CN110348468B - Bearing sub-health identification method for optimized strong reconstruction edge noise reduction automatic encoder of correlation vector machine - Google Patents
Bearing sub-health identification method for optimized strong reconstruction edge noise reduction automatic encoder of correlation vector machine Download PDFInfo
- Publication number
- CN110348468B CN110348468B CN201910388451.XA CN201910388451A CN110348468B CN 110348468 B CN110348468 B CN 110348468B CN 201910388451 A CN201910388451 A CN 201910388451A CN 110348468 B CN110348468 B CN 110348468B
- Authority
- CN
- China
- Prior art keywords
- input
- data
- noise reduction
- output
- vector machine
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
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
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- 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
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Acoustics & Sound (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
- Complex Calculations (AREA)
Abstract
A bearing sub-health identification method of an optimized strong reconstruction edge noise reduction automatic encoder of a correlation vector machine comprises the following steps: (1) signal sampling; (2) denoising; (3) feature extraction; (4) fault diagnosis division; and (5) diagnosing the state of the mechanical equipment. The invention uses the mode of adding the weights of the two kernel functions to replace the single action of the original single kernel function, and improves the correlation vector machine. And solving the variance of the input sample mapped into the feature space under the combined action of the Fisher criterion and the maximum entropy criterion, and determining the weight coefficient of the kernel function. And initializing parameters according to the obtained weight coefficient, estimating hyper-parameters, and finally obtaining the rolling bearing sub-health recognition classification. The invention solves the problems of limited constraint conditions, weak data compression capability, large reconstruction error and weak feature data expression capability of an edge noise reduction automatic encoder, and provides a bearing sub-health identification method for optimizing a strong reconstruction edge noise reduction automatic encoder by a correlation vector machine, which has high identification accuracy.
Description
Technical Field
The invention relates to a bearing sub-health recognition method of an optimized strong reconstruction edge noise reduction automatic encoder of a correlation vector machine, belonging to the field of fault diagnosis of rolling bearings.
Background
The development of modern industrial level is greatly developed along with the improvement of technological level, and the production equipment is replaced by intellectualization instead of the situation of originally relying on manpower. In the process, the requirements on the equipment are continuously increased, the damage of any part can have important influence on the production flow, and the rolling bearing is used as a core component of a rotating mechanical element and is one of the elements commonly used in the mechanical equipment. If the current state of the bearing can be accurately identified, various losses caused by the current state can be avoided to a great extent, and because the fault of the bearing is not caused instantaneously but is a process which is accumulated continuously, the research on the sub-health identification of the bearing is paid extensive attention by experts and scholars.
Disclosure of Invention
In order to solve the existing problems, a bearing sub-health identification method of a correlation vector machine optimized strong reconstruction edge noise reduction automatic encoder is provided.
The invention adopts the technical scheme that: the method comprises the following steps:
1) Signal sampling: respectively placing an acceleration sensor in a twelve-o-clock direction of a motor shell and a fan, processing a vibration signal acquired by a 16-channel data recorder by using MATLAB, and collecting a signal at a sampling frequency of 12KHz, wherein the sampling frequency of a signal at a driving end is 48000S/S;
2) And (3) noise reduction treatment: preprocessing a signal by adopting Fourier transform, and normalizing the signal by using a minimum value and a maximum value after the transform, so that the numerical value fluctuation range is stabilized at [0,1] on the basis of keeping the original relative relation of data;
3) Characteristic extraction: determining the number of hidden layers of the edge noise reduction automatic encoder to be three, and inputting the preprocessed data into the strong reconstruction edge noise reduction automatic encoder; extracting features after mapping under the action of improved constraint conditions, wherein the input of a hidden layer is the output of the previous hidden layer plus preprocessed data, and calculating the actual output of the hidden layer; calculating partial derivatives of the loss function to the weight and the bias on the basis of input and output, and updating the weight and the bias; comparing the set loss function with the maximum set times, stopping when the expected value is reached, or recalculating the output of the hidden layer;
4) Fault diagnosis and division: dividing fault diagnosis by using an improved correlation vector machine;
5) And (3) diagnosing the state of the mechanical equipment: the state of the bearing data is identified using the measured bearing data.
In the step 3), the specific method is as follows:
3.1 Preprocessing input data, setting the processed input value and output expected value as x;
3.2 Random initialization weights w and offsets b, using cross entropy to adjust the parameters w, b, defining a loss function:
wherein: d is the input dimension;
a represents the size after batch division;
ik denotes the kth element of the ith input in each batch of data;
summing the input vectors first and then summing each sample of the batch data, the autoencoder constraints are as follows (2):
wherein: n represents the input number;
m represents the number of hidden layer neurons;
z represents a hidden layer feature expression;
l represents a cross entropy function;
β represents a weight coefficient;
ρ represents the degree of activation of each layer of neurons;
3.3 Define the output of the previous hidden layer as x ', then the input of the next hidden layer is (x' + x);
3.4 Computing the actual output z of the output layer processed by the hidden layer;
3.5 On the basis of input and output, partial derivatives of the loss function pair w and b are lost, a scalar form is used in the reconstruction process, and the calculation mode is as follows:
wherein w qp Representing the weight of the nodes q and p;
w kp representing the weight of the nodes k and p;
b yp represents the bias of node p;
b zk represents the bias of node k;
for the ith sample of each batch of data,an input representing the pth hidden layer,is shown ask inputs of the output layer;
f' (x) is the first derivative of sigmoid: f' (x) = f (x) [1-f (x) ]
Calculating a partial derivative of the loss function formula (1) to obtain a formula (6):
wherein w rs Representing the weight of the nodes r and s;
b yr represents the bias of node r;
3.6 Update parameter values according to the rule of equation (7), where η is learning rate, w is weight, b y Is the offset:
3.7 Compare the set loss function with the maximum set number of times, stop when reaching our expectation, otherwise return to 3.6) continue training.
