R of each factor by using a range analysis methodiAnd the average value K of the accuracy at different levels of each factorijWherein i represents the ith influencing factor in the orthogonal experiment parameter table, and j represents the jth level in the orthogonal experiment parameter table;
step 4.2, according to the sequence of the range from big to small, main and secondary factors influencing the performance of the one-dimensional convolution neural network are obtained, and the maximum average value [ K ] of all the factors is usedij]maxThe corresponding level is the principle of the optimal level of the factor, and the optimal level of each influencing factor is obtained;
step 4.3a, if the relative error of all the factor range is not less than 20%, the result obtained by the range analysis method is the optimal factor level combination, wherein the calculation formula of the relative error between the two factor range is as follows:
Figure BDA0002506829750000021
4.3b, if the relative error of the extreme difference of the factors is lower than 20%, further combining an optimization orthogonal experiment to check whether interaction exists among the corresponding factors, and combining a variance analysis method to obtain the optimal level combination of all the influencing factors;
and 4.4, verifying the accuracy of the optimal level combination, and constructing a one-dimensional convolutional neural network corresponding to the optimal level combination to obtain an optimal diagnosis model.
Preferably, the step 4.3b process is as follows: firstly, fixing the factors with the relative error higher than 20% to the optimal level, and removing the worst level with the relative error smaller than 20% to obtain an optimized orthogonal experiment parameter table; then selecting an orthogonal matrix meeting the analysis requirement of the factor interaction according to the optimized orthogonal experiment parameter table, and developing an optimized orthogonal experiment; and finally, determining the significance of the factors and the interaction thereof by using a joint hypothesis test in combination with an analysis of variance method to obtain the optimal level combination of all the influencing factors.
A cross-component fault diagnosis method comprising the steps of:
step 1-4, corresponding to a component A, establishing an optimal fault diagnosis model based on a one-dimensional convolutional neural network according to any one of claims 1-5;
step 5, performing cross-component fault diagnosis modeling based on a network model reuse transfer learning method, transferring the weight of the optimal fault diagnosis model of the component A in a layer-by-layer adaptive mode, and constructing a series of transfer diagnosis models for fault diagnosis of the component B;
and 6, retraining all the migration diagnosis models by using the operating data of the component B, and obtaining the optimal fault diagnosis model of the component B by comparing the accuracy of the migration diagnosis models to realize the cross-component migration fault diagnosis.
Preferably, in step 5, the number of neurons in the last fully-connected layer changes with the difference between diagnosis tasks, and the parameters of the migration diagnosis model for diagnosing the fault of the component B and the optimal fault diagnosis model for diagnosing the fault of the component a are kept consistent.
Preferably, the process of step 5 is as follows:
step 5.1, the optimal fault diagnosis model of the component A based on the one-dimensional convolutional neural network has N in totalCA convolutional layer and NFThe full connection layers randomly initialize a diagnosis model with the same structure as the component A, and modify the neuron number of the last full connection layer according to the fault category of the component B to be used as an initial fault diagnosis model of the component B;
step 5.2, extracting the weight of the convolution layer in the optimal fault diagnosis model of the component A layer by layer, sequentially transferring the weight to the initial fault diagnosis model of the component B, and constructing NCA fault diagnosis model for the component B, wherein the ith fault diagnosis model of the component B comprises the weights of the first i convolutional layers of the optimal fault diagnosis model of the component A, and i is more than or equal to 1 and less than or equal to NC。
Compared with the prior art, the invention has the following remarkable advantages:
(1) compared with the conventional other fault diagnosis methods based on the convolutional neural network, which only provides a specific network structure for a specific research object, the method can quickly, efficiently and accurately acquire the optimal fault diagnosis model for any object, and has stronger universality;
(2) compared with the traditional fault diagnosis method, the fault diagnosis is divided into two stages of fault feature extraction and fault mode identification, the fault features are extracted depending on a large amount of prior experience and cannot be extracted in a self-adaptive manner, the fault diagnosis model provided by the method is composed of a one-dimensional convolutional neural network, the fault feature extraction and the fault mode identification are integrated into a whole while the fault features are extracted in a self-adaptive manner, end-to-end output of a diagnosis result can be directly realized based on original data, and the fault diagnosis efficiency is improved;
(3) compared with the conventional fault diagnosis method which is mostly limited to model migration of a certain type of parts under different working conditions and different operation stages, the migration diagnosis method is widely suitable for migration diagnosis research among different types of mechanical parts such as ball screws and rolling bearings, and can effectively solve the problem of low fault diagnosis accuracy rate of certain parts due to small data volume.
