CN113434973B - Mechanical fault diagnosis method based on improved PSO-SOM-BPNN - Google Patents

Mechanical fault diagnosis method based on improved PSO-SOM-BPNN Download PDF

Info

Publication number
CN113434973B
CN113434973B CN202110686659.7A CN202110686659A CN113434973B CN 113434973 B CN113434973 B CN 113434973B CN 202110686659 A CN202110686659 A CN 202110686659A CN 113434973 B CN113434973 B CN 113434973B
Authority
CN
China
Prior art keywords
fault
bpnn
som
fault diagnosis
wavelet
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
CN202110686659.7A
Other languages
Chinese (zh)
Other versions
CN113434973A (en
Inventor
樊红卫
严杨
张旭辉
曹现刚
马嘉腾
马宁阁
樊奕霖
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian University of Science and Technology
Original Assignee
Xian 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 Xian University of Science and Technology filed Critical Xian University of Science and Technology
Priority to CN202110686659.7A priority Critical patent/CN113434973B/en
Publication of CN113434973A publication Critical patent/CN113434973A/en
Application granted granted Critical
Publication of CN113434973B publication Critical patent/CN113434973B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G06F30/17—Mechanical parametric or variational design
    • G—PHYSICS
    • 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
    • G01M13/02—Gearings; Transmission mechanisms
    • G01M13/021—Gearings
    • 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
    • G01M13/02—Gearings; Transmission mechanisms
    • G01M13/028—Acoustic or vibration analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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
    • G06F30/25—Design optimisation, verification or simulation using particle-based methods
    • 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
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/06—Power analysis or power optimisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/10—Noise analysis or noise optimisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G06F2218/04—Denoising
    • G06F2218/06—Denoising by applying a scale-space analysis, e.g. using wavelet analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

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

Abstract

The invention discloses a mechanical fault diagnosis method based on improved PSO-SOM-BPNN, which combines a signal processing technology and two types of neural networks in series to carry out a scheme of quickly and accurately identifying the state of a rotating machine. Inertia weight and learning factors of a PSO algorithm are adjusted in a self-adaptive mode, a speed item is abandoned to optimize SOM-BPNN parameters, a SOM-BPNN mechanical fault diagnosis method based on self-adaptive speed item-free particle swarm optimization is established, and efficient and accurate diagnosis of rotary machines is achieved.