3.8 The output of the improved edge denoise auto-encoder is input to an improved correlation vector machine for classification.
In the step 4, the specific method is as follows:
4.1 Is provided with two types of input samples X 1 ,X 2 ,
X 1 (x 11 ,x 12 ,...x 1i ),X 2 (x 21 ,x 22 ,...x 2i )(i=1,2,...l 1 ,j=1,2,...l 2 ),μ 1 ,μ 2 Is X 1 ,X 2 The mapped mean vector is calculated by (8) and (9):
the intra-class scatter matrix is solved by equations (10), (11):
wherein K (,) is a kernel function;
K 1 ,K 2 ,K 11 ,K 22 ,K 12 the formula is as follows:
mapping of input samples to variances in feature spaceThe calculation mode is as the formulas (17) and (18):
under the constraint action of the maximum entropy method, the variance of input sample mapping in the feature space is solvedThe weight coefficient λ proposed herein is calculated as formula (19):
4.2 According to the improved kernel function mode, the structure of the correlation vector machine is defined as follows:
weight w = (w) 0 ,...,w N ),K(x,x i ) The expression is as in formula (21):
solving for variances in input sample mappings into feature space under Fisher criterion constraintsSolving for variances in input sample mappings into feature space under maximum entropy criterion constraintsThe kernel function weight coefficient λ is calculated according to the proposed method, in the way as equation (22):
4.3 Hyper-parameter α initialization, weight covariance Σ as in equation (23):
Σ=-H=(φ T Bφ+A) -1 (23)
wherein H is Hessian matrix, and y = [ y = 1 ,y 2 ,Λ,y N ] T ,B=diag(β 1 ,β 2 ,Λ,β N ),β n =σ[y(x n )]{1-σ[y(x n )]},A=diag(α 0 ,α 1 ,Λ,α N );
4.4 Estimate the hyperparameters, by the formula (24):
μ i is X i The mapped mean vector;
4.5 Correlation vector machine classification generally adopts Sigmoid probability function, and its expression form is as follows (25):
wherein: y (x, w) MP )=φ(x)w MP ,w MP Calculated by the formula (26):
Solving to obtain a probability predicted value p (t =1 calx) of the sample pair category, p (t =0 calx), wherein the category corresponding to the larger value of the values is the recognition result of x.
The beneficial effects of the invention are as follows: the method is characterized in that sparsity constraint conditions are added on the basis of original edge noise reduction constraint conditions, the input of a hidden layer is changed into the output of a previous hidden layer and preprocessed data are added, and the problems that an edge noise reduction automatic encoder is limited in constraint conditions, weak in data compression capacity, large in reconstruction error and weak in feature data expression capacity are solved. When the construction information in the characteristic sample data is different and the distribution is uneven, the original single kernel function is replaced by a mode of adding the weights of the two kernel functions, the variance of the input sample mapped into the characteristic space is solved under the combined action of the Fisher criterion and the maximum entropy criterion by using the proposed method, and the weight coefficient of the kernel function is determined. And (4) solving a weight coefficient by the proposed calculation mode, then initializing parameters, estimating hyper-parameters, and finally obtaining sub-health identification classification. The sub-health recognition accuracy of three methods of a correlation vector machine edge noise reduction automatic encoder (MDAE-RVM), a correlation vector machine strong reconstruction edge noise reduction automatic encoder (IMDAE-RVM) and a correlation vector machine optimization strong reconstruction edge noise reduction automatic encoder (IMDAE-IRVM) is compared by utilizing measured bearing data, the sub-health recognition accuracy of the correlation vector machine optimization strong reconstruction edge noise reduction automatic encoder (IMDAE-IRVM) is higher than that of the other two methods, and the improved method can correctly recognize the state of the bearing data so as to verify the effectiveness of the invention.
Description of the drawings:
fig. 1 is a diagram of an improved constraint MDAE topology.
FIG. 2 is a diagram of an IMDAE topology with improved hidden layer input.
FIG. 3 is a flow chart of a method IMDAE-IRVM.
Fig. 4a is a normal state driving end time domain vibration signal.
Fig. 4b is a time domain vibration signal of the inner ring fault driving end.
FIG. 4c is a ball failure drive end time domain vibration signal.
FIG. 4d is a time domain vibration signal of a faulty driving end of the outer ring.
Fig. 5a is a normal state drive-end time domain vibration signal after fourier transform.
Fig. 5b is a time-domain vibration signal of the failed driving end of the inner ring after fourier transform.
FIG. 5c is a Fourier transformed ball fault drive-end time domain vibration signal.
Fig. 5d is the outer ring fault drive end time domain vibration signal after fourier transform.
Figure 6a is a graph of a ball drive end vibration signal before fourier transformation for a sample section.