Drawings
FIG. 1 is a flow chart of a method for establishing a convolutional neural network-based fault diagnosis model and cross-component fault diagnosis.
Fig. 2 is a schematic diagram of an experimental and signal acquisition device for diagnosing and researching a fault of a ball screw in the embodiment.
Fig. 3 is a time domain diagram of the vibration signal in one reciprocating cycle of the ball screw mechanism in the embodiment.
Fig. 4 is a failure diagnosis result of the ball screw migration model in the embodiment.
Tailstock 1, base 2, ball screw 3, acceleration sensor 4, ball nut 5, workstation 6, motor 7, speed sensor 8, temperature sensor 9.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.
As shown in fig. 1, the method mainly includes two parts of establishing a fault diagnosis model based on an orthogonal optimization convolutional neural network and performing cross-component fault diagnosis based on transfer learning, and the method mainly includes the following steps:
step 1, acquiring operation data of a component A and a component B in a stable operation process respectively in a normal state and a fault state, and establishing a fault diagnosis model of the component A and the component B based on a one-dimensional convolutional neural network;
step 2, listing influence factors which obviously influence the performance of the one-dimensional convolutional neural network, determining the typical level of each influence factor in the fault modeling process of the component A, and establishing a corresponding orthogonal experiment parameter table;
step 3, selecting an orthogonal matrix meeting analysis requirements based on an orthogonal optimization theory, obtaining a factor horizontal combination needing to be researched and analyzed in the fault modeling process of the component A, and constructing a one-dimensional convolutional neural network fault diagnosis model corresponding to the factor horizontal combination one by one;
step 4, training and testing the diagnosis model by using the collected component A operation data, obtaining the average accuracy of each diagnosis model of the component A after repeated experiments, and calculating by adopting an extreme difference analysis method and a variance analysis method to obtain an optimal factor level combination to obtain an optimal diagnosis model of the component A;
step 5, performing cross-component fault diagnosis modeling based on a network model reuse transfer learning method, transferring the weight of the optimal fault diagnosis model of the component A in a layer-by-layer adaptive mode, and constructing a series of transfer diagnosis models for fault diagnosis of the component B;
and 6, retraining all the migration diagnosis models by using the operating data of the component B, and obtaining the optimal fault diagnosis model of the component B by comparing the accuracy of the migration diagnosis models to realize the cross-component migration fault diagnosis.
The pretightening force level of the ball screw is an important precondition for stable operation of the ball screw mechanism, but in practical application, severe working conditions and continuous impact of mechanical operation can directly cause abrasion of the ball screw pair, so that the pretightening force of the ball screw is reduced, and the operation efficiency and the positioning accuracy of the ball screw mechanism are further reduced. However, the failure period of the ball screw is long, the cost of the fault experiment is high, and the like, so that the data which can be used for ball screw fault modeling is less, and it is difficult to directly establish a fault diagnosis model with high accuracy. On the other hand, as an important mechanical component, a large amount of data is accumulated in the process of research by a large number of researchers. Therefore, in the embodiment, a cross-component fault diagnosis method based on orthogonal optimization convolutional neural network and transfer learning is explained by taking the transfer fault diagnosis of the rolling bearing to the ball screw as a research object.
