CN109858352A - A kind of method for diagnosing faults based on compressed sensing and the multiple dimensioned network of improvement - Google Patents
A kind of method for diagnosing faults based on compressed sensing and the multiple dimensioned network of improvement Download PDFInfo
- Publication number
- CN109858352A CN109858352A CN201811601591.2A CN201811601591A CN109858352A CN 109858352 A CN109858352 A CN 109858352A CN 201811601591 A CN201811601591 A CN 201811601591A CN 109858352 A CN109858352 A CN 109858352A
- Authority
- CN
- China
- Prior art keywords
- multiple dimensioned
- sample
- feature
- feature mapping
- mapping
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a kind of based on compressed sensing and improves the method for diagnosing faults of multiple dimensioned network can retain interested information using compressed sensing compression samples from original sample signal, improve data analysis efficiency.Using the Gauss measurement Matrix Multiplication of compressed sensing with signal matrix, since Gauss measurement matrix is random matrix, the matrix element generated every time is different, the training sample of generation is also different, generate enough training samples, small sample problem present in depth model training process is further solved, the accurate rate of small sample fault diagnosis is improved.Identical mapping is introduced between the Feature Mapping of Feature Mapping quantity being of the same size, pass through the feature extraction to different layers, the merging features for connecting different layers are that an entirety realizes Fusion Features, obtain depth integration feature vector corresponding with input picture, to solve multiple dimensioned network optimization process difficult problem, fault type can be effectively identified under different operating conditions.
Description
Technical field
The invention belongs to rotary machinery fault diagnosis technical fields, are based on compressed sensing more particularly, to one kind and change
Into the method for diagnosing faults of multiple dimensioned network.
Background technique
Various damages and failure can occur under severe and extreme operating condition for rotating machinery, and operational failure will affect
The performance of whole system simultaneously may cause system-down, cause heavy losses.Generation, guarantee equipment safety in order to avoid failure
Operation reduces economic loss, needs to carry out intelligent diagnostics and prediction to rotatory mechanical system.However traditional method for diagnosing faults
There is limitation, diagnosis performance depends greatly on expertise and priori knowledge, and the model for prediction
Often with the network structure of shallow-layer.
Convolutional neural networks are widely used in fault diagnosis in recent years, have than traditional artificial neural network deeper time
Network structure, more correlated characteristics can be integrated in prediction model in order to diagnostic classification, or large-scale nerve
Network provides partially connected, avoids overfitting.In order to make full use of the advantage of convolutional neural networks, Hu proposes a kind of multiple dimensioned
Network (multi-scale network, MSN) comprising three branch models of a different layers, it can efficiently extract phase
It closes feature and carries out the abstract of different levels, and merge related and sensitive feature by completely connecting.Although multiple dimensioned
Network has advantage in terms of feature learning, but it still has some limitations in terms of fault diagnosis: (1) multiple dimensioned Web vector graphic
The method of convolution adjusts network size size, is easy to extract validity feature, but its model optimization process is difficult；(2) traditional
Shannon's sampling theorem can easily obtain condition monitoring and fault diagnosis of a large amount of data information for rotating machinery, but
Mass data increases the difficulty of data analysis, it is difficult to the quantity and data of signal acquisition are kept in depth characteristic learning process
Balance between analysis efficiency；(3) in some cases, the data training set for deep learning process is limited, training
The size of collection affects the nicety of grading of fault diagnosis model, and how improving nicety of grading, there are also to be solved.
Summary of the invention
In view of the drawbacks of the prior art, it is an object of the invention to solve multiple dimensioned network diagnosis in the prior art to optimize
Journey is difficult, is difficult to the technical issues of being suitable for small sample.
To achieve the above object, in a first aspect, the embodiment of the invention provides one kind based on compressed sensing and to improve more rulers
The method for diagnosing faults of network is spent, method includes the following steps:
Step S1. is for training sample set and sample to be tested, the vibration signal based on compressed sensing compression rotating device；
Step S2. is reassembled as two dimensional image for training sample set and sample to be tested, by signal after compression, to the two dimension
The pixel of image is normalized；
Step S3. is multiple dimensioned as improving using the normalized two dimensional image of pixel for training sample set and sample to be tested
The input of network improves multiple dimensioned network and exports depth integration feature vector corresponding with input picture, and the improvement is multiple dimensioned
Network introduces identical mapping between the Feature Mapping of multiple dimensioned network；
Step S4. using the depth integration feature vector of the training sample set as the input of softmax sorter model,
Using the fault type label of the training sample set as the output of softmax sorter model, training improves multiple dimensioned network
The fault diagnosis model constituted with softmax classifier；
The depth integration feature vector of the sample to be tested is inputted trained fault diagnosis model by step S5., is obtained
The fault diagnosis result of the sample to be tested.
Specifically, for small sample training set, the method for diagnosing faults further includes after step S1, before step S2
Following steps:
Large sample training is repeatedly generated with the signal matrix of training sample using the Gauss measurement Matrix Multiplication of compressed sensing
Collection.
Specifically, in step S1, front signal X, signal Y=Φ X after compression are compressed, wherein Φ ∈ RM×NIndicate compressed sensing
Gauss measurement matrix.
Specifically, described that signal after compression is reassembled as two dimensional image in step S2 specifically: from after compression in signal Y
Randomly select the data sample of n n dimension, the every a line for the image Z that obtained sample is successively constructed.