Description

Mechanical fault diagnosis method based on improved PSO-SOM-BPNN
Technical Field
The invention relates to the field of mechanical fault diagnosis, in particular to a rotary mechanical fault diagnosis direction, and specifically relates to a mechanical fault diagnosis method based on improved PSO-SOM-BPNN (Particle Swarm Optimization-Self Organizing feature Map-Back prediction Neural Network, PSO-SOM-BPNN).
Background
With the continuous development of manufacturing industry in China, the increasing large-scale, complicated, automatic and continuous development of mechanical equipment becomes an important development trend, and the important position of rotary machinery becomes more and more prominent, so that the technical level of fault diagnosis and maintenance of rotary mechanical parts is improved, and the normal and safe operation of the mechanical equipment is an indispensable key part along with the development of the mechanical equipment.
In a mechanical system, a rotating mechanical part is in a rotating state for a long time, so that faults are easy to occur, the working efficiency and the quality are influenced slightly, and accidents of machine damage and people death are caused seriously. The method mainly comprises the steps of firstly performing wavelet denoising on signals and then extracting the characteristic energy of the rotating member by using a wavelet packet, but because the selection of wavelet bases and the number of decomposition layers of the wavelet bases and the decomposition layers is mostly selected by depending on experience, the characteristic extraction is easily inaccurate. In the aspect of network diagnosis models, besides the limitations of the models, the empirical influence of parameter setting by people is involved. The BPNN is a supervised learning network, namely training is performed on the premise of knowing expected output, and the high accuracy is obtained during fault diagnosis, so that the timeliness of diagnosis is increased and the BPNN is easy to fall into a local minimum value through training of a large number of samples; the SOM is an unsupervised network, i.e. the learning of the network depends only on the characteristics of the input data. In the learning process, the network can automatically find out the implicit rule among the input samples along with continuous iteration, and the samples are classified on the basis. If the two are combined, the SOM neural network rather performs a primary classification on the sample, and has a good promotion effect on the training of the secondary network. Theoretically, the training time of the secondary network can be effectively reduced, so that the whole series network can be converged at a higher speed, and the optimization of the BPNN model can be achieved. However, the values of the combined network weight and the threshold have great influence on the final diagnosis result; as a simple, effective and easily-realized group intelligent algorithm, the PSO is mainly used for optimizing parameters of a system model and has penetrated into various engineering fields, but the PSO algorithm has the defects of easy premature convergence, low convergence precision, low later iteration efficiency and the like. Therefore, when the diagnostic model is optimized, the advantages of the optimization algorithm and the disadvantages of the diagnostic model are considered, the diagnostic model is improved, and the accurate fault characteristic signal is combined, so that the diagnostic model can accurately and timely diagnose the fault type.
Disclosure of Invention
The invention aims to provide a mechanical fault diagnosis method based on improved PSO-SOM-BPNN, which aims to solve the problems in the prior art and can realize intelligent diagnosis of the fault state of a bearing in mechanical equipment.
In order to achieve the purpose, the invention adopts the following technical scheme:
the mechanical fault diagnosis method based on the improved PSO-SOM-BPNN comprises the following steps;
1) Constructing a fault characteristic data set of the rotating mechanical part: collecting vibration signals of a rotating machine fault part, denoising the vibration signals by utilizing wavelet analysis, extracting energy characteristics through wavelet packet decomposition, optimizing a basis function and the decomposition layer number in the wavelet denoising and the wavelet packet decomposition, establishing a fault characteristic data set, and normalizing the fault characteristic data set by dividing the fault data set into a training set and a test set;
2) Constructing a fault diagnosis model: constructing a fault diagnosis model for the normalized fault characteristic data set in the step 1) by using the BPNN;
3) Constructing a series network fault diagnosis model: using the SOM to serially connect the BPNN fault diagnosis models in the step 2) to construct an SOM-BPNN fault diagnosis model;
4) Parameter optimization of the SOM-BPNN fault diagnosis model: optimizing the weight and the threshold of the SOM-BPNN in the step 3) by adopting a PSO, adjusting the inertia weight and the learning factor of the PSO in a self-adaptive mode, in addition, eliminating a speed item to avoid the influence of the initial particle speed on the convergence speed and the solving precision of the algorithm, and finally forming an ANVTPSO which is used for optimizing the threshold and the weight of the SOM-BPNN to obtain a fault diagnosis model of the ANVTPSO-SOM-BPNN;
5) Mechanical fault diagnosis based on ANVTPSO-SOM-BPNN: verifying by using the optimized fault diagnosis model in the step 4), collecting vibration signals of different types of rotating machinery fault pieces, eliminating noise by using wavelets, extracting features by using wavelet packets, inputting the extracted features into the fault diagnosis model of the ANVTPSO-SOM-BPNN in the step 4), and performing fault classification on the fault diagnosis model to finish mechanical fault diagnosis based on the ANVTPSO-SOM-BPNN.
Further, acquiring a vibration signal of a rotating machinery fault part in the step 1), processing the vibration signal by using wavelet denoising, and determining the optimal wavelet decomposition layer number and the optimal wavelet basis function in the wavelet denoising;
when selecting the wavelet basis function, taking the signal noise power p as the 1 st index selected by the wavelet basis function, and taking the noise power difference delta p h As an index for 2 nd measure of the noise cancellation effect of the signal.
Further, the steps of selecting the optimal wavelet basis function are as follows:
(a) Determining the optimal decomposition layer number of different wavelet basis functions in the acquired vibration signals of the rotating mechanical fault part;
the calculation formula of the optimal decomposition layer number is as follows:
Figure BDA0003124811490000031
in the formula: j is the maximum number of decomposition layers; f. of 0 Is the center frequency of the wavelet basis function; f. of m The minimum frequency for the useful signal; Δ t is the sampling period;
for vibration signals, in particular fault status signals, the frequencies of the useful signals are divided into two categories: 1) Rotational frequency, 2) failure frequency;
(b) Selecting different wavelet basis functions to perform noise elimination on the acquired vibration signals of the rotating mechanical fault part according to the optimal decomposition layer number, and calculating noise power under the different wavelet basis functions;
the formula for calculating the noise power p is:
Figure BDA0003124811490000032
in the formula: q is the data length of the acquired signal; s is an original signal mixed with noise, namely an acquired vibration signal of a rotating mechanical fault piece; s. the 1 The signal is a denoised signal;
(c) For alternative wavelet basis functions, sequencing the wavelet basis functions from small to large according to noise power, and increasing the wavelet basis functions from h = 1;
(d) Sequentially calculating the noise power difference Deltap h ；
Noise power difference Δ p h The calculation formula of (2) is as follows:
Figure BDA0003124811490000041
in the formula: s is h (q) is the denoised signal; h is the wavelet basis function number; when h =0, s 0 (q) is the original noisy signal;
(e) Selecting the minimum Δ p h The wavelet basis function numbered h-1 is the optimal wavelet basis function.
Further, the wavelet packet fault feature extraction in the step 1) specifically comprises the following steps:
the maximum decomposition layer number calculation formula is as follows:
Figure BDA0003124811490000042
in the formula: j is the maximum number of decomposition layers; f. of s Is the sampling frequency; f. of sf Is the signal frequency; z is an integer set;