Specifically, in the present embodiment, the component a is a rolling bearing, and the component B is a ball screw. According to the steps, the fault diagnosis method specifically introduces the following steps for the failure diagnosis model of the pre-tightening force recession of the rolling bearing and the failure diagnosis of the pre-tightening force recession of the ball screw:
1. ball screw and rolling bearing vibration signal acquisition
1.1 ball screw vibration Signal acquisition
As shown in fig. 2, the experimental device for acquiring the vibration signal of the ball screw is composed of a tailstock 1, a base 2, a ball screw 3, an acceleration sensor 4, a ball nut 5, a workbench 6, a motor 7, a speed sensor 8 of the motor and a temperature sensor 9. The acceleration sensor 4 is used for collecting vibration acceleration signals of the ball screw pair, and the motor 7 is internally provided with a speed sensor and a temperature sensor and used for acquiring rotating speed and temperature data of the motor in real time. In practical application, the motor 7 drives the ball screw 3 to rotate, the ball nut 6 converts the rotation motion of the ball screw 3 into linear motion, and the workbench 6 is driven to move. In this embodiment, according to the characteristic that the early failure symptom of the ball screw 3 appears in a high frequency band, the acceleration sensor 8 with a high resonant frequency is selected, so that it is ensured that the acquired vibration acceleration signal of the ball screw 3 includes the failure characteristic signal of the ball screw 3, and the failure diagnosis of the ball screw 3 is realized.
When the signals are collected, the ball screw is arranged on the workbench, and the ball screw rotates forwards and backwards to drive the ball nut to reciprocate along the horizontal guide rail. In order to simulate the decay process of the pretightening force of the ball screw mechanism in the actual operation process and realize the diagnosis of the health level of the ball screw under different pretightening force levels, the pretightening force of the ball screw is respectively set to be 0%, 4% and 8% of the rated dynamic load (30kN) of the ball screw to carry out three groups of tests, and the vibration signals of the ball screw under three groups of states are collected by a data acquisition system. FIG. 3 is a vibration signal of the ball screw mechanism during one reciprocating cycle acquired at a pre-load level of 8%. Since one reciprocating motion of the ball screw mechanism includes two parts: in both forward and reverse motions, the ball nut will suddenly brake at both ends of the ball screw lever and then reverse, so that the vibration signal collected during one cycle of reciprocation will appear "noisy" in both the intermediate and final stages. According to the foregoing step 1, only the operation data during steady operation is used to build the failure diagnosis model of the component based on the one-dimensional convolutional neural network, and therefore, the steady operation signals in the sampling times of 3 rd to 20 th and 26 th to 43 th seconds shown in fig. 3 are retained. Vibration signals of stable operation stages in three reciprocating periods are collected for each pretightening force level, and 520000 data points are reserved as data for establishing a ball screw diagnosis model.
1.2 Rolling bearing vibration Signal acquisition
The experiment adopts bearing data collected by a mechanical equipment health monitoring combined laboratory rolling bearing accelerated life test bench of the transport university of western Ann (WANG Biao, LEI Yaguo, LI Naipeng, et al. A hybrid diagnostics for evaluating relating using life of rolling elements [ J ]. IEEE Transactions on Reliability, 2018: 1-12.). The experimental platform comprises an alternating current motor, a motor rotating speed controller, a rotating shaft, a supporting bearing, a hydraulic loading system, a testing bearing and the like. The test totally carries out the life cycle experiment to 15 bearings under 3 kinds of operating modes. In the test, a DT9837 data acquisition system is used for acquiring vibration signals at the sampling frequency of 25.6kHz, the sampling time length of each time is 1.28s, and the sampling interval is 1 min. And finally, extracting 2800000 data points in the stable operation process for each category as data for establishing a diagnosis model of the rolling bearing, wherein the failure reasons of the bearing to be tested comprise inner ring abrasion, retainer fracture, outer ring failure and normal operation conditions.