Specifically, described to improve multiple dimensioned network be a convolutional neural networks, and one shares three layers: first layer to image into
Five convolution operations of row, the size reduction of the Feature Mapping after convolution are the 1/4 of a upper Feature Mapping size, Feature Mapping number
Amount increases to 2 times of a Feature Mapping quantity；The second layer carries out secondary convolution operation to image, the Feature Mapping after convolution
Size reduction be a upper Feature Mapping size 1/16, Feature Mapping quantity increases to the 4 of a Feature Mapping quantity
Times；Third layer carries out a convolution operation to image, and the size reduction of the Feature Mapping after convolution is a upper Feature Mapping ruler
Very little 1/16, Feature Mapping quantity increase to 4 times of a Feature Mapping quantity, are being of the same size and Feature Mapping
Identical mapping is introduced between the Feature Mapping of quantity.
Specifically, by the feature extraction to different layers, the merging features for connecting different layers are that an entirety obtains one
Articulamentum realizes Fusion Features by the full connection of output layer, obtains depth integration feature vector corresponding with input picture.
Specifically, the transmission function of the identical mapping are as follows:
Wherein,For l layers of j-th of Feature Mapping, f () is the activation primitive for improving multiple dimensioned network,It is to change
J-th of convolution kernel into l layers in multiple dimensioned network,It is offset parameter, operator * is convolution operation, MjAnd NjIt is respectively
Improve l-1 layers and s layers in multiple dimensioned network of Feature Mapping quantity.
Specifically, the loss function for improving multiple dimensioned network is cross entropy cost function.
Specifically, the fault diagnosis result includes: fault type and fault severity level.
Second aspect, the embodiment of the invention provides a kind of computer readable storage medium, the computer-readable storage mediums
Computer program is stored in matter, which realizes fault diagnosis described in above-mentioned first aspect when being executed by processor
Method.
In general, through the invention it is contemplated above technical scheme is compared with the prior art, have below beneficial to effect
Fruit:
1, the present invention uses compressed sensing technology compression samples, can protect from original sample signal in process of data preprocessing
Interested information is stayed, to improve data analysis efficiency.
2, the present invention utilizes the Gauss measurement Matrix Multiplication of compressed sensing with signal matrix, since Gauss measurement matrix is one
Random matrix, the matrix element generated every time are different, and therefore, the training sample of generation is also different.Expectation is manually set
Multiple proportion between obtained training samples number and original sample quantity, so that enough training samples are generated, into one
Step solves small sample problem present in depth model training process, improves the accurate rate of small sample fault diagnosis.
3, the present invention is being of the same size the introducing identical mapping algorithm between the Feature Mapping of Feature Mapping quantity,
By the feature extraction to different layers, the merging features for connecting different layers are that an entirety obtains an articulamentum, pass through output
Fusion Features are realized in the full connection of layer, obtain depth integration feature vector corresponding with input picture as the multiple dimensioned network of improvement
Output can effectively identify fault type under different operating conditions to solve multiple dimensioned network optimization process difficult problem.
Detailed description of the invention
Fig. 1 is a kind of method for diagnosing faults based on compressed sensing and the multiple dimensioned network of improvement provided in an embodiment of the present invention
Flow chart；
Fig. 2 is signal compression process schematic provided in an embodiment of the present invention；
Fig. 3 is two dimensional image organigram provided in an embodiment of the present invention；
Fig. 4 is the multiple dimensioned schematic network structure of improvement provided in an embodiment of the present invention；
Fig. 5 is the multi-level confusion matrix figure of fault diagnosis provided in an embodiment of the present invention；
Fig. 6 is the difference based on compressed sensing and the method for diagnosing faults for improving multiple dimensioned network that inventive embodiments provide
The relation schematic diagram of hyper parameter and nicety of grading；
Fig. 7 is the difference based on compressed sensing and the method for diagnosing faults for improving multiple dimensioned network that inventive embodiments provide
Hyper parameter and the relation schematic diagram for calculating the time；
Fig. 8 is the diagnostic result schematic diagram of different training sample sizes provided in an embodiment of the present invention；
Fig. 9 is the nicety of grading schematic diagram of different faults diagnostic method provided in an embodiment of the present invention.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.
Based on compressed sensing and improve the method for diagnosing faults of multiple dimensioned network as shown in Figure 1, a kind of, this method include with
Lower step:
Step S1. is for training sample set and sample to be tested, the vibration signal based on compressed sensing compression rotating device；
Step S2. is reassembled as two dimensional image for training sample set and sample to be tested, by signal after compression, to the two dimension
The pixel of image is normalized；
Step S3. is multiple dimensioned as improving using the normalized two dimensional image of pixel for training sample set and sample to be tested
The input of network improves multiple dimensioned network and exports depth integration feature vector corresponding with input picture, and the improvement is multiple dimensioned
Network introduces identical mapping between the Feature Mapping of multiple dimensioned network；
Step S4. using the depth integration feature vector of the training sample set as the input of softmax sorter model,
Using the fault type label of the training sample set as the output of softmax sorter model, training improves multiple dimensioned network
The fault diagnosis model constituted with softmax classifier；
The depth integration feature vector of the sample to be tested is inputted trained fault diagnosis model by step S5., is obtained
The fault diagnosis result of the sample to be tested.