solving respective energy fluctuation parameters E of the vibration signals of the rotating mechanical fault parts after noise elimination flu And according to the energy fluctuation parameter E flu The fluctuation change rate E' between normal and fault bearings is calculated, the maximum fluctuation change rate in various rotating machinery fault parts is known, the characteristic energy is maximum, namely the fault characteristic is obvious, the corresponding wavelet basis function is the optimal wavelet basis function for decomposing the vibration signal wavelet packet of various rotating machinery fault parts, and the pairExtracting fault characteristics from vibration signals of various rotating mechanical fault parts, and establishing a fault characteristic data set;
energy fluctuation parameter E flu The calculation formula is as follows:
Figure BDA0003124811490000043
the fluctuation rate E' is calculated by the formula:
Figure BDA0003124811490000044
in the formula:
Figure BDA0003124811490000045
n＝0,1,L,2 j -1 is the percentage of energy in each frequency band to the total energy; />
Figure BDA0003124811490000046
Is the maximum ratio;
Figure BDA0003124811490000047
is the minimum occupation ratio; />
Figure BDA0003124811490000048
Is the average ratio; e nor Is the energy fluctuation parameter of the normal signal.
Further, in step 1), the fault feature data set is divided into a training set and a test set, and normalization processing is performed, specifically comprising the following steps:
1.1 The fault feature data is concentrated, and the feature quantity extracted by various fault parts is divided into a training set and a test set according to 2:1;
1.2 Normalized calculation formula is:
Figure BDA0003124811490000051
in the formula:
Figure BDA0003124811490000052
to normalize the eigenvalues, x i Is the original characteristic value, x min Is that the characteristic value is minimumValue, x max Is the maximum value of the eigenvalue.
Further, the step 2) of constructing the fault diagnosis model by using the BPNN on the normalized fault feature data set in the step 1) specifically comprises the following steps:
2.1 Inputting the normalized fault feature training set and the test set into a BPNN fault diagnosis model;
2.2 An input node N, a hidden node L and an output node M in the BPNN model are set, a random value is taken between an initial threshold value and a weight value [ -1,1], and the accuracy of the BPNN fault diagnosis model is tested and constructed.
Further, the construction of the SOM-BPNN fault diagnosis model in the step 3) specifically comprises the following steps:
3.1 SOM realizes the primary classification of the normalized fault feature data set in step 1), and training of a secondary network is carried out on the basis of completing the primary classification of the fault feature data set, wherein the essence of the method is that a dimension is added to an original fault feature data set vector, a new fault feature data set is constructed and is used as the input of a secondary network BPNN, and the newly added dimension is used for marking the classification result of the primary network SOM on the original fault feature data set;
3.2 Dividing the new fault characteristic data set into a new training set and a new test set according to 2:1 according to the characteristic quantity extracted by each fault piece, normalizing, and inputting the normalized training set and test set into a BPNN fault diagnosis model;
3.3 Input nodes, network competition layers and learning step lengths in classification stages in the SOM model are set; in the BPNN model, an input node N, a hidden node L and an output node M take random values with an initial threshold value and a weight value between [ -1,1], and the accuracy of the SOM-BPNN fault diagnosis model is tested.
Further, in the step 4), weight and threshold optimization in the SOM-BPNN is realized through a PSO algorithm, inertia weight and learning factor of the PSO are adjusted through a self-adaptive mode, in addition, in order to avoid influence of particle initial velocity on algorithm convergence speed and solving precision, a velocity term is omitted, and finally an ANVTPSO is formed and used for SOM-BPNN threshold and weight optimization to obtain a final fault diagnosis model, wherein a parameter optimization process based on the ANVTPSO-SOM-BPNN algorithm specifically comprises the following steps:
4.1 Extracting fault characteristics of vibration signals according to the wavelet packets, and setting input nodes, network competition layers and learning step lengths in classification stages in the SOM model; normalizing the classification result obtained by the SOM to be used as a dimension, and forming a new fault characteristic data set with the original fault characteristic data set;
4.2 According to a new fault characteristic data set, an input node N, a hidden node L and an output node M in the BPNN are set, an initial threshold value and a weight value are random values between [ -1,1], and the structure of the SOM-BPNN is determined;
4.3 Initializing PSO, calculating the dimension a of search space, setting the number of population and the maximum iteration number T max ；
4.4 Calculating the fitness value of each particle in the PSO, wherein the fitness function is the mean square error function MSE of the training actual output and the expected output;
4.5 Computing the initial individual optimum position P of the PSO i And global optimum position P g ；
4.6 Updating the PSO position, the inertial weight and the learning factor to obtain an individual and global optimal extreme value, and mapping the PSO position to obtain an optimal weight and a threshold;
the calculation formula of the position is:
X ia (t+1)＝w(t)X ia (t)+c 1 r 1 (P ia (t)-X ia (t))+c 2 r 2 (P ga (t)-X ia (t))
the formula for calculating the inertial weight is:
Figure BDA0003124811490000061
the learning factor is calculated as:
Figure BDA0003124811490000062
wherein w (t) is the inertial weight, t is the number of iterations, c 1 And c 2 Is a learning factor, r 1 And r 2 Is [0,1]A random number in between; w is a max Is the maximum inertial weight, w min For minimum inertial weight, f for particle real timeValue of the objective function, f avg And f min Respectively the average value and the minimum value of all the current particles; 2 is a learning factor c 1 、c 2 An initial value of (1);
4.7 Carry the optimized threshold and weight into SOM-BPNN, and continue tuning until the training target is satisfied.
Compared with the prior art, the invention has the following beneficial technical effects:
1) According to the mechanical fault diagnosis method based on the improved PSO-SOM-BPNN, the number of decomposition layers and wavelet bases in wavelet denoising and wavelet packet decomposition are optimally selected respectively when denoising and fault characteristics of a measured vibration signal are extracted, so that the influence on fault characteristic extraction caused by experience selection is avoided, the extracted characteristics can more accurately reflect the running state of a rotating mechanical part, and further mechanical fault diagnosis is realized.
2) Aiming at the self limitation of the BP neural network, the invention provides the method for using the competition layer result of the SOM neural network as the input data of the BP neural network, constructing the SOM-BPNN fault diagnosis model and analyzing the fault diagnosis performance by means of experiments. Inertia weight and learning factors of the PSO algorithm are adjusted in a self-adaptive mode, and a speed item is abandoned for the standard PSO algorithm so as to avoid influence of the initial speed of the particles on the convergence speed and the solving precision of the algorithm. The method is used for SOM-BPNN threshold value and weight optimization to improve the convergence efficiency and the identification accuracy of the network model. The method is suitable for fault diagnosis of the mechanical equipment rotating member, and can effectively reduce the fault occurrence rate of the mechanical equipment rotating member, thereby improving the working efficiency of the mechanical equipment rotating member.
Drawings
FIG. 1 is a schematic flow chart of a mechanical fault diagnosis method based on an improved PSO-SOM-BPNN in the invention;
FIG. 2 is a flow chart of noise cancellation and feature extraction for a vibration signal of a rotating mechanical part;
FIG. 3 is a diagram of energy ratios of sub-bands of each bearing before and after signal noise cancellation, where (a) is the energy ratio of sub-bands of each bearing before noise cancellation, and (b) is the energy ratio of sub-bands of each bearing after noise cancellation;
FIG. 4 is a block diagram of the SOM-BPNN tandem model of the present invention;
FIG. 5 is a graph comparing the error curves of the ANVTPSO-SOM-BPNN tandem model of the present invention with five neural network models.
Detailed Description
The invention is described in further detail below:
the mechanical fault diagnosis method based on the improved PSO-SOM-BPNN comprises the following steps;