2. Ball bearing fault diagnosis orthogonal parameter test table
According to the foregoing step 2, the influencing factors influencing the performance of the fault diagnosis model based on the one-dimensional convolutional neural network mainly include the dimension of the input signal, the size of the one-dimensional convolutional kernel, the number of neurons in the convolutional layer, the learning rate, and the number of batch processes, which are sequentially represented by the capital letters A, B, C, D, E. In addition to the above typical influencing factors, the basic architecture of the constructed one-dimensional convolutional neural network is sequentially an input layer (input), a convolutional layer 1(C1), a pooling layer (P1), a convolutional layer 2(C2), a pooling layer (P2), a convolutional layer 3(C3), a pooling layer (P3), a full connection layer 1(FC1), a full connection layer 2(FC2) and a full connection layer 3(FC3), wherein the pooling layer adopts maximum pooling, the length and the step size of the pooling kernel are both 4, FC1 is the input layer of the full connection layer, and the number of neurons of FC2 and FC3 is sequentially 100 and 4. The convolution layer and the full-connection layer both adopt a linear rectification function (ReLU function) as an activation function, the output layer adopts a sigmoid function to activate and then inputs a cross entropy loss function to carry out error calculation, and a root mean square back propagation method (RMSprop) is used to calculate the gradient of the convolution layer and the full-connection layer to carry out back propagation.
By looking up documents for carrying out rolling bearing fault diagnosis by adopting a one-dimensional convolutional neural network, selecting and determining typical levels of all influence factors in the ball bearing fault diagnosis modeling process, and finally determining an orthogonal parameter experiment table as shown in table 1, wherein the input signal dimension of the influence factor A represents the number of vibration signal data points contained in an input signal, the neuron number (m, n) of the convolution layer of the influence factor C represents that the number of a first convolution layer is m, and the number of a second convolution layer is n.
TABLE 1 rolling bearing fault diagnosis orthogonal parameter experiment table
Figure BDA0002506829750000071
3. Rolling bearing diagnosis model based on orthogonal matrix design
As can be seen from the orthogonal test table shown in Table 1, this orthogonal optimization design belongs to the 5-factor 4 horizontal test. Thus, L as shown in Table 2 was used16(45) The orthonormal matrix was subjected to an orthogonal test. And constructing one-dimensional convolutional neural networks corresponding to the factor level combinations according to the factor level combinations shown in the table 2.
TABLE 2L16(45) Orthonormal matrix
Figure BDA0002506829750000072
4. Rolling bearing optimal diagnosis model establishment
4.1L16(45) Results of orthogonal experiments
According to L shown in Table 216(45) The standard orthogonal matrix has 16 kinds of one-dimensional convolutional neural networks with different structures. The accuracy is taken as an index of orthogonal optimization, the average value of the accuracy is obtained after 20 times of experiments are carried out on each framework, and the experimental results are shown in table 3.
TABLE 3L16(45) Results of orthogonal experiments
Figure BDA0002506829750000073
Figure BDA0002506829750000081
4.2 very poor analysis of test results
Using range analysis to L16(45) The results of the orthogonal experiments were analyzed, and the analysis results are shown in table 4. The extreme difference of all factors can show that the performance of the one-dimensional rolling neural network fault diagnosis model is influenced in the fault diagnosis of the ball bearingThe primary and secondary ordering of the factors is as follows: the learning rate, the number of convolutional layer neurons, the size of a one-dimensional convolution kernel, the dimension of an input signal and the number of batch processing are combined into D4C2B2A1E3. According to a cross-component fault diagnosis method based on orthogonal optimization convolutional neural network and transfer learning, the method comprises the following steps of: if there is a relative error of the factor badly less than or equal to 20%, the interaction between the corresponding factors should be further considered. In this example, the relative errors between the factor a input signal dimension, the factor B one-dimensional convolution kernel dimension, and the factor C convolution layer neuron number are respectively 15.5%, 3.6%, and 18.6%, which are all lower than 20%, so it is necessary to further combine the optimization orthogonal experiment, and analyze whether there is an interaction between the factor a, the factor B, and the factor C by using an analysis of variance method.