For small sample training set, the method for diagnosing faults after step S1, before step S2 further include following step
It is rapid:
Large sample training is repeatedly generated with the signal matrix of training sample using the Gauss measurement Matrix Multiplication of compressed sensing
Collection.
Step S1. is for training sample set and sample to be tested, the vibration signal based on compressed sensing compression rotating device.
Signal compression ratioM indicates signal dimension after compression, and N indicates compression front signal dimension (original sample signal
Dimension), M < N.Suitable compression ratio is selected, interested letter can be retained from original sample signal in process of data preprocessing
Breath.
Original sample signal comes from acquisition device, the acquisition rotation of the acquisition devices such as acceleration transducer, pressure sensor, sound emission
The vibration signal of rotary device obtains vibration signalWherein, Ψ={ ψ1,ψ2,...,ψN}T∈RN×NIt indicates
Sparse basis array, Θ={ θ1,θ2,...,θN}T∈RNIndicate sparse coefficient, N indicates original sample signal dimension.Rotating device includes
The mechanical devices such as blower, motor, rolling bearing.
Compress original sample signal X, signal Y=Φ X after compression, wherein Φ ∈ RM×NIndicate the Gauss measurement square of compressed sensing
Battle array.In compressed sensing, signal X is projected on calculation matrix (random matrix) Φ, and obtains measured value Y, i.e., with one
It is a that linear measurement Y is obtained to signal X progress linear projection with transformation matrix incoherent M × N calculation matrix Φ.Such as Fig. 2
Shown, vibration signal X is time-domain signal, and the signal Y after the compression that compression processing obtains, dimension is M × 1, sample dimension
It reduces.
Using the Gauss measurement Matrix Multiplication of compressed sensing with signal matrix X, since Gauss measurement matrix is a random square
Battle array, the matrix element generated every time are different, and therefore, the training sample Y of generation is also different.It is manually set desired
Multiple proportion between training samples number and original sample quantity further solves to generate enough training samples
Small sample problem present in depth model training process.Fault diagnosis model parameter is trained, Gauss measurement matrix is
It is generated by gauss of distribution function.
Step S2. is reassembled as two dimensional image for training sample set and sample to be tested, by signal after compression, to the two dimension
The pixel of image is normalized.
(2.1) signal Y after compression is converted into two dimensional image Z
Assuming that the dimension of signal Y is n after compression2, n here2=M, first from randomly selecting n n dimension after compression in signal Y
Data sample, the every a line for the image Z that then obtained sample successively constructs, at this point, the dimension of image Z be n × n.
(2.2) pixel of image Z is normalized
Normalizing calculation method isWherein, Y (i, j) is the signal Y after compression
J-th of value of i-th of sample data section of middle selection, min (Y) are the minimum value of signal Y after compression, and max (Y) is to believe after compressing
The maximum value of number Y, p (i, j) are the pixel value after normalizing in image Z, and value range is [0,1].
Step S3. is multiple dimensioned as improving using the normalized two dimensional image of pixel for training sample set and sample to be tested
The input of network improves multiple dimensioned network and exports depth integration feature vector corresponding with input picture, and the improvement is multiple dimensioned
Network introduces identical mapping between the Feature Mapping of multiple dimensioned network.
Using multiple dimensioned network (identical mapping is integrated into multiple dimensioned network) is improved, image Z after normalization is carried out
Depth characteristic is extracted and depth characteristic fusion.Using the normalized two dimensional image of pixel as the input for improving multiple dimensioned network, with
The corresponding depth integration feature vector of input picture is set as C as the output for improving multiple dimensioned network, the quantity of Feature Mapping,
The neuronal quantity of output layer is Noutput。
As shown in figure 4, improving multiple dimensioned network is a convolutional neural networks, one shares three layers: first layer to image into
Five convolution operations of row, the size reduction of the Feature Mapping after convolution are the 1/4 of a upper Feature Mapping size, Feature Mapping number
Amount increases to 2 times of a Feature Mapping quantity；The second layer carries out secondary convolution operation to image, the Feature Mapping after convolution
Size reduction be a upper Feature Mapping size 1/16, Feature Mapping quantity increases to the 4 of a Feature Mapping quantity
Times；Third layer carries out a convolution operation to image, and the size reduction of the Feature Mapping after convolution is a upper Feature Mapping ruler
Very little 1/16, Feature Mapping quantity increase to 4 times of a Feature Mapping quantity.Such convolution operation is known as feature and mentions
It takes.Identical mapping algorithm is introduced being of the same size between the Feature Mapping of Feature Mapping quantity, by different layers
Feature extraction, the merging features for connecting different layers are that an entirety obtains an articulamentum, pass through the full connection of output layer reality
Existing Fusion Features obtain depth integration feature vector corresponding with input picture as the output for improving multiple dimensioned network.
It improves multiple dimensioned network to learn to improve diagnosis performance by depth characteristic, the loss function for improving multiple dimensioned network is
Cross entropy cost function.
The transmission function of identical mapping are as follows:
Wherein,For l layers of j-th of Feature Mapping, f () is the activation primitive for improving multiple dimensioned network,It is to change
J-th of convolution kernel into l layers in multiple dimensioned network,It is offset parameter, operator * is convolution operation, MjAnd NjIt is respectively
Improve l-1 layers and s layers in multiple dimensioned network of Feature Mapping quantity.