Figure 6b is a graph of ball drive end vibration signal after fourier transformation of a portion of the sample.
Fig. 7a is a graph of the inner ring drive end vibration signal before fourier transform of a portion of a sample.
Fig. 7b is a graph of the inner ring drive end vibration signal after fourier transformation of a portion of the sample.
Fig. 8a is a graph of the outer ring drive end vibration signal before fourier transform of a portion of the sample.
Figure 8b is a graph of the outer ring drive end vibration signal after fourier transformation of a portion of the sample.
Figure 9a is a graph of the normal drive end vibration signal before fourier transformation of a portion of the sample.
Figure 9b is a graph of the normal drive end vibration signal after fourier transformation of a portion of the sample.
FIG. 10 is a graph of the error variation of the hidden layer number with increasing neuron number.
Fig. 11 is a mean square error trend graph.
Detailed Description
A bearing sub-health identification method for an optimized strong reconstruction edge noise reduction automatic encoder of a correlation vector machine comprises the following steps:
1) Signal sampling: the acceleration sensor is respectively placed in the twelve o' clock direction of the motor shell and the fan, the MATLAB processes vibration signals collected by the 16-channel data recorder, signals are collected at the sampling frequency of 12KHz, and the sampling frequency of signals at the driving end is 48000S/S.
2) And (3) noise reduction treatment: and preprocessing the signals by adopting Fourier transform, and normalizing the signals by using a minimum value and a maximum value after the Fourier transform, so that the numerical fluctuation range is stabilized at [0,1] on the basis of keeping the original relative relationship of the data.
3) Feature extraction: determining the number of hidden layers of an edge noise reduction automatic encoder to be three, wherein the number of neurons of the hidden layers is 400, inputting preprocessed data into a strong reconstruction edge noise reduction automatic encoder (IMDAE), extracting characteristics after mapping under the action of improved constraint conditions, inputting the hidden layers by adding the preprocessed data to the output of the previous hidden layer, calculating the actual output of the hidden layers, calculating partial derivatives of a loss function to weight and bias on the basis of input and output, updating the weight and bias, comparing the set loss function with the maximum set number of times, stopping when the expected value is reached, or recalculating the output of the hidden layers. Wherein, the number of the neurons of the hidden layer is obtained when the best effect can be obtained by testing and adjusting.
The specific method comprises the following steps:
3.1 Pre-processing the input data, and setting the processed input value and output expected value as x.
3.2 Random initialization weights w and offsets b, using cross entropy to adjust the parameters w, b, defining a loss function:
wherein: d is the input dimension;
a represents the size after batch division;
ik represents the kth element of the ith input in each batch of data;
summing the input vectors first and then summing each sample of the batch data, the autoencoder constraints are as follows (2):
wherein: n represents the input number;
m represents the number of hidden layer neurons;
z represents a hidden layer feature expression;
l represents a cross entropy function;
β represents a weight coefficient;
ρ represents the degree of activation of neurons in each layer.
3.3 Define the output of the previous hidden layer as x ', then the input of the next hidden layer is (x' + x).
3.4 Compute the actual output z of the output layer processed through the hidden layer.
3.5 On the basis of input and output, partial derivatives of the loss function to w and b are lost, a scalar form is used in the reconstruction process, and the calculation mode is as follows:
wherein, w qp Representing the weight of the nodes q and p;
w kp representing the weight of the nodes k and p;
b yp represents the bias of node p;
b zk represents the bias of node k;
for the ith sample of each batch of data,represents the input of the pth hidden layer,represents the input of the kth output layer;
f' (x) is the first derivative of sigmoid: f' (x) = f (x) [1-f (x) ]
Calculating a partial derivative of the loss function formula (1) to obtain a formula (6):
wherein w rs Representing the weight of the nodes r and s;
b yr representing the bias of node r.
3.6 Update the parameter values according to the rule of equation (7), where η is the learning rate, w is the weight, b y Is the offset:
3.7 Compare the set loss function with the maximum set number of times, stop when reaching our expectation, otherwise return to 3.6) continue training.
3.8 The output of the improved edge noise reduction auto-encoder is input to an improved correlation vector machine for classification.
4) Fault diagnosis and division: dividing fault diagnosis by using an improved correlation vector machine; on the basis of a sub-health Identification Method (IMDAE) of a strong reconstruction edge noise reduction automatic encoder, a correlation vector machine is improved, an original single kernel function is replaced by a mode of adding weights of two kernel functions, and a weight coefficient in an improved kernel function is solved by using sample variance in a feature space under the constraint of a Fisher criterion and a maximum entropy criterion, so that the performance of processing data by the kernel function can be improved, and the identification accuracy rate is improved.