TABLE 4L16(45) Extreme analysis of experimental results
Figure BDA0002506829750000082
4.3 analysis of variance of influencing factor interactions
According to the results of the range analysis, the optimum levels of the learning rate and the number of batches were 0.00001 and 64, respectively, and thus the learning rate and the number of batches were fixed to the optimum levels in the optimization orthogonal experiment. The worst level of the input signal dimension, the one-dimensional convolution kernel size and the convolution layer neuron number is further removed, and the finally obtained optimized orthogonal experiment parameter table is shown in table 5.
TABLE 5 optimized orthogonal parameters Experimental Table
Figure BDA0002506829750000091
Selecting L according to the established orthogonal parameter experiment table27(313) And analyzing the interaction among the dimension of the input signal of the factor A, the dimension of the one-dimensional convolution kernel of the factor B and the number of convolution layer neurons of the factor C by using the standard orthogonal matrix. Parameter setting and optimization orthogonal experimental nodeAs shown in table 6. Wherein each group of experiments are repeated for 20 times, the average accuracy is taken as the final experiment result, and irrelevant factors and blank columns in the table are not listed.
TABLE 6 optimized orthogonal experimental results
Figure BDA0002506829750000092
For the optimized orthogonal experiment results, the analysis of the factor a (input signal dimension), the factor B (one-dimensional convolution kernel size), the factor C (convolution layer neuron number) and the interaction between the two factors is performed by the analysis of variance, and the analysis results are shown in table 7. As can be seen from table 7, in the fault diagnosis of the rolling bearing, the primary and secondary orders of the influence of the input signal dimension, the one-dimensional convolution kernel dimension, the convolution layer neuron number and the interaction between the convolution layer neuron number on the performance of the fault diagnosis model are as follows: factor B, factor C, factor a × B (interaction between input signal dimension and one-dimensional convolution kernel size), factor a × C (interaction between input signal dimension and convolutional layer neuron number), and factor B × C (interaction between one-dimensional convolution kernel size and convolutional layer neuron number). Therefore, the optimal levels corresponding to the input signal dimension, the one-dimensional convolution kernel size, and the number of convolution layer neurons are determined by factor B, factor C, and factor a × C, respectively. Further, by comparing the factor B, the factor C, and the average value of the accuracy of the factor a × C at different levels, the optimal levels corresponding to the dimension of the input signal, the dimension of the one-dimensional convolution kernel, and the number of convolution layer neurons are 1280, 13, and (64,32), respectively.
TABLE 7 results of analysis of variance in optimized orthogonal experiments
Figure BDA0002506829750000101
4.4 optimal Fault diagnosis model of ball bearing
The analysis results of range analysis and variance analysis are integrated, the optimal fault diagnosis model of the ball bearing and the diagnosis accuracy of the optimal fault diagnosis model are shown in table 8, the test results show that the accuracy of the optimal horizontal combination is up to 100%, and the effectiveness of orthogonal optimization design in establishing the performance of the optimal fault diagnosis model based on the one-dimensional convolutional neural network is verified. The network weight of the convolution layer in the optimal fault diagnosis model of the ball bearing is transferred to a transfer model in the fault diagnosis of the ball screw, so that the fault diagnosis of the cross-component is realized.
TABLE 8 optimal factor level combination and Fault diagnosis accuracy
Figure BDA0002506829750000102
5. Method for establishing ball screw migration fault diagnosis model
According to the step 5 of the cross-component fault diagnosis method based on the orthogonal optimization convolutional neural network and the transfer learning, two ball screw diagnosis models with the parameters consistent with the optimal model of the rolling bearing except the number of the neurons in the last full-connection layer are initialized. The method is divided into three types according to the pretightening force level in the fault diagnosis of the ball screw, and the number of the neurons in the last full-connection layer is adjusted to 3 from 4. That is, the initialized diagnosis model of the ball screw consists of a one-dimensional convolutional neural network including one input layer, two convolutional layers, and three fully-connected layers, wherein the number of neurons of the input layer is 1280, the number of neurons of the first convolutional layer is 64, the number of neurons of the second convolutional layer is 32, the number of neurons of the first fully-connected layer is 2432, the number of neurons of the second fully-connected layer is 100, and the number of neurons of the last fully-connected layer is 3. The learning rate in the diagnostic model training process of the ball screw was 0.00001, and the number of batches was 64. And then, based on a transfer learning method for reusing the network model, transferring the convolution layer weight of the optimal ball bearing model to the initialized ball screw diagnosis model in a layer-by-layer adaptive mode, and constructing two transfer diagnosis models for ball screw fault diagnosis. The ball screw failure diagnosis model which only transfers the weights of 64 neurons in the first convolution layer of the ball bearing optimal model is M1, and the ball screw failure diagnosis model which transfers the weights of 64 neurons in the first convolution layer and 32 neurons in the second convolution layer of the ball bearing optimal model is M2.