Step S4. using the depth integration feature vector of the training sample set as the input of softmax sorter model,
Using the fault type label of the training sample set as the output of softmax sorter model, training improves multiple dimensioned network
The fault diagnosis model constituted with softmax classifier.
The parameter for improving multiple dimensioned network may be configured as: Epoch:10, Batch size:64, Learning rate:
0.0002,Decay(Learning rate):0.9,Optimizer:RMSProp,Leaky(ReLU):0.2。
The depth integration feature vector of the sample to be tested is inputted trained fault diagnosis model by step S5., is obtained
The fault diagnosis result of the sample to be tested.
Sample to be tested also needs to pre-process, by the handling by step S1-S3 of raw sensory signal under different operating conditions
To depth integration feature, substitution obtains obtaining corresponding fault diagnosis result in trained fault diagnosis model through step S4,
Fault diagnosis result includes: fault type and fault severity level.
Embodiment 1: validation verification
In order to verify the validity of method for diagnosing faults proposed by the present invention, the present embodiment uses U.S.'s Case Western Reserve University
Bearing data center provide open experimental data set verified.Experimental bench includes: driving motor, a drive end axle
It holds, exchanger encoder and dynamometer, setting driving motor revolving speed is 1750 revs/min, loads as 2Hp, acquires driving motor end 6
O'clock position vibration signal, sample rate 12kHz.The fault type of rolling bearing is inner ring defect, outer ring defect and ball
Defect, wherein fault severity level is simulated by electrical discharge machining, and fault diameter is respectively 7,14,21 mils (mil).10
The details of kind bearing state type is as shown in table 1.2400 data samples are acquired under every kind of bearing state type, wherein
Each data sample includes 4096 data points.
Table 1
Its specific steps is as shown in Figure 1, be described as follows:
(1) data acquisition and compression
The original sensor data collected is pre-processed using compressed sensing.It is adopted from every kind of bearing state type
In 2400 data samples of collection, 2000 data samples are randomly selected as training sample, 400 data samples in addition are made
For test sample.Since each data sample includes 4096 data points, when selecting compression ratio CR is 0.25, signal after compression
The data length of Y is 1024.It is as shown in Figure 2 using the signal compression process of compressed sensing.
(2) building of two dimensional image
Signal Y after compression is converted into two dimensional image Z, it is assumed that the dimension of signal Y is n after compression2, n=32, first herein
First from 32 32 dimension sample data sections are randomly selected after compression in signal Y, then obtained sample data section is successively constructed
Every a line of image Z, the dimension of image Z are 32 × 32, and two dimensional image organigram is as shown in figure 3, finally to the picture of image Z
Element is normalized.
(3) depth characteristic is extracted
Normalize two dimensional image using multiple dimensioned network processes pixel is improved, to two dimensional compaction image carry out feature learning and
Depth characteristic is extracted.As shown in figure 4, improving multiple dimensioned network one shares three layers, first layer carries out five convolution operations to image,
The quantity of Feature Mapping C is set as 16, and the size W of the Feature Mapping after convolution is reduced into the 1/4 of a Feature Mapping size,
Respectively 16 × 16,8 × 8,4 × 4,2 × 2,1 × 1, quantity increases to 2 times of a Feature Mapping quantity, and respectively 16,
32,64,128,256；The second layer carries out secondary convolution operation to image, and the size reduction of the Feature Mapping after convolution is upper one
The 1/16 of Feature Mapping size, respectively 8 × 8,2 × 2, quantity increases to 4 times of a Feature Mapping quantity, distinguish 32,
128；Third layer carries out a convolution operation to image, and the size reduction of the Feature Mapping after convolution is a upper Feature Mapping ruler
Very little 1/16 is 8 × 8, and quantity increases to 32；Such convolution operation is known as feature extraction.It is being of the same size and special
Identical mapping algorithm is introduced between the Feature Mapping of sign mapping amount, and different layers are connected by the feature extraction to different layers
Merging features are that an entirety obtains an articulamentum, realize Fusion Features by the full connection of output layer, obtain and input figure
Output as corresponding depth integration feature vector as the multiple dimensioned network of improvement, the neuronal quantity of output layer are set as 512.
As the network number of plies changes, the quantity of Feature Mapping is in exponential increase, and the size of Feature Mapping will become smaller, with same color
Feature Mapping is of the same size and quantity.Such as: the Feature Mapping 12,21 and 31 in Fig. 4 is of the same size, feature
Mapping 14 and 22 is identical, which is to realize the premise of identical mapping.The parameter setting for improving multiple dimensioned network is as shown in table 2.
Table 2
(4) diagnostic model training
Using the feature vector through depth integration in step (3) as input, softmax sorter model is substituted into, through model
Initial setting up and training are parameterized, optimal diagnostic model is obtained.
(5) state recognition
Raw sensory signal is obtained into depth integration feature by the processing of step (1)~(3), substitutes into and is instructed through step (4)
In experienced optimal diagnosis model, corresponding fault diagnosis result is obtained, the multi-level confusion matrix figure of fault diagnosis is as shown in Figure 5.