The specific method comprises the following steps:
4.1 Is provided with two types of input samples X 1 ,X 2 ,
X 1 (x 11 ,x 12 ,...x 1i ),X 2 (x 21 ,x 22 ,...x 2i )(i=1,2,...l 1 ,j=1,2,...l 2 ),μ 1 ,μ 2 Is X 1 ,X 2 The mapped mean vector is calculated by (8) and (9):
the intra-class scatter matrix is solved by equations (10), (11):
wherein K (,) is a kernel function;
K 1 ,K 2 ,K 11 ,K 22 ,K 12 the formula is as follows:
mapping of input samples to variances in feature spaceThe calculation mode is as the formulas (17) and (18):
under the constraint action of the maximum entropy method, the same method is used for solvingSolving the variance in the mapping of input samples into feature spaceThe weight coefficient λ proposed herein is calculated as formula (19):
4.2 According to the improved kernel function mode, the structure of the correlation vector machine is defined as follows:
weight w = (w) 0 ,...,w N ),K(x,x i ) The expression is as in formula (21):
solving for variances in input sample mappings into feature space under Fisher criterion constraintsSolving for variances in input sample mappings into feature space under maximum entropy criterion constraintsThe kernel function weight coefficient λ is calculated according to the proposed method, in the way as shown in equation (22):
4.3 Hyper-parameter α initialization, weight covariance Σ as in equation (23):
Σ=-H=(φ T Bφ+A) -1 (23)
wherein H is Hessian matrix, and y = [ y = 1 ,y 2 ,Λ,y N ] T ,B=diag(β 1 ,β 2 ,Λ,β N ),β n =σ[y(x n )]{1-σ[y(x n )]},A=diag(α 0 ,α 1 ,Λ,α N )。
4.4 Estimate the hyperparameters, by the formula (24):
μ i is X i The mapped mean vector;
4.5 Correlation vector machine classification generally adopts Sigmoid probability function, and its expression form is as follows (25):
wherein: y (x, w) MP )=φ(x)w MP ,w MP Calculated by equation (26):
solving to obtain a predicted probability value p (t =1 calx) of the sample pair category, p (t =0 calx), wherein the category corresponding to the larger value in the values is the recognition result of x.
5) And (3) diagnosing the state of the mechanical equipment: the state of the bearing data is identified using the measured bearing data.
Example 1:
according to statistics, more than 70% of faults of the rolling bearing are expressed in a vibration mode. Generally, the rolling bearing has many reasons for faults, but it is not late as long as the problems can be found before the faults occur and measures can be taken in time, and common fault problems mainly include inner ring faults, outer ring faults and ball faults, and have great significance in monitoring the health condition of the rolling bearing. Therefore, the experimental data is actually measured data of the rolling bearing, the real state in work is reflected, and the data distribution conditions under various fault states can be seen more clearly by comparing the data before and after preprocessing.
The proposed method selects a vibration signal for fault feature extraction. Aiming at the situation that rolling bearing signals contain a large number of complex signal features, a traditional feature extraction method usually depends on manual processing, and meanwhile a special preprocessing process is needed for vibration signals, so that a sub-health recognition method of a strong reconstruction edge noise reduction automatic encoder improved by a correlation vector machine is provided, the constraint conditions and the input of a hidden layer of a prototype automatic encoder are improved, and data reconstructed by the strong reconstruction edge noise reduction automatic encoder (IMDAE) are input to the improved correlation vector machine for sub-health recognition. The problems that the construction information in the characteristic sample data is different and uneven in distribution and the sample data with a complex data structure cannot be effectively processed by using a single kernel function can be well solved.
1. The theoretical basis of the scheme of the invention is as follows:
1. automatic encoder
The automatic encoder as an unsupervised learning neural network model has a three-layer neural network structure comprising an input layer, a hidden layer and an output layer. It can be divided into two parts, an encoder and a decoder. Compared with other three-layer neural networks, the form is the same, but partial limitation is made on input and output, high-dimensional input data can be converted into low-dimensional hidden layer nodes through an encoder part, the low-dimensional hidden layer is reconstructed into original input data through a decoder part, and another expression of the data can be obtained through the hidden layer due to the fact that the hidden layer reconstructs an input signal.
The automatic encoder operates as follows:
(1) X is input, and the output desired value is set to x.
(2) And (3) constructing an encoder and a decoder, and randomly initializing to generate a weight w and an offset b, b'.
(3) The respective hidden layers are calculated from the given input and output, and the layer output is output.
(4) Using the inputs and outputs, an error function is calculated to calculate the partial derivatives of each neural unit.
(5) The weight w and the offsets b, b' are updated.
(6) The training is stopped when the desired conditions are reached by comparing the set loss function with the maximum set number. Otherwise, selecting the next learning sample, returning to the step (3), and entering the next round of learning.
The noise reduction automatic encoder can effectively remove the noise of original features as the improvement of the automatic encoder, and has better effect on feature extraction. The edge noise reduction automatic encoder overcomes the defects of the noise reduction automatic encoder in the aspects of calculated amount, processing time and the like in a mode of marginalizing noise interference. The most key part is the expansion mode of the loss function, the aim of approximating the expected loss function of the model can be achieved by using the Taylor expansion mode, and the sensitivity problem caused by the input vector to the hidden layer neuron is fully considered.
The strong reconstruction edge noise reduction automatic encoder solves the problems that the edge noise reduction automatic encoder is weak in data compression capability, limited in constraint condition capability and large in error before and after reconstruction, on one hand, sparsity constraint conditions are added on the basis of the constraint conditions of the edge noise reduction automatic encoder, the defects of the edge noise reduction automatic encoder in the aspects of the constraint conditions and data compression are overcome, and the MDAE topological structure of the improved constraint conditions is shown in an attached drawing 1; on the other hand, the input of the hidden layer is improved into the output of the previous hidden layer and the preprocessed data are added, the expression capability of a characteristic sample is enhanced, the reconstruction error is reduced, and the MDAE topological structure input by the hidden layer is improved as shown in the attached figure 2; and finally, identifying the output of an automatic coder (IMDAE) for strong edge noise reduction by using a correlation vector machine (SVM).
2. Improved IMDAE-RVM training process
The operation flow of IMDAE-RVM is shown in the attached figure 3, and the specific operation is as follows:
(1) Input data is preprocessed using fourier transform and normalization methods.
(2) Defining an expected value of a strong reconstruction edge noise reduction automatic encoder (IMDAE), randomly initializing a weight and a bias, setting a constraint condition of the IMDAE, taking preprocessed data as input of the strong reconstruction edge noise reduction automatic encoder (IMDAE), mapping under the action of the constraint condition, extracting features, calculating actual output of a hidden layer by adding the preprocessed data to the output of a previous hidden layer, calculating a partial derivative of a loss function to the weight and the bias on the basis of input and output, updating the weight and the bias, comparing the set loss function with the maximum set number of times, stopping when the expected value is reached, or recalculating the output of the hidden layer.
(3) Inputting data reconstructed by a strong reconstruction edge noise reduction automatic encoder (IMDAE) into a correlation vector machine for improving a kernel function to perform sub-health identification, firstly defining the structure of the correlation vector machine, wherein the kernel function is the sum of a Gaussian kernel function multiplied by a weight coefficient and a linear kernel function multiplied by the weight coefficient, calculating the variance of a sample mapped into a sample space under the constraint of a Fisher criterion and a maximum entropy criterion for determining the weight coefficient of the kernel function, solving the weight coefficient by a proposed calculation mode, initializing parameters, estimating hyper-parameters, and finally obtaining sub-health identification classification.
The technical scheme of the invention comprises the following implementation processes:
1. signal acquisition: the experimental data acquisition work is completed through an acceleration sensor, the sensor is respectively placed in twelve o' clock directions of a motor shell and a fan, MATLAB processes vibration signals acquired by a data recorder 16 channel, signals are collected under the sampling frequency of 12KHz, the sampling frequency of a drive end signal is 48000S/S, data are divided into 0HP and 3HP loads, a bearing is in a normal state, an outer ring fault, an inner ring fault and a rolling fault are respectively in corresponding rotating speed conditions of 1797r/min and 1730r/min, the fault diameter is selected to be 0.007mm, and vibration signals of four faults are listed as shown in figures 4 a-d. The experimental data were selected as shown in table 1:
TABLE 1 eight groups of bearing Experimental data parameters
2. And (3) noise reduction treatment: since the neural network is not able to process time series under normal conditions and the vibration signal of the rolling bearing is mostly a regular periodic signal, the vibration signal of the rolling bearing is preprocessed by fourier transform, and the converted signals are shown in fig. 5 a-d. And then, carrying out corresponding transformation and then carrying out normalization processing by using the minimum value-maximum value, so that the numerical value fluctuation range is stabilized in [0,1] on the basis of keeping the original relative relationship of the data, and the formula is as shown in formula (27):
here, the minimum value min A Max, max A V is the original of the attribute a and v' is the normalized value. As can be seen from the above equation, the maximum value and the minimum value of the characteristic values correspond to 1 and 0, respectively. In order to more clearly see the difference between the Fourier processing and the Fourier processing, the data (1000) of a part of the experiments are intercepted and compared, and the results are respectively shown in the attached figures 6a-b, 7a-b, 8a-b and 9 a-b.
3. Feature extraction: aiming at the selection of the experimental model topological structure, comparing the change conditions of the number of hidden layers of one layer, two layers and three layers, which are increased along with the number of neurons, and the error percentage of the sample set, as shown in the attached drawing 10, within a certain range, the error of the sample set is in a descending trend along with the increase of the number of neurons under the condition that the number of hidden layers is not changed, for the two layers of hidden layers, the error is the lowest when the number of neurons reaches 500 and is about 0.07, compared with the two layers of hidden layers, the three layers of hidden layers reach the lowest point when the number of neurons is 400, and the error is about 0.03, so that in order to ensure the accuracy of the final experiment, the number of hidden layers of the edge noise reduction automatic encoder selected in the experiment is three layers, and the number of neurons in the hidden layers is 400.
4. Fault diagnosis and division: different faults can be more effectively divided through the improved RVM, aiming at the improved kernel function, the weight coefficient of the kernel function needs to be calculated, a method for calculating the weight coefficient according to the variance of the input sample in the characteristic space is provided, and the variance of the input sample in the characteristic space under the constraint of the Fisher criterion is calculated in the table 2And the variance of the input samples in the feature space under the constraint of the maximum entropy methodLambda =0.52, 1-lambda =0.48 was calculated.