6. Optimal migration fault diagnosis model for ball screw
According to the cross-component fault diagnosis method based on the orthogonal optimization convolutional neural network and the transfer learning, step 6, small sample data of the ball screw collected in step 1 under three different pretightening force levels are utilized to perform cross-component fault diagnosis on the M constructed in step 51And M2Retraining was performed, in which each model was repeatedly trained 20 times, taking the average accuracy as the final experimental result, as shown in fig. 4. The accuracy (93.9%) of fault modeling by utilizing small sample data of a ball screw based on a one-dimensional convolutional neural network directly without using the method is marked in a figure, and the accuracy can be used for judging whether a cross-component fault diagnosis method based on an orthogonal optimization convolutional neural network and transfer learning is effective or not. By comparing the accuracy of the migration diagnosis model, the optimal fault diagnosis model is M1The failure diagnosis accuracy is 95.7%. In summary, the cross-component fault diagnosis method based on the orthogonal optimization convolutional neural network and the transfer learning in the embodiment can effectively realize the transfer of diagnosis knowledge between different components.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (8)

1. A fault diagnosis model based on a convolutional neural network is characterized by comprising the following establishing steps:
step 1, acquiring operation data of a component in a stable operation process in a normal state and a fault state respectively, and establishing a fault diagnosis model of the component based on a one-dimensional convolutional neural network;
step 2, listing influence factors which obviously influence the performance of the one-dimensional convolutional neural network, determining the typical level of each influence factor in the component fault modeling process, and establishing a corresponding orthogonal experiment parameter table;
step 3, selecting an orthogonal matrix meeting analysis requirements based on an orthogonal optimization theory, obtaining a factor horizontal combination needing to be researched and analyzed in the component fault modeling process, and constructing a one-dimensional convolutional neural network fault diagnosis model corresponding to the factor horizontal combination one by one;
and 4, training and testing the diagnostic models by using the collected component operation data, obtaining the average correct rate of each diagnostic model of the component after repeated experiments, and calculating by adopting a range analysis method and an variance analysis method to obtain an optimal factor level combination so as to obtain an optimal diagnostic model of the component.
2. The convolutional neural network-based fault diagnosis model of claim 1, wherein in the step 1, the operation data during the steady operation of the component is one or more of force, current, vibration, temperature or acoustic emission data.
3. The convolutional neural network-based failure diagnosis model of claim 1, wherein in the step 2, the influencing factors significantly influencing the performance of the one-dimensional convolutional neural network include the number of convolutional layer neurons, the dimension of the input signal, the size of the one-dimensional convolutional kernel, the learning rate and the number of batch processes.