Fig. 5 shows the invention proposes based on compressed sensing and improving the method for diagnosing faults of multiple dimensioned network to the 1st, 2,3,4,7,
The nicety of grading of 8,9,10 kinds of bearing state types is 100%, and the nicety of grading to the 5th kind of bearing state type is 99%, is had
1% probability meeting mistaken diagnosis is the 6th kind of bearing state type, and the nicety of grading to the 6th kind of bearing state type is 96%, there is 4%
Probability can mistaken diagnosis be the 5th kind of bearing state type, have preferable diagnosis performance.
Embodiment 2: hyper parameter setting
The present invention compresses original sensor data using compressed sensing technology, improves data to a certain extent
Efficiency has studied using compressed sensing, using the performance of the data-driven method for diagnosing faults of different hyper parameters,
Including calculating time and nicety of grading；Time and calculating when without using compressed sensing are calculated when compared using compressed sensing
Time.Its specific implementation step is as shown in Figure 1, the setting of deep learning procedure parameter as shown in table 3, has carried out 10 repetition examinations altogether
It tests, average result is calculated for data.
Table 3
Experimental result is as shown in Figure 6, Figure 7, the experimental results showed that, for specific Feature Mapping quantity C, with output layer
Neuronal quantity increase, rolling bearing nicety of grading improve；When the quantity of Feature Mapping increases, 32,64 and are such as selected
When 128, rolling bearing nicety of grading nearly reaches 100%.On on the other hand, when use less Feature Mapping such as 8 and 16
When, depth characteristic learning ability can be significantly improved using compressed sensing technology；32,64 and when using more Feature Mappings for example
When 128, effect of the compressed sensing in terms of improving nicety of grading is not protruded.When using compressed sensing, the improvement that is proposed
Multiple dimensioned network can be by using suitable Feature Mapping quantity C and more neuronal quantities, raising to change in output layer
Into the depth characteristic learning ability of multiple dimensioned network.
The experimental results showed that the neuronal quantity of output layer does not have an impact substantially to the time is calculated, with Feature Mapping number
The increase of C is measured, the calculating time also increases as.
Table 4 and table 5 are shown in depth model training process using compressed sensing technology and without using compressed sensing skill
The time loss of art improves depth model the experimental results showed that facilitating to reduce depth model learning time using compressed sensing
Learning efficiency.
The calculating time (unit: second) of the different hyper parameter drag training of table 4
The calculating time (unit: second) of the different hyper parameter drag training (not using compressed sensing) of table 5
Embodiment 3: small sample problem
The size of sensing data collection is depended on generally, based on the performance of the data-driven method for diagnosing faults of depth model,
When using more training datasets, the accuracy of fault diagnosis will be improved, however be difficult to obtain foot in practical applications
Enough monitoring sensing datas.The present invention uses the calculation matrix of compressed sensing to generate enough training samples for subsequent first
Deep learning analysis.Its specific implementation step is carried out altogether as shown in Figure 1, small sample training process parameter setting is as shown in table 6
10 repetitions are tested, and average result is calculated for data.
Different bearing state types (being shown in Table 1) randomly select 200 training samples, then the training of 10 kinds of bearing state types
Total sample number is 2000.This example generates enough samples for training pattern, the instruction of generation using Gauss measurement matrix
The multiple for practicing sample is set as 1,2,3 ..., and 19,20 times, so the size of the training sample generated is 2000,4000,
6000 ..., 38000,40000.Last diagnostic result is as shown in figure 8, being plotted in the nicety of grading of test sample with ellipse
Red line on, utilize black error bars to draw corresponding nicety of grading error.It is according to figure 8 as a result, different conditions type
Rolling bearing nicety of grading all with training sample quantity increase and improves, standard deviation with training sample quantity increase
And reduce, diagnosis performance improves.Therefore enough training samples are generated using the Gauss measurement matrix of compressed sensing, can solve
Small sample problem present in depth model training process.
Table 6
Embodiment 4: conventional fault diagnosis method comparison
In order to verify the validity proposed by the present invention for improving multiple dimensioned network, multiple dimensioned network will be improved and be based on more rulers
It spends network, LeCun Net-5, depth confidence network, stack the progress of the classical faults diagnostic methods such as self-encoding encoder, support vector machines
Comparative experiments.Its specific implementation step is carried out altogether as shown in Figure 1, conventional fault diagnosis method condition setting is as shown in table 7
10 repetitions are tested, and average result is calculated for data.
Using 20000 sample training fault diagnosis models of 10 kinds of different bearing state types, remaining 4000 samples
This is for testing.As shown in figure 9, experiment shows the nicety of grading based on the method for diagnosing faults for improving multiple dimensioned network better than it
His method, while its standard deviation is relatively small.
Table 7
The characteristics of present invention is according to compressed sensing technology and improvement multiple dimensioned network, realizes and extracts from non-stationary signal
Depth integration feature efficiently solves the small sample problem in depth model training process.Method proposed by the invention is
Based on method, using rolling bearing as case verification, in actual application, extend to general rotating machinery, processing system
It makes, plant maintenance etc., there is good engineering practicability.
More than, the only preferable specific embodiment of the application, but the protection scope of the application is not limited thereto, and it is any
Within the technical scope of the present application, any changes or substitutions that can be easily thought of by those familiar with the art, all answers
Cover within the scope of protection of this application.Therefore, the protection scope of the application should be subject to the protection scope in claims.