TABLE 2 feature space sample variance and weight coefficient under different constraints
5. And (3) diagnosing the state of the mechanical equipment: selecting four experimental results to calculate a mean value, comparing and analyzing the identification accuracy rates of different methods, wherein the table 3 shows the accuracy rate comparison before and after the improvement of the method, the sub-health identification method of the MDAE-RVM is characterized in that the sub-health identification method is input through the reconstruction of an edge noise reduction automatic encoder (MDAE) and then is classified by using a Related Vector (RVM), the sub-health identification accuracy rates of the four experiments are 85.6%, 80.3%, 82.7% and 83.1%, and the mean value of the four experiments is 82.9%; according to the IMDAE-RVM sub-health identification method, input information is reconstructed by using IMDAE, then a correlation vector machine (RVM) is used for classification, the accuracy of four experiments is respectively 90.3%, 90.7%, 89.9% and 91.4%, the mean value of the four experiments is 90.55%, and compared with an MDAE-RVM method, the improved IMDAE-RVM method is improved by 7.65 percentage points in identification accuracy; the sub-health identification method of the IMDAE-IRVM is characterized in that a single kernel function is improved on the basis of a strong reconstruction edge noise reduction automatic encoder, the sample variance of a feature space is used for solving the weight coefficient of the improved kernel function, the identification accuracy rates of four experiments are respectively 94.2%, 95.7%, 94.6% and 94.3%, the four mean values are 94.7%, and compared with the IMDAE-RVM, the sub-health identification method of the IMDAE-IRVM is improved by 11.8 and 4.15 percentage points in accuracy rate. The trend of the mean square error with increasing number of iterations is shown in figure 11.
TABLE 3 method before and after improvement classification accuracy comparison
Table 4 shows the comparison of the accuracy of the bearing sub-health recognition by several different improved methods, and it can be seen from the table that the sub-health recognition method (IMDAE-IRVM) for the RVM optimized strong reconstruction MDAE improves the accuracy of the particle swarm optimized BP neural network (IPSO-BP), the ant colony optimized support vector machine (IACO-SVM) and the genetic optimized BP neural network (IGA-BP) by 7.2%, 2.2% and 3.46%, respectively.
TABLE 4 comparison of accuracy rates for different improved methods
The data used in the experiment is actually measured bearing data, and the data of a drive end with the sampling frequency of 12KHz, the bearing load of 0HP and the experiment rotating speed of 1797r/min is taken as a normal state; the data of a sampling frequency of 12KHz, a bearing load of 3HP and an experimental rotating speed of 1730r/min are taken as the sub-health state; the sampling frequency is 12KHz, the bearing load is 0HP, the experimental rotating speed is 1797r/min, and the fault diameter selects the drive end fault data of 0.007mm as the fault state. The output data is binary coded, i.e. the normal state is [ 10 ], the sub-health state is [ 00 ], and the fault state is [ 01 ]. The results of the experiment are shown in Table 5, 6:
TABLE 5 comparison of the results
As can be seen from Table 5, the expected output of the second set of data for the test sample is the normal state of [ 10 ], the output of the sub-health recognition method using IMDAE-RVM is [ 0.2062.6535.0001 ], it can be seen from the data that the sub-health state should be classified, which is not the expected output, and the output of the sub-health recognition method using IMDAE-IRVM is [0.79030.0134 0.0988], it can be seen from the data that the classification into the normal state is the expected output, indicating that the improved method is correctly recognized.
TABLE 6 comparison of the results
From Table 6, it can be seen that the expected output of the seventh set of data of the test sample is a fault state of [ 01], the output of the sub-health identification method using IMDAE-RVM is [ 0.0245.5783 0.1591], the data shows that the sub-health state should be classified, which is not the expected output, the output of the sub-health identification method using IMDAE-IRVM is [ 0.0351.2769.0.8984 ], and the data shows that the classification into the fault state is the expected output, which indicates that the improved method is correctly identified. In addition, when the fifth group of data of the test sample is observed, the sub-health state output is expected to be [ 0], the sub-health recognition method output using the IMDAE-RVM is [ 0.3826.4427.0032 ], the sub-health recognition method output using the IMDAE-IRVM is [ 0.0001.6352.0093 ], and as can be seen from the data, both methods can successfully recognize the test sample, but the output values of the IMDAE-RVM method, namely 0.3826 and 0.4427, are not greatly different, so that the problem of inaccurate classification is easily caused, and compared with the IMDAE-RVM method, the difference between the values of the output values of the IMDAE-IRVM method, namely [ 0.0001.6352.0093 ], is large, so that the situation of wrong classification is not easily caused.
By comparing tables 5 and 6, the fact that the correct state of the test sample can be accurately identified by the IMDAE-IRVM method compared with the IMDAE-RVM method, the size difference between the numerical values is obvious, the situation of wrong classification cannot occur, and the effectiveness of the sub-health identification method (IMDAE-IRVM) for optimizing the strongly-reconstructed MDAE by the RVM is proved.