4. The convolutional neural network-based fault diagnosis model of claim 1, wherein the process of step 4 is as follows:
step 4.1, with the accuracy as an index, respectively calculating the range R of each factor by using a range analysis methodiAnd the average value K of the accuracy at different levels of each factorijWherein i represents the ith influencing factor in the orthogonal experiment parameter table, and j represents the jth level in the orthogonal experiment parameter table;
step 4.2, according to the sequence of the range from big to small, main and secondary factors influencing the performance of the one-dimensional convolution neural network are obtained,according to the maximum average value [ K ] of each factorij]maxThe corresponding level is the principle of the optimal level of the factor, and the optimal level of each influencing factor is obtained;
step 4.3a, if the relative error of all the factor range is not less than 20%, the result obtained by the range analysis method is the optimal factor level combination, wherein the calculation formula of the relative error between the two factor range is as follows:
Figure FDA0002811390040000011
4.3b, if the relative error of the extreme difference of the factors is lower than 20%, further combining an optimization orthogonal experiment to check whether interaction exists among the corresponding factors, and combining a variance analysis method to obtain the optimal level combination of all the influencing factors;
and 4.4, verifying the accuracy of the optimal level combination, and constructing a one-dimensional convolutional neural network corresponding to the optimal level combination to obtain an optimal diagnosis model.
5. The convolutional neural network-based fault diagnosis model of claim 4, wherein the step 4.3b process is as follows: firstly, fixing the factors with the relative error higher than 20% to the optimal level, and removing the worst level with the relative error smaller than 20% to obtain an optimized orthogonal experiment parameter table; then selecting an orthogonal matrix meeting the analysis requirement of the factor interaction according to the optimized orthogonal experiment parameter table, and developing an optimized orthogonal experiment; and finally, determining the significance of the factors and the interaction thereof by using a joint hypothesis test in combination with an analysis of variance method to obtain the optimal level combination of all the influencing factors.
6. A cross-component fault diagnosis method, comprising the steps of:
step 1-4, corresponding to a component A, establishing an optimal fault diagnosis model based on a one-dimensional convolutional neural network according to any one of claims 1-5;
step 5, performing cross-component fault diagnosis modeling based on a network model reuse transfer learning method, transferring the weight of the optimal fault diagnosis model of the component A in a layer-by-layer adaptive mode, and constructing a series of transfer diagnosis models for fault diagnosis of the component B;
and 6, retraining all the migration diagnosis models by using the operating data of the component B, and obtaining the optimal fault diagnosis model of the component B by comparing the accuracy of the migration diagnosis models to realize the cross-component migration fault diagnosis.
7. The cross-component fault diagnosis method according to claim 6, wherein in the step 5, the migration diagnosis model of the component B fault diagnosis and the component A optimal fault diagnosis model are consistent in parameters except that the number of neurons of the last fully-connected layer changes with the difference between diagnosis tasks.
8. The cross-component fault diagnosis method according to claim 6, wherein the process of step 5 is as follows:
step 5.1, the optimal fault diagnosis model of the component A based on the one-dimensional convolutional neural network has N in totalCA convolutional layer and NFThe full connection layers randomly initialize a diagnosis model with the same structure as the component A, and modify the neuron number of the last full connection layer according to the fault category of the component B to be used as an initial fault diagnosis model of the component B;
step 5.2, extracting the weight of the convolution layer in the optimal fault diagnosis model of the component A layer by layer, sequentially transferring the weight to the initial fault diagnosis model of the component B, and constructing NCA fault diagnosis model for the component B, wherein the ith fault diagnosis model of the component B comprises the weights of the first i convolutional layers of the optimal fault diagnosis model of the component A, and i is more than or equal to 1 and less than or equal to NC。
CN202010449011.3A 2020-05-25 2020-05-25 Fault diagnosis model based on convolutional neural network and cross-component fault diagnosis method Active CN111562110B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010449011.3A CN111562110B (en) 2020-05-25 2020-05-25 Fault diagnosis model based on convolutional neural network and cross-component fault diagnosis method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010449011.3A CN111562110B (en) 2020-05-25 2020-05-25 Fault diagnosis model based on convolutional neural network and cross-component fault diagnosis method

Publications (2)

Publication Number Publication Date
CN111562110A CN111562110A (en) 2020-08-21
CN111562110B true CN111562110B (en) 2021-02-26

Family

ID=72075044

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010449011.3A Active CN111562110B (en) 2020-05-25 2020-05-25 Fault diagnosis model based on convolutional neural network and cross-component fault diagnosis method

Country Status (1)

Country Link
CN (1) CN111562110B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112179481B (en) * 2020-09-01 2023-01-06 国网山西省电力公司电力科学研究院 Mechanical fault diagnosis method and system for GIS equipment
CN114611564A (en) * 2020-11-24 2022-06-10 国机工业互联网研究院(河南)有限公司 Steam pipeline drain valve fault detection method based on convolutional neural network and temperature
CN113283532B (en) * 2021-06-09 2022-09-09 重庆大学 Water chilling unit fault diagnosis method based on deep migration learning
CN113869451B (en) * 2021-10-12 2024-05-24 中国民航大学 Rolling bearing fault diagnosis method under variable working condition based on improved JGSA algorithm

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20190018798A (en) * 2017-08-16 2019-02-26 강병수 car noise sound with Convolution Nueral Network classification method
CN108344574B (en) * 2018-04-28 2019-09-10 湖南科技大学 A kind of Wind turbines Method for Bearing Fault Diagnosis based on depth joint adaptation network
KR102093929B1 (en) * 2018-12-05 2020-03-26 중앙대학교 산학협력단 Apparatus and Method for Diagnosing Mechanical System Health based on CIM
CN109918752B (en) * 2019-02-26 2022-12-16 华南理工大学 Mechanical fault diagnosis method, equipment and medium based on migration convolution neural network
CN109873610B (en) * 2019-03-19 2020-04-10 福州大学 Photovoltaic array fault diagnosis method based on IV characteristic and depth residual error network

Also Published As

Publication number Publication date
CN111562110A (en) 2020-08-21

Similar Documents

Publication Publication Date Title
CN111562110B (en) Fault diagnosis model based on convolutional neural network and cross-component fault diagnosis method
CN109918752B (en) Mechanical fault diagnosis method, equipment and medium based on migration convolution neural network
CN112417954A (en) Bearing fault mode diagnosis method and system for small sample data set
CN108171263A (en) Based on the Fault Diagnosis of Roller Bearings for improving variation mode decomposition and extreme learning machine
CN110031227A (en) A kind of Rolling Bearing Status diagnostic method based on binary channels convolutional neural networks
Zhang et al. Deep convolutional neural network using transfer learning for fault diagnosis
CN111523081B (en) Aeroengine fault diagnosis method based on enhanced gate control circulating neural network
CN111289250A (en) Method for predicting residual service life of rolling bearing of servo motor
CN109543905A (en) Rolling bearing method for predicting residual useful life based on improved two dimension CNN model
CN112308147A (en) Rotating machinery fault diagnosis method based on integrated migration of multi-source domain anchor adapter
CN116593157A (en) Complex working condition gear fault diagnosis method based on matching element learning under small sample
CN111459144A (en) Airplane flight control system fault prediction method based on deep cycle neural network
CN112729834A (en) Bearing fault diagnosis method, device and system
CN114993677A (en) Rolling bearing fault diagnosis method and system based on unbalanced small sample data
CN113095179A (en) Metametric learning driven helicopter planetary gearbox fault diagnosis method
CN115452376A (en) Bearing fault diagnosis method based on improved lightweight deep convolution neural network
CN117473411A (en) Bearing life prediction method based on improved transducer model
CN115587290A (en) Aero-engine fault diagnosis method based on variational self-coding generation countermeasure network
CN112146880B (en) Intelligent diagnosis method for internal structure faults of rolling bearing at different rotating speeds
CN117315328B (en) Gear fault diagnosis method for improving convolutional neural network under attention mechanism
CN117664575A (en) Rolling bearing residual service life prediction method based on double-graph attention and GRU
CN113221996A (en) Bearing fault diagnosis method and system based on multi-head attention mechanism
CN112444395A (en) CMWPE and SaE-ELM based locomotive wheel pair bearing fault diagnosis method
Xing et al. Multi-task learning for the bearing based on a one-dimensional convolutional neural network with attention guidance mechanism and multi-scale feature extraction
CN117191396A (en) Gear box fault diagnosis method based on two-stage migration

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