Claims (10)
- Based on compressed sensing and improve the method for diagnosing faults of multiple dimensioned network 1. a kind of, which is characterized in that this method include with Lower step:Step S1. is for training sample set and sample to be tested, the vibration signal based on compressed sensing compression rotating device；Step S2. is reassembled as two dimensional image for training sample set and sample to be tested, by signal after compression, to the two dimensional image Pixel be normalized；Step S3. is for training sample set and sample to be tested, using the normalized two dimensional image of pixel as the multiple dimensioned network of improvement Input, improve multiple dimensioned network and export depth integration feature vector corresponding with input picture, the multiple dimensioned network of improvement Identical mapping is introduced between the Feature Mapping of multiple dimensioned network；Step S4. is using the depth integration feature vector of the training sample set as the input of softmax sorter model, by institute State output of the fault type label of training sample set as softmax sorter model, training by improve multiple dimensioned network and The fault diagnosis model that softmax classifier is constituted；The depth integration feature vector of the sample to be tested is inputted trained fault diagnosis model by step S5., is obtained described The fault diagnosis result of sample to be tested.
- 2. method for diagnosing faults as described in claim 1, which is characterized in that be directed to small sample training set, the fault diagnosis Method after step S1, it is further comprising the steps of before step S2:Large sample training set is repeatedly generated with the signal matrix of training sample using the Gauss measurement Matrix Multiplication of compressed sensing.
- 3. method for diagnosing faults as described in claim 1, which is characterized in that in step S1, compress front signal X, believe after compression Number Y=Φ X, wherein Φ ∈ RM×NIndicate the Gauss measurement matrix of compressed sensing.
- 4. method for diagnosing faults as described in claim 1, which is characterized in that described to recombinate signal after compression in step S2 For two dimensional image specifically: from the data sample for randomly selecting n n dimension after compression in signal Y, obtained sample is successively constructed Image Z every a line.
- 5. method for diagnosing faults as described in claim 1, which is characterized in that the multiple dimensioned network of improvement is a convolution mind Through network, one shares three layers: first layer carries out five convolution operations to image, and the size reduction of the Feature Mapping after convolution is upper The 1/4 of one Feature Mapping size, Feature Mapping quantity increase to 2 times of a Feature Mapping quantity；The second layer is to image Secondary convolution operation is carried out, the size reduction of the Feature Mapping after convolution is the 1/16 of a upper Feature Mapping size, and feature is reflected Penetrate 4 times that quantity increases to a Feature Mapping quantity；Third layer carries out a convolution operation to image, the feature after convolution The size reduction of mapping is the 1/16 of a upper Feature Mapping size, and Feature Mapping quantity increases to a Feature Mapping quantity 4 times, introduce identical mapping between the Feature Mapping of Feature Mapping quantity being of the same size.
- 6. method for diagnosing faults as claimed in claim 5, which is characterized in that by the feature extraction to different layers, connection is not The merging features of same layer are that an entirety obtains an articulamentum, by the complete connection of output layer realization Fusion Features, obtain and The corresponding depth integration feature vector of input picture.
- 7. method for diagnosing faults as described in claim 1, which is characterized in that the transmission function of the identical mapping are as follows:Wherein,For l layers of j-th of Feature Mapping, f () is the activation primitive for improving multiple dimensioned network,It is that improvement is more J-th of convolution kernel in scale network in l layers,It is offset parameter, operator * is convolution operation, MjAnd NjIt is to improve respectively L-1 layers and s layers of Feature Mapping quantity in multiple dimensioned network.
- 8. method for diagnosing faults as described in claim 1, which is characterized in that the loss function for improving multiple dimensioned network is Cross entropy cost function.
- 9. method for diagnosing faults as described in claim 1, which is characterized in that the fault diagnosis result includes: fault type And fault severity level.