1) Constructing a fault characteristic data set of the rotating mechanical part: collecting vibration signals of a rotating machine fault piece, denoising the vibration signals by utilizing wavelet analysis, extracting energy characteristics through wavelet packet decomposition, optimizing a basis function and decomposition layer number in the wavelet denoising and the wavelet packet decomposition, establishing a fault characteristic data set, and normalizing the fault characteristic data set by dividing the fault characteristic data set into a training set and a test set;
acquiring a vibration signal of a rotating machinery fault part, processing the vibration signal by utilizing wavelet denoising, and determining the optimal wavelet decomposition layer number and the optimal wavelet basis function in the wavelet denoising;
when selecting the wavelet basis function, taking the signal noise power p as the 1 st index selected by the wavelet basis function, and taking the noise power difference delta p h As an index for 2 nd measure of the noise cancellation effect of the signal.
The specific steps of selecting the optimal wavelet basis function in the step 1) are as follows:
(a) Determining the optimal decomposition layer number of different wavelet basis functions in the acquired vibration signals of the rotating mechanical fault part;
the calculation formula of the optimal decomposition layer number is as follows:
Figure BDA0003124811490000081
in the formula: j is the maximum number of decomposition layers; f. of 0 Is the center frequency of the wavelet basis function; f. of m Is the minimum frequency of the useful signal; Δ t is the sampling period;
for vibration signals, in particular fault status signals, the frequencies of the useful signals are divided into two categories: 1) Rotational frequency, 2) failure frequency.
(b) Selecting different wavelet basis functions to perform noise elimination on the acquired vibration signals of the rotating mechanical fault part according to the optimal decomposition layer number, and calculating noise power under the different wavelet basis functions;
the formula for calculating the noise power p is:
Figure BDA0003124811490000082
in the formula: q is the data length of the acquired signal; s is an original signal mixed with noise, namely an acquired vibration signal of a rotating mechanical fault piece; s. the 1 The signal is a denoised signal;
(c) For alternative wavelet basis functions, sequencing the wavelet basis functions from small to large according to noise power, and increasing the alternative wavelet basis functions from h = 1;
(d) Sequentially calculating the noise power difference Deltap h ；
Noise power difference Δ p h The calculation formula of (c) is:
Figure BDA0003124811490000091
in the formula: s h (q) is the denoised signal; h is the wavelet basis function number; when h =0, s 0 (q) is the original noisy signal;
(e) Selecting the minimum Δ p h The wavelet basis function numbered h-1 is the optimal wavelet basis function.
Because the phenomenon of over-noise elimination exists in the noise elimination process, the wavelet basis function with the highest noise power is not considered as the optimal wavelet basis function and only participates in calculation.
In addition, the fault feature extraction specifically comprises the following steps: the number of decomposition layers influences the extraction of fault features, the dimension of a feature vector is also determined, and the number of wavelet packet decomposition layers needs to consider the characteristics of a signal.
The maximum decomposition layer number calculation formula is as follows:
Figure BDA0003124811490000092
in the formula: j is the maximum number of decomposition layers; f. of s Is the sampling frequency; f. of sf Is the signal frequency; z is an integer set;
the purpose of wavelet packet decomposition is to find the fault signature, so the "signal frequency" can be replaced with the fault signature frequency.
Solving respective energy fluctuation parameters E of the vibration signals of the rotating mechanical fault parts after noise elimination flu And according to the energy fluctuation parameter E flu And solving the fluctuation change rate E' between the normal bearing and the fault bearing, wherein the fluctuation change rate is the largest among various rotating mechanical fault parts, the characteristic energy is the largest, namely the fault characteristic is obvious, the corresponding wavelet basis function is the optimal wavelet basis function for decomposing the wavelet packet of the vibration signals of various rotating mechanical fault parts, the fault characteristic is extracted from the vibration signals of various rotating mechanical fault parts, and a fault characteristic data set is established.
Energy fluctuation parameter E flu The calculation formula is as follows:
Figure BDA0003124811490000093
the fluctuation rate E' is calculated by the formula:
Figure BDA0003124811490000101
in the formula:
Figure BDA0003124811490000102
n＝0,1,L,2 j -1 is the percentage of energy in each frequency band to the total energy; />
Figure BDA0003124811490000103
Is the maximum ratio;
Figure BDA0003124811490000104
is the minimum occupation ratio; />
Figure BDA0003124811490000105
Is the average ratio; e nor Is the energy fluctuation parameter of the normal signal.
The normalization process is to effectively reduce the calculation amount of the model and improve the accuracy of the model.
Dividing a fault characteristic data set into a training set and a test set, and carrying out normalization processing, wherein the method comprises the following specific steps:
1.1 ) the fault characteristic data is concentrated, and the characteristic quantity of each fault piece is divided into a training set and a test set according to 2:1;
1.2 Normalized calculation formula is:
Figure BDA0003124811490000106
in the formula:
Figure BDA0003124811490000107
to normalize the eigenvalues, x i Is the original characteristic value, x min Is the minimum value of the eigenvalue, x max Is the maximum value of the eigenvalue.
2) Constructing a fault diagnosis model: constructing a fault diagnosis model for the normalized fault characteristic data set in the step 1) by using BPNN;
the specific steps of constructing the BPNN fault diagnosis model in the step 1) are as follows:
2.1 Inputting the normalized fault feature training set and the test set into a BPNN fault diagnosis model;
2.2 Setting an input node N, a hidden node L and an output node M in the BPNN model, taking a random value between an initial threshold and a weight [ -1,1], and checking and constructing the precision of the BPNN fault diagnosis model.
3) Constructing a series network fault diagnosis model: using the SOM to serially connect the BPNN fault diagnosis models in the step 2) to construct an SOM-BPNN fault diagnosis model;
the SOM-BPNN fault diagnosis model in the step 3) is constructed by the following specific steps:
3.1 SOM realizes the primary classification of the normalized fault feature data set in step 1), and training of a secondary network is carried out on the basis of completing the primary classification of the fault feature data set, wherein the essence of the method is that a dimension is added to an original fault feature data set vector, a new fault feature data set is constructed and is used as the input of a secondary network BPNN, and the newly added dimension is used for marking the classification result of the primary network SOM on the original fault feature data set;
3.2 Dividing a new fault feature data set into a new training set and a new test set according to 2:1 by the feature quantity extracted by each fault piece, normalizing, and inputting the normalized training set and the normalized test set into a BPNN fault diagnosis model;
3.3 Input nodes, network competition layers and learning step length of classification stages in the SOM model are set; in the BPNN model, an input node N, a hidden node L and an output node M take random values with an initial threshold value and a weight value between [ -1,1], and the accuracy of the SOM-BPNN fault diagnosis model is tested.
4) Optimizing parameters of the SOM-BPNN fault diagnosis model: optimizing the weight and the threshold of the SOM-BPNN in the step 3) by adopting a PSO (particle swarm optimization), adjusting the inertia weight and the learning factor of the PSO in a self-Adaptive mode, and in addition, eliminating a speed item to avoid the influence of the initial particle speed on the convergence speed and the solving precision of the algorithm to finally form a self-Adaptive non-speed item PSO (Adaptive No Velocity Term PSO, ANVTPSO) which is used for optimizing the threshold and the weight of the SOM-BPNN to obtain a fault diagnosis model of the ANVTPSO-SOM-BPNN;
step 4) the parameter optimization process based on the ANVTPSO-SOM-BPNN algorithm comprises the following specific steps:
4.1 Extracting fault characteristics of vibration signals according to the wavelet packets, and setting input nodes, network competition layers and learning step lengths in classification stages in the SOM model; normalizing the classification result obtained by the SOM to be used as a dimension, and forming a new fault characteristic data set with the original fault characteristic data set;
4.2 According to a new fault characteristic data set, an input node N, a hidden node L and an output node M in the BPNN are set, an initial threshold value and a weight value are random values between [ -1,1], and the structure of the SOM-BPNN is determined;
4.3 Initialize PSO, calculate its search space dimension a, set population number, maximum number of iterations T max ；
4.4 Calculating the fitness value of each particle in the PSO, wherein the fitness function is the mean square error function MSE of the training actual output and the expected output;
4.5 Computing the initial individual optimum position P of the PSO i And a global optimum position P g ；
4.6 Updating the PSO position, the inertial weight and the learning factor to obtain an individual and global optimal extreme value, and mapping the PSO position to obtain an optimal weight and a threshold;
the calculation formula of the position is as follows:
X ia (t+1)＝w(t)X ia (t)+c 1 r 1 (P ia (t)-X ia (t))+c 2 r 2 (P ga (t)-X ia (t))
the formula for calculating the inertial weight is as follows:
Figure BDA0003124811490000121
/>
the formula for calculating the learning factor is as follows:
Figure BDA0003124811490000122
wherein w (t) is the inertial weight, t is the number of iterations, c 1 And c 2 Is a learning factor, r 1 And r 2 Is [0,1]A random number in between; w is a max Is the maximum inertial weight, w min Is the minimum inertial weight, f is the real-time objective function value of the particle, f avg And f min Respectively the average value and the minimum value of all the current particles; 2 is a learning factor c 1 、c 2 An initial value of (1);
4.7 Carry the optimized threshold and weight into SOM-BPNN, and continue tuning until the training target is satisfied.
5) Mechanical fault diagnosis based on improved PSO-SOM-BPNN: verifying by using the optimized fault diagnosis model in the step 4), collecting vibration signals of different types of rotating machinery fault pieces, eliminating noise by using wavelets, extracting features by using wavelet packets, inputting the extracted features into the fault diagnosis model of the ANVTPSO-SOM-BPNN in the step 4), and performing fault classification on the fault diagnosis model to finish mechanical fault diagnosis based on the ANVTPSO-SOM-BPNN.
The invention is described in further detail in the following examples, which are intended to be illustrative of the invention, but these should not be construed as limiting the scope of the invention, which is defined by the appended claims, any modification which comes within the scope of the invention being covered thereby.
Referring to the attached FIG. 1, the mechanical fault diagnosis method based on the improved PSO-SOM-BPNN of the invention mainly comprises the following steps:
1) The fault simulation platform of the gear transmission system is utilized, and the experiment table consists of a three-phase variable frequency motor, a rotor bearing system, a radial loading device and a parallel shaft gear box which are supported by rolling bearings; the load was simulated by a magnetic particle brake. Selecting a rolling bearing in a bearing seat on the left side of a rotor as a test object, wherein the model of the rolling bearing is NSK-6205; the structural parameters are shown in table 1, the fault size is shown in table 2, and the fault characteristic frequency of the rolling bearing is shown in table 3, wherein an inner ring and an outer ring are processed with cracks by adopting an electric spark forming machine, and a rolling body is subjected to 3-second pitting corrosion by a TH-RFT300 high-speed laser welding machine; setting the rotating speed of a motor to be 1800r/min, carrying out no-load in the radial direction and the axial direction, and acquiring shell vibration by adopting an acceleration sensor, wherein the sensitivity of the sensor is 103mV/g (g is gravity acceleration); during collection, the sampling time is set to be 1s, the sampling rate is set to be 10.24kHz, and 60 groups of vibration acceleration signals are collected, wherein 15 groups of bearings with normal bearing, inner ring fault bearing, outer ring fault bearing and rolling body fault bearing are collected. The bearing data in different states are divided into 5 sets of 10240 point-long subsamples which do not overlap with each other. Therefore, the vibration acceleration signals collected in the four states are divided into 300 groups of sub-samples in total, wherein 75 groups of the four types of bearings divide the signal samples into a training set and a test set by 2:1.
TABLE 1 Rolling bearing construction parameters
Figure BDA0003124811490000131
TABLE 2 Rolling bearing failure size
Figure BDA0003124811490000132
/>
TABLE 3 failure characteristic frequency of rolling bearing
Figure BDA0003124811490000141
The method comprises the steps of selecting 5 common wavelet basis functions including system 8, db3, db4, db5 and db10 as candidate wavelet bases for determining the optimal wavelet decomposition layer number and the optimal wavelet basis function, calculating the optimal decomposition layer number of various bearing signals based on different wavelet basis functions through formula (1) as shown in table 4, and calculating an adaptive threshold value by adopting a stein-based unbiased likelihood estimation principle and finishing signal denoising through a wden function in combination with a soft threshold value in view of different noise intensity of various bearing faults as shown in table 5. Selecting 10 groups of four types of bearing data without data segmentation, performing wavelet de-noising according to the determined uniform decomposition layer number in the table 5, and solving a noise power p from an alternative wavelet basis function according to a formula (2), wherein a representative group of data is as follows, and the result of the noise power p is shown in a table 6; calculating the noise power difference Δ p according to equation (3) h The results are shown in Table 7. Based on the calculation, the optimal wavelet basis functions corresponding to various types of data in different states are determined, the occurrence frequency of various types of wavelet basis functions in 10 groups of data is shown in table 8, and the wavelet basis function with the largest occurrence frequency in different states is taken as the optimal wavelet basis function with unified four types of bearing signals.
TABLE 4 wavelet basis center frequencies
Figure BDA0003124811490000142
The calculation formula of the optimal decomposition layer number is as follows:
Figure BDA0003124811490000143
in the formula: j is the maximum number of decomposition layers; f. of 0 Is the center frequency of the wavelet basis function; f. of m Is the minimum frequency of the useful signal; Δ t is the sampling period;
the "minimum frequency of a useful signal" in the present invention refers to a fault characteristic frequency of a corresponding bearing.
TABLE 5 optimal number of decomposition levels based on wavelet basis
Figure BDA0003124811490000144
Figure BDA0003124811490000151
As can be seen from Table 5, the optimal number of decomposition layers of various bearing signals is different based on different wavelet basis functions. Therefore, in order to retain useful information of four types of bearing vibration signals to the maximum extent, in the selection of the unified decomposition layer number, if the decomposition layer number exceeds 6 layers, the inner ring signal may cause over-noise cancellation to cause loss of the useful information, and the unified hierarchy is taken as 6 through comprehensive consideration.
Noise power p:
Figure BDA0003124811490000152
noise power difference Δ p i ：
Figure BDA0003124811490000153
In the formula: q is the data length of the acquired signal; s is an original signal mixed with noise, namely an acquired vibration signal of a rotating mechanical fault piece; s 1 The signal is a denoised signal; s h (q) is the denoised signal; h is the wavelet basis function number; when h =0, s 0 (q) is the original noisy signal;
TABLE 6 calculated noise Power
Figure BDA0003124811490000154
TABLE 7 noise Power Difference calculation
Figure BDA0003124811490000155
Figure BDA0003124811490000161
TABLE 8 optimal wavelet basis number for bearing data
Figure BDA0003124811490000162
As can be seen from Table 7, the optimal wavelet basis functions corresponding to the four types of bearing signals are db5; the optimal wavelet basis function selected in wavelet denoising herein is db5, as shown in table 8.