Claims (1)
1. A bearing sub-health identification method for a correlation vector machine optimized strong reconstruction edge noise reduction automatic encoder is characterized by comprising the following steps:
1) Signal sampling: respectively placing an acceleration sensor in a twelve-o-clock direction of a motor shell and a fan, processing a vibration signal acquired by a 16-channel data recorder by using MATLAB, and collecting a signal at a sampling frequency of 12KHz, wherein the sampling frequency of a signal at a driving end is 48000S/S;
2) And (3) noise reduction treatment: preprocessing a signal by adopting Fourier transform, and normalizing by using a minimum value-a maximum value after the Fourier transform, so that the numerical value fluctuation range is stabilized at 0,1 on the basis of keeping the original relative relation of data;
3) Feature extraction: determining the number of hidden layers of the edge noise reduction automatic encoder to be three layers, and inputting the preprocessed data into the strong reconstruction edge noise reduction automatic encoder; extracting features after mapping under the action of improved constraint conditions, wherein the input of a hidden layer is the output of the previous hidden layer plus preprocessed data, and calculating the actual output of the hidden layer; calculating partial derivatives of the loss function to the weight and the bias on the basis of input and output, and updating the weight and the bias; comparing the set loss function with the maximum set times, stopping when the expected value is reached, or recalculating the output of the hidden layer;
4) Fault diagnosis and division: dividing fault diagnosis by using an improved correlation vector machine;
5) And (3) diagnosing the state of the mechanical equipment: identifying the state of the bearing data by using the actually measured bearing data;
in the step 3), the specific method is as follows:
3.1 Pre-processing input data, and setting the processed input value and output expected value as x;
3.2 Random initialization weights w and offsets b, using cross entropy to adjust the parameters w, b, defining a loss function:
wherein: d is the input dimension;
a represents the size after batch division;
ik denotes the kth element of the ith input in each batch of data;
summing the input vectors first and then summing each sample of the batch data, the autoencoder constraints are as follows (2):
wherein: n represents the input number;
m represents the number of hidden layer neurons;
z represents a hidden layer feature expression;
l represents a cross entropy function;
β represents a weight coefficient;
ρ represents the degree of activation of each layer of neurons;
3.3 Define the output of the previous hidden layer as x ', then the input of the next hidden layer is (x' + x);
3.4 Computing the actual output z of the output layer processed by the hidden layer;
3.5 On the basis of input and output, partial derivatives of the loss function pair w and b are lost, a scalar form is used in the reconstruction process, and the calculation mode is as follows:
wherein, w qp Representing the weight of the nodes q and p;
w kp representing the weight of the nodes k and p;
b yp represents the bias of node p;
b zk represents the offset of node k;
for the ith sample of each batch of data,represents the input of the pth hidden layer,represents the input of the kth output layer;
f' (x) is the first derivative of sigmoid: f' (x) = f (x) [1-f (x) ]
Calculating a partial derivative of the loss function formula (1) to obtain a formula (6):
wherein, w rs Representing the weight of the nodes r and s;
b yr represents the offset of node r;
3.6 Update the parameter values according to the rule of equation (7), where η is the learning rate, w is the weight, b y Is the offset:
3.7 Compare the set loss function with the maximum set number of times, stop when reaching our expectation, otherwise return to 3.6) continue training.
3.8 The output of the improved edge noise reduction auto-encoder is input to an improved correlation vector machine for classification.
In the step 4), the specific method is as follows:
4.1 Is provided with two types of input samples X 1 ,X 2 ,
X 1 (x 11 ,x 12 ,...x 1i ),X 2 (x 21 ,x 22 ,...x 2i )(i=1,2,...l 1 ,j=1,2,...l 2 ),μ 1 ,μ 2 Is X 1 ,X 2 The mapped mean vector is calculated by (8), (9):
the intra-class scatter matrix is solved by equations (10), (11):
wherein K (,) is a kernel function;
K 1 ,K 2 ,K 11 ,K 22 ,K 12 the formula is as follows:
mapping of input samples to variances in feature spaceThe calculation mode is as the formulas (17) and (18):
under the constraint action of the maximum entropy method, the input sample mapping is solved in the feature spaceVariance of (2)The weight coefficient λ proposed herein is calculated as formula (19):
4.2 According to the improved kernel function mode, the structure of the correlation vector machine is defined as follows:
weight w = (w) 0 ,...,w N ),K(x,x i ) The expression is as in formula (21):
solving for variances in input sample mappings into feature space under Fisher criterion constraintsSolving for variances in input sample mappings into feature space under maximum entropy criterion constraintsThe kernel function weight coefficient λ is calculated according to the proposed method, in the way as equation (22):
4.3 Hyper-parameter α initialization, weight covariance Σ as in equation (23):
Σ=-H=(φ T Bφ+A) -1 (23)
wherein H is Hessian matrix, and y = [ y = 1 ,y 2 ,Λ,y N ] T ,B=diag(β 1 ,β 2 ,Λ,β N ),β n =σ[y(x n )]{1-σ[y(x n )]},A=diag(α 0 ,α 1 ,Λ,α N );
4.4 Estimate the hyperparameter again, in the manner of equation (24):
μ i is X i The mapped mean vector;
4.5 Correlation vector machine classification generally adopts Sigmoid probability function, and its expression form is as follows (25):
wherein: y (x, w) MP )=φ(x)w MP ,w MP Calculated from equation (26):
solving to obtain a probability predicted value p (t =1 calx) of the sample pair category, p (t =0 calx), wherein the category corresponding to the larger value of the values is the recognition result of x.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910388451.XA CN110348468B (en) | 2019-05-10 | 2019-05-10 | Bearing sub-health identification method for optimized strong reconstruction edge noise reduction automatic encoder of correlation vector machine |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910388451.XA CN110348468B (en) | 2019-05-10 | 2019-05-10 | Bearing sub-health identification method for optimized strong reconstruction edge noise reduction automatic encoder of correlation vector machine |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110348468A CN110348468A (en) | 2019-10-18 |
CN110348468B true CN110348468B (en) | 2023-01-20 |
Family
ID=68174645
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910388451.XA Active CN110348468B (en) | 2019-05-10 | 2019-05-10 | Bearing sub-health identification method for optimized strong reconstruction edge noise reduction automatic encoder of correlation vector machine |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110348468B (en) |
Families Citing this family (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110967188B (en) * | 2019-11-08 | 2020-11-24 | 珠海格力电器股份有限公司 | Rolling bearing residual life prediction method and system based on iterative correlation vector machine |
CN111626361B (en) * | 2020-05-28 | 2023-08-11 | 辽宁大学 | Bearing sub-health identification method for improving capsule network optimization hierarchical convolution |
CN111759316A (en) * | 2020-07-10 | 2020-10-13 | 郑州轻工业大学 | Lower limb artificial limb gait recognition method based on muscle sound signal |
CN113328755B (en) * | 2021-05-11 | 2022-09-16 | 内蒙古工业大学 | Compressed data transmission method facing edge calculation |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106017879A (en) * | 2016-05-18 | 2016-10-12 | 河北工业大学 | Universal circuit breaker mechanical fault diagnosis method based on feature fusion of vibration and sound signals |
CN107101828A (en) * | 2017-05-27 | 2017-08-29 | 辽宁大学 | The bearing residual life Forecasting Methodology that a kind of encoder health degree is evaluated |
CN108106844A (en) * | 2017-11-22 | 2018-06-01 | 辽宁大学 | A kind of Method for Bearing Fault Diagnosis of the automatic coding machine of self-tuning parameter adjustment |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107832687A (en) * | 2017-10-27 | 2018-03-23 | 武汉大学 | Fault diagnostic method for transformer winding based on wireless identification sensing |
-
2019
- 2019-05-10 CN CN201910388451.XA patent/CN110348468B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106017879A (en) * | 2016-05-18 | 2016-10-12 | 河北工业大学 | Universal circuit breaker mechanical fault diagnosis method based on feature fusion of vibration and sound signals |
CN107101828A (en) * | 2017-05-27 | 2017-08-29 | 辽宁大学 | The bearing residual life Forecasting Methodology that a kind of encoder health degree is evaluated |
CN108106844A (en) * | 2017-11-22 | 2018-06-01 | 辽宁大学 | A kind of Method for Bearing Fault Diagnosis of the automatic coding machine of self-tuning parameter adjustment |
Non-Patent Citations (3)
Title |
---|
基于融合型深度学习的滚动轴承亚健康识别算法;张利等;《计算机应用》;20180416;全文 * |
深度学习的亚健康识别算法的研究;徐骁;《中国优秀博硕士学位论文全文数据库(硕士)信息科技》;20170215(第02期);全文 * |
融合型深度学习亚健康识别算法的研究;李大伟;《中国优秀博硕士学位论文全文数据库(硕士)信息科技》;20180315(第03期);全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN110348468A (en) | 2019-10-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110348468B (en) | Bearing sub-health identification method for optimized strong reconstruction edge noise reduction automatic encoder of correlation vector machine | |
CN110849627B (en) | Width migration learning network and rolling bearing fault diagnosis method based on same | |
CN109726524B (en) | CNN and LSTM-based rolling bearing residual service life prediction method | |
CN110132554B (en) | Rotary machine fault diagnosis method based on deep Laplace self-coding | |
CN110057584B (en) | Degradation monitoring method for locomotive traction motor bearing | |
CN109827777B (en) | Rolling bearing fault prediction method based on partial least square method extreme learning machine | |
CN113255848B (en) | Water turbine cavitation sound signal identification method based on big data learning | |
CN112257530B (en) | Rolling bearing fault diagnosis method based on blind signal separation and support vector machine | |
CN108106844B (en) | A kind of Method for Bearing Fault Diagnosis of the automatic coding machine of self-tuning parameter adjustment | |
CN111967486A (en) | Complex equipment fault diagnosis method based on multi-sensor fusion | |
CN115017945A (en) | Mechanical fault diagnosis method and system based on enhanced convolutional neural network | |
CN111597651A (en) | Rolling bearing performance degradation evaluation method based on HWPSO-SVDD model | |
CN112729834B (en) | Bearing fault diagnosis method, device and system | |
CN113822139A (en) | Equipment fault diagnosis method based on improved 1DCNN-BilSTM | |
CN111753891A (en) | Rolling bearing fault diagnosis method based on unsupervised feature learning | |
CN115901249B (en) | Rolling bearing performance degradation evaluation method combining feature optimization and multi-strategy optimization SVDD | |
Li et al. | Gear pitting fault diagnosis using raw acoustic emission signal based on deep learning | |
CN114091525A (en) | Rolling bearing degradation trend prediction method | |
CN112182912B (en) | Manufacturing equipment spindle bearing health assessment method based on probability description and spectrum analysis | |
CN115290326A (en) | Rolling bearing fault intelligent diagnosis method | |
CN114741945B (en) | On-line fault diagnosis method for aero-engine | |
CN115587290A (en) | Aero-engine fault diagnosis method based on variational self-coding generation countermeasure network | |
CN115238736A (en) | Method for identifying early fault of rotary machine | |
CN113126489B (en) | Intelligent BIT design method of heavy-duty gas turbine control system based on CNN-GRU-BINN | |
CN117458955A (en) | Operation control method and system of motor |
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 |