- 10. a kind of computer readable storage medium, which is characterized in that be stored with computer on the computer readable storage medium Program, the computer program realize method for diagnosing faults as described in any one of claim 1 to 9 when being executed by processor.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811601591.2A CN109858352B (en) | 2018-12-26 | 2018-12-26 | Fault diagnosis method based on compressed sensing and improved multi-scale network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811601591.2A CN109858352B (en) | 2018-12-26 | 2018-12-26 | Fault diagnosis method based on compressed sensing and improved multi-scale network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109858352A true CN109858352A (en) | 2019-06-07 |
| CN109858352B CN109858352B (en) | 2020-09-18 |
Family
ID=66892491
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811601591.2A Active CN109858352B (en) | 2018-12-26 | 2018-12-26 | Fault diagnosis method based on compressed sensing and improved multi-scale network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109858352B (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110555478A (en) * | 2019-09-05 | 2019-12-10 | 东北大学 | Fan multi-fault diagnosis method based on depth measurement network of difficult sample mining |
| CN110595780A (en) * | 2019-09-20 | 2019-12-20 | 西安科技大学 | Bearing fault identification method based on vibration gray level image and convolution neural network |
| CN110928237A (en) * | 2019-12-20 | 2020-03-27 | 华中科技大学 | Vibration signal-based numerical control machining center flutter online identification method |
| CN110940889A (en) * | 2019-12-06 | 2020-03-31 | 西安锐驰电器有限公司 | Fault detection method for high-voltage power equipment |
| CN111240279A (en) * | 2019-12-26 | 2020-06-05 | 浙江大学 | Confrontation enhancement fault classification method for industrial unbalanced data |
| CN111397902A (en) * | 2020-03-22 | 2020-07-10 | 华南理工大学 | Rolling bearing fault diagnosis method based on feature alignment convolutional neural network |
| CN112614089A (en) * | 2020-12-04 | 2021-04-06 | 淮阴工学院 | FPC defect detection method based on Bayesian compressed sensing and deep learning |
| CN112634391A (en) * | 2020-12-29 | 2021-04-09 | 华中科技大学 | Gray level image depth reconstruction and fault diagnosis system based on compressed sensing |
| CN113281029A (en) * | 2021-06-09 | 2021-08-20 | 重庆大学 | Rotating machinery fault diagnosis method and system based on multi-scale network structure |
| CN113609939A (en) * | 2021-07-26 | 2021-11-05 | 吉林大学 | Moving target efficient identification method based on compressed observation |
| CN114964782A (en) * | 2022-06-02 | 2022-08-30 | 昆明理工大学 | Rolling bearing fault diagnosis acoustic emission signal processing method and system based on compressed sensing |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7639025B2 (en) * | 2002-04-10 | 2009-12-29 | Dcg Systems, Inc. | Collection optics integrating an objective and a SIL |
| CN102435934A (en) * | 2011-09-20 | 2012-05-02 | 北京工业大学 | Random sampling analog circuit compressed sensing measurement and signal reconstruction method |
| EP2679936A1 (en) * | 2012-06-25 | 2014-01-01 | Whirlpool Corporation | Fault detection and diagnosis for refrigerator from compressor sensor |
| CN103645029A (en) * | 2013-12-09 | 2014-03-19 | 江苏大学 | Powder flow pattern detection apparatus in spiral conveyer pipe and compression perception flow pattern identification method |
| CN105023282A (en) * | 2014-04-30 | 2015-11-04 | 华中科技大学 | Sparse projection ultrasonic CT image reconstruction method based on CS |
| CN105258947A (en) * | 2015-11-06 | 2016-01-20 | 北京航空航天大学 | Compressed sensing-based antifriction bearing fault diagnosis method under working condition disturbance condition |
| CN105372087A (en) * | 2015-11-30 | 2016-03-02 | 南通大学 | Fault diagnosis method based on multi-sensor signal analysis |
| CN105761290A (en) * | 2016-03-01 | 2016-07-13 | 北京理工大学 | Adaptive multi-scale partitioning compression sensing sampling method |
| CN106997463A (en) * | 2017-03-29 | 2017-08-01 | 浙江师范大学 | A kind of guideboard recognition methods based on compressed sensing domain and convolutional neural networks |
| US20180121723A1 (en) * | 2015-12-31 | 2018-05-03 | International Business Machines Corporation | Visual object and event detection and prediction system using saccades |
| CN108896296A (en) * | 2018-04-18 | 2018-11-27 | 北京信息科技大学 | A kind of wind turbine gearbox method for diagnosing faults based on convolutional neural networks |
-
2018
- 2018-12-26 CN CN201811601591.2A patent/CN109858352B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7639025B2 (en) * | 2002-04-10 | 2009-12-29 | Dcg Systems, Inc. | Collection optics integrating an objective and a SIL |
| CN102435934A (en) * | 2011-09-20 | 2012-05-02 | 北京工业大学 | Random sampling analog circuit compressed sensing measurement and signal reconstruction method |
| EP2679936A1 (en) * | 2012-06-25 | 2014-01-01 | Whirlpool Corporation | Fault detection and diagnosis for refrigerator from compressor sensor |
| CN103645029A (en) * | 2013-12-09 | 2014-03-19 | 江苏大学 | Powder flow pattern detection apparatus in spiral conveyer pipe and compression perception flow pattern identification method |
| CN105023282A (en) * | 2014-04-30 | 2015-11-04 | 华中科技大学 | Sparse projection ultrasonic CT image reconstruction method based on CS |
| CN105258947A (en) * | 2015-11-06 | 2016-01-20 | 北京航空航天大学 | Compressed sensing-based antifriction bearing fault diagnosis method under working condition disturbance condition |
| CN105372087A (en) * | 2015-11-30 | 2016-03-02 | 南通大学 | Fault diagnosis method based on multi-sensor signal analysis |
| US20180121723A1 (en) * | 2015-12-31 | 2018-05-03 | International Business Machines Corporation | Visual object and event detection and prediction system using saccades |
| CN105761290A (en) * | 2016-03-01 | 2016-07-13 | 北京理工大学 | Adaptive multi-scale partitioning compression sensing sampling method |