Extracting characteristic energy by using a wavelet packet, optimizing a basis function and the number of decomposition layers, calculating by the formula (4) to obtain the number of decomposition layers, selecting 10 groups of four types of bearing data as shown in table 9, decomposing five types of wavelet bases of sym8, db3, db4, db5 and db10 by using the wavelet packet with the number of 3 layers, and calculating by the formula (5) to obtain an energy fluctuation parameter E corresponding to the four types of bearing data flu And the average value of the parameters under various states is calculated, as shown in table 10; equation (6) calculates the fluctuation change rate E' between the normal and failed bearings as shown in table 11.
Optimal number of decomposed layers of wavelet packet:
Figure BDA0003124811490000163
in the formula: j is the maximum number of decomposition layers; f. of s Is the sampling frequency; f. of sf Is the signal frequency; z is an integer set;
TABLE 9 number of wavelet packet decomposition layers
Figure BDA0003124811490000164
As can be seen from Table 9, the optimal value of the number of wavelet packet decomposition layers is 3-5 according to the characteristic frequency of different fault positions of the bearing. When the number of decomposition layers is too small, information of each frequency band cannot be completely decomposed, and bearing characteristic information is not accurately extracted, so that the fault diagnosis accuracy is influenced. Although the fault signal can be analyzed more finely by increasing the number of wavelet packet decomposition layers, the number of decomposed signals increases, and when the number of decomposition layers is too large, the number of dimensions of the feature vector is too large, thereby affecting the fault identification efficiency and the like. Therefore, the number of wavelet packet decomposition layers is 3 in the present invention.
Energy fluctuation parameter E flu ：
Figure BDA0003124811490000171
Rate of change of fluctuation E':
Figure BDA0003124811490000172
in the formula:
Figure BDA0003124811490000173
n＝0,1,L,2 j -1 is the percentage of energy in each frequency band to the total energy; />
Figure BDA0003124811490000174
Is the maximum ratio;
Figure BDA0003124811490000175
is the minimum occupation ratio; />
Figure BDA0003124811490000176
Is the average ratio; e nor Is the energy fluctuation parameter of the normal signal.
TABLE 10 energy fluctuation parameters for four types of bearings
Figure BDA0003124811490000177
TABLE 11 fluctuating Rate of Change of failed bearing
Figure BDA0003124811490000178
As can be seen from Table 11, the largest fluctuation change rate of the three types of failed bearings is db4, and the characteristic energy of the failed bearing is the largest, namely the failure characteristic is obvious. Therefore, comprehensively considered, db4 is selected as the optimal wavelet basis function of the wavelet packet decomposition of the four types of bearing signals in the text.
And (4) normalizing the extracted fault characteristic data by an equation (7).
The normalized calculation formula is as follows:
Figure BDA0003124811490000179
in the formula:
Figure BDA00031248114900001710
to normalize the eigenvalues, x i Is the original characteristic value, x min Is the minimum value of the eigenvalue, x max Is the maximum value of the eigenvalue.
1) And constructing a BPNN fault diagnosis model by adopting the normalized fault characteristic data set.
2) The BPNN excessively depends on the sample, and the SOM is introduced to utilize the characteristic that the SOM does not need a large amount of sample data, so that the training time of the BPNN is reduced, the whole series network is converged at a higher speed, and an SOM-BPNN fault diagnosis model is constructed.
3) And realizing initial classification of input samples aiming at the SOM, adding a dimensionality to a training sample vector according to an initial classification result, and taking a newly formed feature vector as the input of the SOM-BPNN. However, the initial network connection weight and the node threshold of the SOM-BPNN are the same as those of the BPNN, are usually determined empirically, and are prone to fall into a locally optimal solution, so that the convergence efficiency of the network model is limited. Adjusting inertia weight and learning factor of PSO in a self-adaptive mode, avoiding influence of initial particle speed on algorithm convergence speed and solving precision, and eliminating speed items, namely a so-called ANVTPSO new algorithm, wherein the new ANVTPSO algorithm is used for SOM-BPNN threshold value and weight optimization to obtain an ANVTPSO-SOM-BPNN-based mechanical fault diagnosis model, and the parameter optimization process based on the ANVTPSO-SOM-BPNN algorithm specifically comprises the following steps:
4.1 Extracting fault characteristics of vibration signals according to the wavelet packets, and setting input nodes, network competition layers and learning step lengths in classification stages in the SOM model; normalizing the classification result obtained by the SOM, and forming a new fault characteristic data set with the original fault characteristic data set by using the normalized classification result as a dimension;
4.2 According to a new fault characteristic data set, an input node N, a hidden node L and an output node M in the BPNN are set, an initial threshold value and a weight value are random values between [ -1,1], and the structure of the SOM-BPNN is determined;
4.3 Initialize PSO, calculate its search space dimension a, set population number, maximum number of iterations T max ；
4.4 Calculating the fitness value of each particle in the PSO, wherein the fitness function is the mean square error function MSE of the training actual output and the expected output;
4.5 Computing the initial individual optimum position P of the PSO i And a global optimum position P g ；
4.6 Updating the PSO position, the inertial weight and the learning factor to obtain an individual and global optimal extreme value, and mapping the PSO position to obtain an optimal weight value and a threshold value;
the calculation formula of the position is as follows:
X ia (t+1)＝w(t)X ia (t)+c 1 r 1 (P ia (t)-X ia (t))+c 2 r 2 (P ga (t)-X ia (t))
the formula for calculating the inertial weight is:
Figure BDA0003124811490000191
the formula for calculating the learning factor is as follows:
Figure BDA0003124811490000192
wherein w (t) is the inertial weight, t is the number of iterations, c 1 And c 2 Is a learning factor, r 1 And r 2 Is [0,1]A random number in between; w is a max Is the maximum inertial weight, w min Is the minimum inertial weight, f is the real-time objective function value of the particle, f avg And f min Respectively the average value and the minimum value of all the current particles; 2 is a learning factor c 1 、c 2 An initial value of (1);
4.7 Carry the optimized threshold and weight into SOM-BPNN, and continue tuning until the training target is satisfied.
4) And classifying the fault characteristics by adopting ANVTPSO-SOM-BPNN to realize fault diagnosis of the bearing, wherein the ANVTPSO-SOM-BPNN parameters are shown in a table 12.
TABLE 12 ANVTPSO-SOM-BPNN parameters
Figure BDA0003124811490000193
The desired outputs for the four types of bearing signal settings are shown in table 13.
Table 13 fault flag table
Figure BDA0003124811490000201
In order to prove the superiority of the ANVTPSO-SOM-BPNN algorithm constructed by the invention, six different BPNNs are adopted for comparison. Firstly, extracting energy characteristics of signals before and after denoising by performing wavelet denoising and optimizing the number of basis functions and decomposition layers in a wavelet packet, and verifying the accuracy of wavelet denoising to characteristic extraction under an optimized condition by comparing BPNN (Raw Signal BPNN, RSBPNN) of original signals with standard BPNN; secondly, respectively carrying out standard BPNN and PSO-BPNN on the denoised signals, and verifying the superiority of the standard PSO-BPNN compared with the BPNN; then, unified 7-layer decomposition is adopted for four types of bearing signals subjected to noise elimination, 7 layers of PSO-BPNN (Unified Layering PSO-BPNN, ULBPNN) subjected to Unified decomposition are compared with the three schemes BPNN, and the influence of the number of decomposition layers on a diagnosis result in signal noise elimination is researched; comparing the ANVTPSO-BPNN with the four methods, and verifying the influence on the fault diagnosis result after improving the PSO; finally, comparing the ANVTPSO-SOM-BPNN constructed by the invention with the five methods, the superiority is verified, and the bearing fault diagnosis result is shown in a table 14.
TABLE 14 comparison of bearing fault diagnosis results
Figure BDA0003124811490000202
The present invention has been described in further detail with reference to specific examples thereof, which are given by way of illustration and are not to be construed as limiting the scope of the invention, which is defined by the appended claims, as any variation which comes within the scope of the claims is intended to be covered thereby.

Claims (7)

1. The mechanical fault diagnosis method based on the improved PSO-SOM-BPNN is characterized by comprising the following steps;
1) Constructing a fault characteristic data set of the rotating mechanical part: collecting vibration signals of a rotating machine fault part, denoising the vibration signals by utilizing wavelet analysis, extracting energy characteristics through wavelet packet decomposition, optimizing a basis function and the decomposition layer number in the wavelet denoising and the wavelet packet decomposition, establishing a fault characteristic data set, and normalizing the fault characteristic data set by dividing the fault data set into a training set and a test set;
2) Constructing a fault diagnosis model: constructing a fault diagnosis model for the normalized fault characteristic data set in the step 1) by using BPNN;
3) Constructing a serial network fault diagnosis model: using the SOM to serially connect the BPNN fault diagnosis models in the step 2) to construct an SOM-BPNN fault diagnosis model;
4) Optimizing parameters of the SOM-BPNN fault diagnosis model: optimizing the weight and the threshold of the SOM-BPNN in the step 3) by adopting a PSO, adjusting the inertia weight and the learning factor of the PSO in a self-adaptive mode, in addition, eliminating a speed item to avoid the influence of the initial particle speed on the convergence speed and the solving precision of the algorithm, and finally forming an ANVTPSO which is used for optimizing the threshold and the weight of the SOM-BPNN to obtain a fault diagnosis model of the ANVTPSO-SOM-BPNN;
the method comprises the following steps of optimizing weight and threshold in the SOM-BPNN through a PSO algorithm, adjusting inertial weight and learning factor of the PSO in a self-adaptive mode, eliminating a speed item to avoid influence of initial particle speed on algorithm convergence speed and solving precision, and finally forming ANVTPSO which is used for optimizing SOM-BPNN threshold and weight to obtain a final fault diagnosis model, wherein the parameter optimization process based on the ANVTPSO-SOM-BPNN algorithm specifically comprises the following steps:
4.1 Extracting fault characteristics of vibration signals according to the wavelet packets, and setting input nodes, network competition layers and learning step lengths in classification stages in the SOM model; normalizing the classification result obtained by the SOM, and forming a new fault characteristic data set with the original fault characteristic data set by using the normalized classification result as a dimension;
4.2 According to a new fault characteristic data set, an input node N, a hidden node L and an output node M in the BPNN are set, an initial threshold value and a weight value are random values between [ -1,1], and the structure of the SOM-BPNN is determined;
4.3 Initialize PSO, calculate its search space dimension a, set population number, maximum number of iterations T max ；
4.4 Calculating the fitness value of each particle in the PSO, wherein the fitness function is the mean square error function MSE of the training actual output and the expected output;
4.5 Computing the initial individual optimum position P of the PSO i And a global optimum position P g ；
4.6 Updating the PSO position, the inertial weight and the learning factor to obtain an individual and global optimal extreme value, and mapping the PSO position to obtain an optimal weight and a threshold;
the calculation formula of the position is:
X ia (t+1)＝w(t)X ia (t)+c 1 r 1 (P ia (t)-X ia (t))+c 2 r 2 (P ga (t)-X ia (t))
the formula for calculating the inertial weight is:
Figure FDA0004046257360000021
the learning factor is calculated as:
Figure FDA0004046257360000022
wherein w (t) is the inertial weight, t is the number of iterations, c 1 And c 2 As a learning factor, r 1 And r 2 Is [0,1]A random number in between; w is a max Is the maximum inertial weight, w min Is the minimum inertial weight, f is the real-time objective function value of the particle, f avg And f min Respectively the average value and the minimum value of all the current particles; 2 is a learning factor c 1 、c 2 An initial value of (1);
4.7 Carry the threshold and weight after optimizing into SOM-BPNN, continue tuning and optimizing, until meeting and training the goal;
5) Mechanical fault diagnosis based on ANVTPSO-SOM-BPNN: verifying by using the optimized fault diagnosis model in the step 4), collecting vibration signals of different types of rotating machinery fault pieces, eliminating noise by using wavelets, extracting features by using wavelet packets, inputting the extracted features into the fault diagnosis model of the ANVTPSO-SOM-BPNN in the step 4), and performing fault classification on the fault diagnosis model to finish mechanical fault diagnosis based on the ANVTPSO-SOM-BPNN.
2. The improved PSO-SOM-BPNN-based mechanical fault diagnosis method according to claim 1, characterized in that vibration signals of rotating mechanical fault parts are collected in the step 1), wavelet denoising is used for processing the vibration signals, and the optimal wavelet decomposition layer number and the optimal wavelet basis function in the wavelet denoising are determined;
when selecting the wavelet basis function, taking the signal noise power p as the 1 st index selected by the wavelet basis function, and taking the noise power difference delta p h As an index for 2 nd measure of the noise cancellation effect of the signal.
3. The improved PSO-SOM-BPNN-based mechanical fault diagnosis method of claim 2, wherein the optimal wavelet basis function selection step is as follows:
(a) Determining the optimal decomposition layer number of different wavelet basis functions in the acquired vibration signals of the rotating mechanical fault part;
the calculation formula of the optimal decomposition layer number is as follows:
Figure FDA0004046257360000031
in the formula: j is the maximum number of decomposition layers；f 0 Is the center frequency of the wavelet basis function; f. of m Is the minimum frequency of the useful signal; Δ t is the sampling period;
for vibration signals, the frequencies of the useful signal are divided into two categories: 1) Rotational frequency, 2) failure frequency;
(b) Selecting different wavelet basis functions to perform noise elimination on the acquired vibration signals of the rotating mechanical fault part according to the optimal decomposition layer number, and calculating noise power under the different wavelet basis functions;
the formula for calculating the noise power p is:
Figure FDA0004046257360000032
in the formula: q is the data length of the acquired signal; s is an original signal mixed with noise, namely an acquired vibration signal of a rotating mechanical fault piece; s 1 The signal is a signal after noise elimination;
(c) For alternative wavelet basis functions, sequencing the wavelet basis functions from small to large according to noise power, and increasing the wavelet basis functions from h = 1;
(d) Sequentially calculating the noise power difference Deltap h ；
Noise power difference Δ p h The calculation formula of (2) is as follows:
Figure FDA0004046257360000041
in the formula: s is h (q) is the denoised signal; h is the wavelet basis function number; s 0 (q) is the original noisy signal;
(e) Selecting the minimum Δ p h The wavelet basis function numbered h-1 is the optimal wavelet basis function.
4. The improved PSO-SOM-BPNN-based mechanical fault diagnosis method as claimed in claim 1, wherein the wavelet packet fault feature extraction in the step 1) is specifically as follows:
the maximum decomposition layer number calculation formula is as follows:
Figure FDA0004046257360000042
in the formula: j is the maximum number of decomposition layers; f. of s Is the sampling frequency; fsf is the signal frequency; z is an integer set;
solving respective energy fluctuation parameters E of the vibration signals of the rotating mechanical fault parts after noise elimination flu And according to the energy fluctuation parameter E flu Solving the fluctuation change rate E' between normal and fault bearings, wherein the fluctuation change rate is the largest among various rotating mechanical fault parts, the characteristic energy is the largest, namely the fault characteristic is obvious, the corresponding wavelet basis function is the optimal wavelet basis function for decomposing the wavelet packet of the vibration signals of various rotating mechanical fault parts, the fault characteristic is extracted from the vibration signals of various rotating mechanical fault parts, and a fault characteristic data set is established;
energy fluctuation parameter E flu The calculation formula is as follows:
Figure FDA0004046257360000043
the fluctuation rate E' is calculated by the formula:
Figure FDA0004046257360000044
in the formula:
Figure FDA0004046257360000045
the energy of each frequency band accounts for the total energy percentage, wherein n =0,1, …,2 j -1；/>
Figure FDA0004046257360000046
Is the maximum ratio;
Figure FDA0004046257360000047
is the minimum occupation ratio; />
Figure FDA0004046257360000048
Is the average ratio; e nor Is the energy fluctuation parameter of the normal signal.
5. The improved PSO-SOM-BPNN-based mechanical fault diagnosis method according to claim 1, wherein in step 1), the fault feature data set is divided into a training set and a test set, and normalization processing is performed, and the specific steps are as follows:
1.1 ) the fault feature data are concentrated, and the feature quantity of each fault piece is divided into a training set and a testing set according to the ratio of 2: 1;
1.2 Normalized computational formula) is:
Figure FDA0004046257360000051
in the formula:
Figure FDA0004046257360000052
to normalize the eigenvalues, x i Is the original characteristic value, x min Is the minimum value of the eigenvalue, x max Is the maximum value of the eigenvalue.
6. The improved PSO-SOM-BPNN-based mechanical fault diagnosis method as claimed in claim 1, wherein the step 2) of constructing the fault diagnosis model by using the BPNN to the fault feature data set normalized in the step 1) specifically comprises the following steps:
2.1 Inputting the normalized fault feature training set and the test set into a BPNN fault diagnosis model;
2.2 Setting an input node N, a hidden node L and an output node M in the BPNN model, taking a random value between an initial threshold and a weight [ -1,1], and checking and constructing the precision of the BPNN fault diagnosis model.
7. The improved PSO-SOM-BPNN-based mechanical fault diagnosis method as claimed in claim 1, wherein the SOM-BPNN fault diagnosis model construction in the step 3) is specifically as follows:
3.1 Carrying out primary classification on the normalized fault feature data set in the step 1) by the SOM, and training a secondary network on the basis of finishing the primary classification on the fault feature data set, wherein the essence of the method is that a dimension is added to an original fault feature data set vector, a new fault feature data set is constructed and used as the input of a secondary network BPNN, and the newly added dimension is used for marking the classification result of the primary network SOM on the original fault feature data set;
3.2 The new fault feature data set is divided into a plurality of fault feature data sets according to the feature quantity of each fault element, and the data is divided into 2:1, dividing the training set and the test set into a new training set and a new test set, normalizing the training set and the test set, and inputting the normalized training set and the normalized test set into a BPNN fault diagnosis model;
3.3 Input nodes, network competition layers and learning step lengths in classification stages in the SOM model are set; in the BPNN model, an input node N, a hidden node L and an output node M take random values with an initial threshold value and a weight value between [ -1,1], and the accuracy of the SOM-BPNN fault diagnosis model is tested.
CN202110686659.7A 2021-06-21 2021-06-21 Mechanical fault diagnosis method based on improved PSO-SOM-BPNN Active CN113434973B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110686659.7A CN113434973B (en) 2021-06-21 2021-06-21 Mechanical fault diagnosis method based on improved PSO-SOM-BPNN

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110686659.7A CN113434973B (en) 2021-06-21 2021-06-21 Mechanical fault diagnosis method based on improved PSO-SOM-BPNN

Publications (2)

Publication Number Publication Date
CN113434973A CN113434973A (en) 2021-09-24
CN113434973B true CN113434973B (en) 2023-04-07

Family

ID=77756694

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110686659.7A Active CN113434973B (en) 2021-06-21 2021-06-21 Mechanical fault diagnosis method based on improved PSO-SOM-BPNN

Country Status (1)

Country Link
CN (1) CN113434973B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114492196B (en) * 2022-02-14 2024-05-28 瑶声科技(苏州)有限责任公司 Method and system for rapidly detecting faults based on reduced wave energy ratio theory

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112347588A (en) * 2020-11-26 2021-02-09 中国舰船研究设计中心 Rotary machine fault diagnosis method based on wavelet packet decomposition

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106548230B (en) * 2016-10-14 2019-08-06 云南电网有限责任公司昆明供电局 Diagnosis Method of Transformer Faults based on Modified particle swarm optimization neural network
CN107894564B (en) * 2017-11-09 2020-02-18 合肥工业大学 Analog circuit fault diagnosis method based on cross wavelet characteristics
CN108414923A (en) * 2018-02-05 2018-08-17 武汉大学 A kind of analog-circuit fault diagnosis method based on the extraction of depth confidence network characterization
CN108388860B (en) * 2018-02-12 2020-04-28 大连理工大学 Aero-engine rolling bearing fault diagnosis method based on power entropy spectrum-random forest
WO2020000248A1 (en) * 2018-06-27 2020-01-02 大连理工大学 Space reconstruction based method for predicting key performance parameters of transition state acceleration process of aircraft engine
CN109470477B (en) * 2018-09-27 2020-08-25 昆明理工大学 Rolling bearing fault diagnosis method for optimizing FSVM (free space vector optimization) based on improved PSO (particle swarm optimization) algorithm

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112347588A (en) * 2020-11-26 2021-02-09 中国舰船研究设计中心 Rotary machine fault diagnosis method based on wavelet packet decomposition

Also Published As

Publication number Publication date
CN113434973A (en) 2021-09-24

Similar Documents

Publication Publication Date Title
CN107784325B (en) Spiral fault diagnosis method based on data-driven incremental fusion
CN107316046B (en) Fault diagnosis method based on incremental compensation dynamic self-adaptive enhancement
CN111037365B (en) Cutter state monitoring data set enhancing method based on generative countermeasure network
CN110849626B (en) Self-adaptive sparse compression self-coding rolling bearing fault diagnosis system
CN112257530B (en) Rolling bearing fault diagnosis method based on blind signal separation and support vector machine
CN110070060B (en) Fault diagnosis method for bearing equipment
CN111539152B (en) Rolling bearing fault self-learning method based on two-stage twin convolutional neural network
CN102243133B (en) High-speed automaton fault diagnosis method based on movement patterns and impact signal analysis
CN113927371A (en) Cutter wear prediction method based on multi-sensor feature fusion
CN113505655B (en) Intelligent bearing fault diagnosis method for digital twin system
CN111562108A (en) Rolling bearing intelligent fault diagnosis method based on CNN and FCMC
CN113188794B (en) Gearbox fault diagnosis method and device based on improved PSO-BP neural network
CN110969096A (en) Motor fault mode diagnosis method based on particle swarm optimization support vector machine
CN113434973B (en) Mechanical fault diagnosis method based on improved PSO-SOM-BPNN
CN114429152A (en) Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption
CN115146687B (en) Fault feature extraction method based on second-order variable-scale parameter self-matching stochastic resonance
CN115545070A (en) Intelligent diagnosis method for unbalance-like bearing based on comprehensive balance network
CN115587290A (en) Aero-engine fault diagnosis method based on variational self-coding generation countermeasure network
CN114417924B (en) Rolling bearing fault diagnosis method based on undirected graph adjacent matrix of mixed features
CN115683604A (en) Engineering machinery fault diagnosis method based on unbalanced data
CN114154254A (en) Fault diagnosis method for electric actuator of gas turbine
CN114964783B (en) Gearbox fault detection model based on VMD-SSA-LSSVM
CN114897016A (en) Fan bearing fault intelligent diagnosis method based on multi-source frequency spectrum characteristics
CN111360578A (en) Method for identifying lubricating state of ball screw pair
CN117407698B (en) Hybrid distance guiding field self-adaptive fault diagnosis method

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