| CN106997463A (en) * | 2017-03-29 | 2017-08-01 | 浙江师范大学 | A kind of guideboard recognition methods based on compressed sensing domain and convolutional neural networks |
| CN108896296A (en) * | 2018-04-18 | 2018-11-27 | 北京信息科技大学 | A kind of wind turbine gearbox method for diagnosing faults based on convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| MIN XIA 等: "Fault Diagnosis for Rotating Machinery Using Multiple Sensors and Convolutional Neural Networks", 《IEEE/ASME TRANSACTIONS ON MECHATRONICS》 * |
| 王攀 等: "《优化与控制中的软计算方法研究》", 31 January 2017, 湖北科学技术出版社 * |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110555478A (en) * | 2019-09-05 | 2019-12-10 | 东北大学 | Fan multi-fault diagnosis method based on depth measurement network of difficult sample mining |
| CN110555478B (en) * | 2019-09-05 | 2023-02-03 | 东北大学 | Fan multi-fault diagnosis method based on depth measurement network of difficult sample mining |
| CN110595780B (en) * | 2019-09-20 | 2021-12-14 | 西安科技大学 | Bearing fault identification method based on vibration gray level image and convolution neural network |
| CN110595780A (en) * | 2019-09-20 | 2019-12-20 | 西安科技大学 | Bearing fault identification method based on vibration gray level image and convolution neural network |
| CN110940889A (en) * | 2019-12-06 | 2020-03-31 | 西安锐驰电器有限公司 | Fault detection method for high-voltage power equipment |
| CN110928237A (en) * | 2019-12-20 | 2020-03-27 | 华中科技大学 | Vibration signal-based numerical control machining center flutter online identification method |
| CN111240279A (en) * | 2019-12-26 | 2020-06-05 | 浙江大学 | Confrontation enhancement fault classification method for industrial unbalanced data |
| CN111397902A (en) * | 2020-03-22 | 2020-07-10 | 华南理工大学 | Rolling bearing fault diagnosis method based on feature alignment convolutional neural network |
| CN112614089A (en) * | 2020-12-04 | 2021-04-06 | 淮阴工学院 | FPC defect detection method based on Bayesian compressed sensing and deep learning |
| CN112634391A (en) * | 2020-12-29 | 2021-04-09 | 华中科技大学 | Gray level image depth reconstruction and fault diagnosis system based on compressed sensing |
| CN112634391B (en) * | 2020-12-29 | 2023-12-29 | 华中科技大学 | Gray image depth reconstruction and fault diagnosis system based on compressed sensing |
| CN113281029A (en) * | 2021-06-09 | 2021-08-20 | 重庆大学 | Rotating machinery fault diagnosis method and system based on multi-scale network structure |
| CN113281029B (en) * | 2021-06-09 | 2022-03-15 | 重庆大学 | Rotating machinery fault diagnosis method and system based on multi-scale network structure |
| CN113609939A (en) * | 2021-07-26 | 2021-11-05 | 吉林大学 | Moving target efficient identification method based on compressed observation |
| CN114964782A (en) * | 2022-06-02 | 2022-08-30 | 昆明理工大学 | Rolling bearing fault diagnosis acoustic emission signal processing method and system based on compressed sensing |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109858352B (en) | 2020-09-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109858352A (en) | A kind of method for diagnosing faults based on compressed sensing and the multiple dimensioned network of improvement | |
| CN109765053B (en) | Rolling bearing fault diagnosis method using convolutional neural network and kurtosis index | |
| CN109186973B (en) | Mechanical fault diagnosis method of unsupervised deep learning network | |
| CN106980822B (en) | A kind of rotary machinery fault diagnosis method based on selective ensemble study | |
| CN110110768B (en) | Rolling bearing fault diagnosis method based on parallel feature learning and multiple classifiers | |
| CN113255078A (en) | Bearing fault detection method and device under unbalanced sample condition | |
| CN111914883B (en) | Spindle bearing state evaluation method and device based on deep fusion network | |
| CN109782603A (en) | The detection method and monitoring system of rotating machinery coupling fault | |
| CN106017876A (en) | Wheel set bearing fault diagnosis method based on equally-weighted local feature sparse filter network | |
| CN110334764A (en) | Rotating machinery intelligent failure diagnosis method based on integrated depth self-encoding encoder | |
| CN111458144B (en) | Wind driven generator fault diagnosis method based on convolutional neural network | |
| CN110188774A (en) | A kind of current vortex scan image classifying identification method based on deep learning | |
| CN110263767A (en) | In conjunction with the intelligent Rotating Shaft Fault method of compressed data acquisition and deep learning | |
| CN112304614A (en) | End-to-end rolling bearing intelligent fault diagnosis method adopting multi-attention machine system | |
| CN111595541A (en) | Multi-dimensional structure damage identification method based on mass vibration transmissibility data convolutional neural network processing | |
| CN114091504A (en) | Rotary machine small sample fault diagnosis method based on generation countermeasure network | |
| CN113516638A (en) | Neural network internal feature importance visualization analysis and feature migration method | |
| CN112541510A (en) | Intelligent fault diagnosis method based on multi-channel time series data | |
| CN112766200A (en) | Planetary gear box small sample fault diagnosis method based on one-dimensional convolutional neural network | |
| CN113705096A (en) | Small sample deep learning-based impact fault diagnosis | |
| CN115060494A (en) | Fault diagnosis method for rolling bearing | |
| CN113916535A (en) | Bearing diagnosis method, system, equipment and medium based on time frequency and CNN | |
| CN116541771A (en) | Unbalanced sample bearing fault diagnosis method based on multi-scale feature fusion | |
| CN117056678A (en) | Machine pump equipment operation fault diagnosis method and device based on small sample | |
| CN116644273A (en) | Fault diagnosis method and system based on interpretability multiplication convolution network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |