CN109596326A - Rotary machinery fault diagnosis method based on optimization structure convolutional neural networks - Google Patents
Rotary machinery fault diagnosis method based on optimization structure convolutional neural networks Download PDFInfo
- Publication number
- CN109596326A CN109596326A CN201811451754.3A CN201811451754A CN109596326A CN 109596326 A CN109596326 A CN 109596326A CN 201811451754 A CN201811451754 A CN 201811451754A CN 109596326 A CN109596326 A CN 109596326A
- Authority
- CN
- China
- Prior art keywords
- layer
- neural networks
- convolutional neural
- grayscale image
- input
- 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
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
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
- G01M13/02—Gearings; Transmission mechanisms
- G01M13/028—Acoustic or vibration analysis
-
- G—PHYSICS
- 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
- 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
Abstract
The invention discloses a kind of rotary machinery fault diagnosis methods based on optimization structure convolutional neural networks, collecting work signal first under rotating machinery normal condition and under malfunction, it is then converted into grayscale image, is trained using grayscale image and corresponding faulty tag as convolutional neural networks of the training sample to building；In the rotating machinery course of work, collecting work signal is simultaneously converted to grayscale image, inputs trained convolutional neural networks and carries out fault diagnosis.The present invention completes more classification tasks to rotating machinery fault using grayscale image is converted by the rotating machinery working signal collected, by convolutional neural networks.
Description
Technical field
The invention belongs to Construction Machinery System fault diagnosis technology fields, more specifically, are related to a kind of based on optimization
The rotary machinery fault diagnosis method of structure convolutional neural networks.
Background technique
Rotating machinery is most popular machinery in the industry, with the development of modern industry and mechanical automation degree
It improves, reliability, maintainable and safety all has been to be concerned by more and more people.Rolling bearing is as rotating machinery
One of core component, according to statistics in the rotating machinery using rolling bearing, there are about 30% mechanical breakdowns and bearing to damage
Hurt related.And compared with other machine components components, rolling bearing has the big feature of service life discreteness, thus in real work
In, some bearings, which already exceed its projected life, but can still work normally, and some bearings, which reach far away its service life, but to be occurred
Various failures.If bearing fault without find in time meeting so that machine operation accuracy decline, results even in entire mechanical disorder,
Cause the accident even casualties.Meanwhile rolling bearing is extremely complex dynamic system.When bearing breaks down,
Its dynamic behavior typically exhibits more complicated nonlinear characteristic.Signal not only shows non-stationary, and usually companion
With complicated self-similarity, the nonlinear characteristics such as chaos and point shape are shown, in this case, to be shaken from the bearing of non-stationary
The characteristic parameter that characterization bearing system dynamics behavior is extracted in dynamic signal, to identify that the severity of damage of the bearing just becomes very
It is difficult.But equipment is always undergone during the work time by normally to the process for degenerating to ultimate failure, if it is possible to prison in real time
Measurement equipment health and fitness information has positive meaning for the formulation of maintenance strategy, reduction maintenance cost and production loss.
Although traditional intelligent Fault Diagnose Systems are theoretical mature, method multiplicity, in face of what is become increasingly complex now
Intelligent industrial equipment has been unable to satisfy requirement.Firstly, the method for machine learning can not directly be needed using original signal mostly
Feature extraction is carried out by the feature extractor artificially designed, and feature extraction relies on priori knowledge, and during the extraction process can
Lose many information.And existing most Fault Classification is all by pretreatment (wigner-ville distribution, wavelet transformation, Empirical Mode
State decompose etc.) after, coarse abstraction sequence feature.Time-frequency characteristics are sent into support vector machines, extract MFDFA spy after extracting
K clustering and discriminant is sent into after being sent into geneva judgement system, the extraction data information entropy based on informatics after sign.These methods can be with
Effective processing nonstationary time series, but due to having concealed original time series, then use the side for relying on tagsort
Formula, it is difficult to which the information for comprehensively expressing original time series cannot achieve high-precision classification.Traditional machine learning model is most
Whole accuracy of identification depends critically upon the feature extracted, and feature extraction has two: first is that treated data
Whether with the good ability for stating signal characteristic, second is that preprocessing process extremely expends the time.Therefore make engineering is practical
Used time, traditional rote learning model have that speed is slow, accuracy rate is low, need to research and solve.
Summary of the invention
It is an object of the invention to overcome the deficiencies of the prior art and provide a kind of based on optimization structure convolutional neural networks
Rotary machinery fault diagnosis method converts grayscale image for the rotating machinery working signal collected, passes through convolutional Neural net
Network completes more classification tasks to rotating machinery fault.
For achieving the above object, the present invention is based on the rotary machinery fault diagnosis sides of optimization structure convolutional neural networks
Method the following steps are included:
S1: random intercepted length is M under rotating machinery normal condition and under R kind malfunction respectively2Working signal Ln
(m), wherein n=1,2 ..., N, N indicate the quantity of working signal, m=0,1 ..., M2- 1, M2=k × 2d, and M2> T, T table
Show the period of working signal, remembers every section of working signal Ln(m) corresponding label is Yn, label YnIt is corresponding for identification work signal
Rotating machinery working condition；The grayscale image I that size is M × M is converted by every section of working signaln, pixel in grayscale image
The pixel value f of (i, j)n(i, j) is calculated using the following equation:
Wherein, i, j=0,1 ..., M-1, round () indicate round-off function, Maxn、MinnRespectively indicate one-dimensional acceleration
Spend vibration signal Ln(m) maximum value and minimum value；
S2: building obtains convolutional neural networks in accordance with the following methods:
1st layer is convolutional layer Conv1, and the 2nd layer is maximum pond layer Pool1, and the 3rd layer is convolutional layer Conv2, and the 4th layer is
Maximum pond layer Pool2, the 5th layer be convolutional layer Conv3, the 6th layer be cccp layer Cccp1, the 7th layer be cccp layers of Cccp2, the 5th,
6,7 layers of MlpConv layers of composition, the 8th layer is global average pond layer Pool3, and the 9th layer is output layer softmax, wherein
Using the linear unit R elu of amendment as activation primitive in Conv3, Cccp1, Cccp2；
S3: by every width grayscale image I obtained in step S1nAs input, corresponding label YnAs desired output, to step
The convolutional neural networks of rapid S2 building are trained；
S4: in the rotating machinery course of work, acquiring a segment length as needed is M2Working signal L ' (m), use
The working signal is converted the grayscale image I ' that size is M × M by same procedure in step S1n, then by grayscale image I 'nInput
Into the trained convolutional neural networks of step S3, diagnostic result is obtained.
The present invention is based on optimization structure convolutional neural networks rotary machinery fault diagnosis method, first rotating machinery just
With collecting work signal under malfunction under normal state, it is then converted into grayscale image, grayscale image and corresponding faulty tag are made
It is trained for convolutional neural networks of the training sample to building；In the rotating machinery course of work, collecting work signal simultaneously turns
It is changed to grayscale image, trained convolutional neural networks is inputted and carries out fault diagnosis.
The invention has the following advantages:
1) present invention does not do feature extraction to original signal, but is converted into grayscale image, avoids feature and mentions
Bring information is taken to lose, the information for remaining original signal and including of high degree；
2) the characteristics of working signal of the invention for being directed to rotating machinery, improves traditional convolutional neural networks, makes
It with convolution-pond module and MLPConv, that is, ensure that integrally-built simple, and improve the feature representation ability of network,
Traditional full articulamentum is replaced by using the average pond layer of the overall situation simultaneously, avoid because the parameter amount of full articulamentum it is excessive caused by
Over-fitting situation；
3) convolutional neural networks of optimization structure proposed by the present invention have the characteristics that structure is simple, parameter amount is few, therefore compared with
General requirement of the deep learning network to hardware resource is low, deployment use is carried out in mainstream hardware system with can be convenient, together
When experiments have shown that can effectively improve failure modes accuracy rate based on the convolutional neural networks.
Detailed description of the invention
Fig. 1 is that the present invention is based on the specific embodiment parties of the rotary machinery fault diagnosis method of optimization structure convolutional neural networks
Formula flow chart；
Fig. 2 is convolutional neural networks structural schematic diagram constructed in the present embodiment；
Fig. 3 is that two three kinds for comparing LeNIN network in convolutional neural networks and the present invention deform in the present embodiment
Failure modes accuracy rate curve graph；
Fig. 4 is the failure modes accuracy rate curve graph that LeNIN network different parameters combine in the present invention；
Fig. 5 is failure point of the convolutional neural networks in different size input gray level figures of the convolution kernel combination of 64-64-192
Class accuracy rate curve graph；.
Fig. 6 is the failure modes accuracy rate histogram of the present embodiment Chinese and Western storage university's bearing fault data；
Fig. 7 is the failure modes accuracy rate histogram of self-built platform bearing fault data in the present embodiment.
Specific embodiment
A specific embodiment of the invention is described with reference to the accompanying drawing, preferably so as to those skilled in the art
Understand the present invention.Requiring particular attention is that in the following description, when known function and the detailed description of design perhaps
When can desalinate main contents of the invention, these descriptions will be ignored herein.
Fig. 1 is that the present invention is based on the specific embodiment parties of the rotary machinery fault diagnosis method of optimization structure convolutional neural networks
Formula flow chart.As shown in Figure 1, the present invention is based on the specific of the rotary machinery fault diagnosis method of optimization structure convolutional neural networks
Step includes:
S101: training sample is obtained:
Firstly the need of obtain training sample, the present invention do not use directly the working signal in the rotating machinery course of work as
Sample, but grayscale image first is converted by original working signal data, using grayscale image as training sample, then it is based on grayscale image
Carry out subsequent processing.The working signal of rotating machinery generally has one-dimensional acceleration vibration signal, displacement signal etc., can be according to reality
Border needs to select.
Training sample is obtained in the present invention method particularly includes: respectively under rotating machinery normal condition and R kind failure shape
Random intercepted length is M under state2Working signal Ln(m), wherein n=1,2 ..., N, N indicate the quantity of working signal, m=0,
1,...,M2- 1, M2=k × 2d, k > 1, d > 1, and M2> T, T indicate the period of working signal, remember every section of working signal Ln
(m) corresponding label is Yn, label YnWorking condition for the corresponding rotating machinery of identification work signal, it is clear that label YnIt deposits
In R+1 value.The grayscale image I that size is M × M is converted by every section of working signaln, the picture of pixel (i, j) in grayscale image
Plain value fn(i, j) is calculated using the following equation:
Wherein, i, j=0,1 ..., M-1, round () indicate round-off function, Maxn、MinnRespectively indicate working signal
Ln(m) maximum value and minimum value.
Obviously, according to the above operation, the pixel value in grayscale image can be made to be located at 0-255.As it can be seen that the present invention is by work
Signal is divided into M subsequence, each subsequence corresponding grey scale figure InIn one-row pixels point.
The size of grayscale image is come in conjunction with Practical Project system failure classification accuracy needs and hardware device performance characteristic
It chooses.In general, if original data volume is small, grayscale image size is with regard to less than normal, and vice versa.In addition it is also necessary in conjunction with
Design parameter in convolutional neural networks is arranged, and selection makes the higher size of final classification accuracy rate.
S102: building optimization structure/lightweight convolutional neural networks:
In the present invention, the characteristics of one-dimensional acceleration vibration signal in engineering machine in working process, is analyzed, in conjunction with
Using the MLPConv layer in traditional convolution-pond structure and NIN network, and it is refreshing with the average pond layer substitution exemplary convolution of the overall situation
Through the full articulamentum in network, a novel convolutional neural networks are constructed --- LeNIN network, specific structure are as follows: the 1st
Layer is convolutional layer Conv1, and the 2nd layer is maximum pond layer Pool1, and the 3rd layer is convolutional layer Conv2, and the 4th layer is maximum pond layer
Pool2, the 5th layer is convolutional layer Conv3, and the 6th layer is cccp (cascaded cross channel parametric
Pooling cascades across channel parameters change ponds) layer Cccp1, the 7th layer is cccp layers of Cccp2, the 5th, 6,7 layer of composition MlpConv
Layer, the 8th layer is global average pond layer Pool3, and the 9th layer is output layer softmax, wherein convolutional layer Conv1, Conv2,
Using the linear unit R elu of amendment as activation primitive in Conv3.
Since the outstanding advantage of CNN network is that local sensing and weight are shared, network institute in training is considerably reduced
The parameter that need to learn, therefore the present invention still uses traditional convolution-pond structure in the first half of LeNIN network.And
MlpConv layers are improved on the basis of common convolutional layer, be joined the convolutional layer (cccp layers) of 1*1, are equivalent to and are passing
Using local receptor field as the input of a miniature neural network on the basis of system convolutional layer, this micronet is equivalent to one
Multi-layer perception (MLP), i.e., one sub-network being made of multiple full articulamentums.The present invention is combined in the latter half of LeNIN network to be made
With MLPConv layers, the feature representation ability of LeNIN network can be effectively promoted.
In addition, it is exactly that parameter amount is excessive that the full articulamentum in tradition CNN network, which but has a very typical weakness, especially
It is the full articulamentum being connected with the last one convolutional layer.The present invention is substituted in LeNIN network using global average pond layer complete
Articulamentum eliminates the process that characteristic pattern is unfolded compared with traditional full articulamentum, directly uses and characteristic pattern equidimension
Sliding window carries out average pond, smaller compared to full articulamentum calculation amount, alleviates hardware burden while effectively preventing connecting entirely
Layer parameter amount it is excessive and caused by over-fitting situation.
S103: training convolutional neural networks:
By every width grayscale image I obtained in step S101nAs input, corresponding label YnAs desired output, to step
The convolutional neural networks of S102 building are trained.
Presently, there are the training method of a variety of convolutional neural networks, back-propagation algorithm is selected in the present embodiment and accelerates ladder
Degree descent method Nesterov is trained.
S104: fault diagnosis:
In the rotating machinery course of work, acquiring a segment length as needed is M2Working signal L ' (m), using step
The working signal is converted the grayscale image I ' that size is M × M by same procedure in S101n, then by grayscale image I 'nIt is input to
In the trained convolutional neural networks of step S103, diagnostic result is obtained.
Embodiment
Technical solution and technical effect in order to better illustrate the present invention, using a specific example to work of the invention
Make process and technical effect carries out analytic explanation.Bearing fault is one of rotating machinery typical fault, therefore the present embodiment
The open data of bearing fault of U.S.'s Case Western Reserve University and the bearing fault number of self-built fault simulation platform acquisition are respectively adopted
According to experiment test is carried out, used data are one-dimensional acceleration vibration signal data.
For Xi Chu university bearing fault data, choosing 3 kinds of fault types is respectively rolling element failure (B), inner ring failure
(IR), outer ring failure (OR) and one group of normal data (NR)；Every kind of fault type is divided into 4 kinds of fault degrees again, respectively
0.18mm,0.36mm,0.54mm,1mm；Step of the present invention is used to the one-dimensional acceleration vibration signal of above several fault modes
Method in S101 is handled to obtain grayscale image.By way of randomly selecting, training set totally 8000 samples are obtained, are tested
Collect totally 2000 samples；Each sample includes 784 (for the grayscale image for constituting 28*28) a sampled points.Table 1 is base in the present embodiment
In the bearing fault database of Xi Chu university bearing fault data building.
Table 1
Bearing inner race failure (IR), outer ring failure (OR) and gear sun gear failure is respectively set in self-built fault simulation platform
(Sun), and every kind of bearing fault type there are two types of different faults degree: I2 (failure axial dimension Δ φf=35.7 °, failure is deep
Spend d=0.3mm), I3 (Δ φf=64.3 °, d=0.3mm), O2 (Δ φf=38.6 °, d=0.3mm), O4 (Δ φf=1 °, d
=0.3mm).800 samples are intercepted at random to every kind of fault degree under each type of every class failure part under different rotating speeds,
Sample length is the grayscale image matrix of 4096 data points (facilitate and constitute 64*64), and training set includes 9600 samples, test set
Include 2800 samples.Table 2 is the bearing fault database based on the acquisition building of self-built fault simulation platform in the present embodiment.
Table 2
Obtained two parts bearing fault data are converted into grayscale image, are then divided into training set and test set, training
Collection accounts for the 75% of overall data, and training set is used for the training of convolutional neural networks (LeNIN network), and test set is used for convolution mind
It is tested through network, with statistical classification accuracy rate.
Fig. 2 is the structural schematic diagram of convolutional neural networks constructed in the present embodiment.Table 3 is convolution mind in the present embodiment
Parameter configuration table through network.
| Layer | Port number | Core size | Step-length | Zero padding number |
| Conv1 | 64 | 5 | 1 | 0 |
| Pool1 | 64 | 2 | 2 | 0 |
| Conv2 | 64 | 5 | 1 | 0 |
| Pool2 | 64 | 2 | 2 | 0 |
| Conv3 | 192 | 3 | 1 | 1 |
| Cccp1 | 192 | 1 | 1 | 0 |
| Cccp2 | 192 | 1 | 1 | 0 |
| Pool3 | 192 | 4 | 1 | 0 |
Table 3
As shown in table 3, the convolution kernel that convolutional layer Conv1 is 5 × 5 using 64 sizes in the present embodiment, to the ash of input
Degree figure does convolution with step-length for 1；The weights initialisation of each convolutional layer is Xavier in the present embodiment, and bias is initialized as often
Number 0.
It is the 2 maximum pondization operations for doing 2 × 2 that maximum pond layer Pool1, which is input to the data of Pool1 to Conv1 with step-length,；
The convolution kernel that convolutional layer Conv2 is 5 × 5 using 64 sizes, is input to the data of Conv2 to Pool1 with step-length
Convolution is done for 1.
It is the 2 maximum pondization operations for doing 2 × 2 that maximum pond layer Pool2, which is input to the data of Pool2 to Conv2 with step-length,.
The convolution kernel that convolutional layer Conv3 is 3 × 3 using 192 sizes, is input to the data of Conv2 to Pool1 with step-length
Convolution is done for 1, and carries out 1 zero padding.Zero padding operation is to prevent dimension from losing with zero padding for controlling feature size,
Its concrete operations are as follows: vacation lets m represent the size (wide or high) of input unit, and W indicates the size (wide or high) of output unit, F table
Show convolution kernel size (kernel size), S indicates that stride, P indicate zero padding (zero padding) quantity.Then output unit
Size are as follows:
The convolution kernel that cccp layers of Cccp1 are 1 × 1 using 192 sizes, is input to the data of Cccp1 to Conv3 with step-length
Convolution is done for 1.
The convolution kernel that cccp layers of Cccp2 are 1 × 1 using 192 sizes, is input to the data of Cccp2 to Cccp1 with step-length
Convolution is done for 1.
The average pond layer Pool3 of the overall situation exports the convolution kernel of characteristic pattern equidimension using 192 sizes and Cccp2, right
Cccp2 is input to the data of Pool3 and does average pond with step-length for 1.
Output layer Softmax handles Pool3 layers of linear convergent rate, and the grayscale image inputted belongs to R+1 event
Hinder the relative probability of status categories.
In the present embodiment, LeNIN network is instructed using back-propagation algorithm and accelerating gradient descent method Nesterov
Practice, is then tested using test set.
Technical effect in order to better illustrate the present invention, using in two comparison convolutional neural networks and the present invention
Three kinds of deformations of LeNIN network compare experiment, count failure modes accuracy rate, wherein each network carries out 10 instructions respectively
Practice and tests.Fig. 3 is that two three kinds for comparing LeNIN network in convolutional neural networks and the present invention deform in the present embodiment
Failure modes accuracy rate curve graph.Table 4 is LeNIN network in two comparison convolutional neural networks and the present invention in the present embodiment
Three kinds deformation failure modes accuracy rate statistical forms.
Table 4
In table 4, LeNet is traditional LeNet network, wherein convolution nuclear volume is 1024 in full articulamentum fc；LeNe+
2fc is the LeNet network using two full articulamentums of cascade, wherein nuclear volume is 1536 in full articulamentum fc1, full articulamentum
Nuclear volume is 32 in fc2；LeNIN (1) indicates the volume that Conv1 convolution nuclear volume is 20, Conv2 in convolutional neural networks of the present invention
Product nuclear volume is 50；LeNIN (2) indicates that the convolution nuclear volume of Conv1_1, Conv1_2 are 20, and convolution kernel size is 3 × 3,
The convolution nuclear volume of Conv2_1, Conv2_2 are 50, and convolution kernel size is 3 × 3；It is equivalent to the convolutional layer etc. with two 3 × 3
The convolutional layer of effect substitution original 5 × 5, keeps convolution nuclear volume constant；LeNIN (3) indicates holding LeNIN (2), and remaining is constant, will
The convolution nuclear volume of Conv1_1, Conv1_2, Conv2_1, Conv2_2 become 64.
From Fig. 3 and table 4 as can be seen that LeNIN network failure modes accuracy rate with higher proposed by the invention.
Next the LeNIN network under different parameters (input gray level figure size, convolution nuclear volume) combination is tested,
LeNIN network under each parameter combination carries out 10 training and test respectively.Fig. 4 is that LeNIN network difference is joined in the present invention
The failure modes accuracy rate curve graph that array is closed.Table 5 is that the failure modes in the present invention under the combination of LeNIN network different parameters are quasi-
True rate statistical form.
Table 5
As shown in table 5, under the convolution kernel combination of 64-64-192, classification accuracy is best.Fig. 5 is the volume of 64-64-192
Failure modes accuracy rate curve graph of the convolutional neural networks of product core combination in different size input gray level figures.It can from table 5 and Fig. 5
To find, under consolidated network structure, when inputting various sizes of grayscale image, classification accuracy can also be had differences, therefore in reality
The size of grayscale image can be determined in the application of border according to experiment.
Next two bearing fault databases in the present embodiment are respectively adopted and constitute two groups of training sets and sample set, select
Grayscale image be respectively size 48*48,64*64, the convolution kernel that convolutional neural networks structure is 64-64-192 combines, and uses two
A database carries out 10 training and test respectively.Fig. 6 is the failure modes of the present embodiment Chinese and Western storage university's bearing fault data
Accuracy rate histogram.Fig. 7 is the failure modes accuracy rate histogram of self-built platform bearing fault data in the present embodiment.Table 6 is
Classification accuracy statistical form in the present embodiment under two bearing fault databases.
Table 6
As shown in table 6, the present invention has good classifying quality for bearing fault, and passes through self-built platform rotating machinery
The verifying of (including gear) fault data, it can be seen that LeNIN network proposed by the invention has generalization ability, in face of multiple
It closes failure modes and also maintains preferable classifying quality.
Although the illustrative specific embodiment of the present invention is described above, in order to the technology of the art
Personnel understand the present invention, it should be apparent that the present invention is not limited to the range of specific embodiment, to the common skill of the art
For art personnel, if various change the attached claims limit and determine the spirit and scope of the present invention in, these
Variation is it will be apparent that all utilize the innovation and creation of present inventive concept in the column of protection.
Claims (2)
1. a kind of rotary machinery fault diagnosis method based on optimization structure convolutional neural networks, which is characterized in that including following
Step:
S1: random intercepted length is M under rotating machinery normal condition and under R kind malfunction respectively2Working signal Ln(m),
Wherein n=1,2 ..., N, N indicate the quantity of working signal, m=0,1 ..., M2- 1, M2=k × 2d, and M2> T, T indicate work
Make the period of signal, remembers every section of working signal Ln(m) corresponding label is Yn, label YnFor the corresponding rotation of identification work signal
The working condition of favourable turn tool；The grayscale image I that size is M × M is converted by every section of working signaln, pixel (i, j) in grayscale image
Pixel value fn(i, j) is calculated using the following equation:
Wherein, i, j=0,1 ..., M-1, round () indicate round-off function, Maxn、MinnRespectively indicate one-dimensional acceleration vibration
Signal Ln(m) maximum value and minimum value；
S2: building obtains convolutional neural networks in accordance with the following methods:
1st layer is convolutional layer Conv1, and the 2nd layer is maximum pond layer Pool1, and the 3rd layer is convolutional layer Conv2, and the 4th layer is maximum
Pond layer Pool2, the 5th layer is convolutional layer Conv3, and the 6th layer is cccp layers of Cccp1, and the 7th layer is cccp layers of Cccp2, the 5th, 6,7
Layer constitutes MlpConv layer, and the 8th layer is global average pond layer Pool3, and the 9th layer is output layer softmax, wherein Conv3,
Using the linear unit R elu of amendment as activation primitive in Cccp1, Cccp2；
S3: by every width grayscale image I obtained in step S1nAs input, corresponding label YnAs desired output, to step S2
The convolutional neural networks of building are trained；
S4: in the rotating machinery course of work, acquiring a segment length as needed is M2Working signal L ' (m), using step
The working signal is converted the grayscale image I ' that size is M × M by same procedure in S101n, then by grayscale image I 'nIt is input to
In the trained convolutional neural networks of step S103, diagnostic result is obtained.
2. Construction Machinery System method for diagnosing faults according to claim 1, which is characterized in that the convolutional neural networks
Design parameter it is as follows:
The convolution kernel that convolutional layer Conv1 is 5 × 5 using 64 sizes does convolution with step-length to the grayscale image of input for 1；
It is the 2 maximum pondization operations for doing 2 × 2 that maximum pond layer Pool1, which is input to the data of Pool1 to Conv1 with step-length,；
The convolution kernel that convolutional layer Conv2 is 5 × 5 using 64 sizes is input to the data of Conv2 to Pool1 and is done with step-length for 1
Convolution；
It is the 2 maximum pondization operations for doing 2 × 2 that maximum pond layer Pool2, which is input to the data of Pool2 to Conv2 with step-length,；
The convolution kernel that convolutional layer Conv3 is 3 × 3 using 192 sizes is input to the data of Conv2 to Pool1 with step-length as 1
Convolution is done, and carries out 1 zero padding；
The convolution kernel that cccp layers of Cccp1 are 1 × 1 using 192 sizes is input to the data of Cccp1 to Conv3 with step-length as 1
Do convolution；
The convolution kernel that cccp layers of Cccp2 are 1 × 1 using 192 sizes is input to the data of Cccp2 to Cccp1 with step-length as 1
Do convolution；
The average pond layer Pool3 of the overall situation is input to Cccp2 using the convolution kernel of the sizes such as 192 and Cccp2 output characteristic pattern
The data of Pool3 do average pond with step-length for 1；
Output layer Softmax handles Pool3 layers of linear convergent rate, and the grayscale image inputted belongs to R+1 failure shape
The relative probability of state classification.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811451754.3A CN109596326B (en) | 2018-11-30 | 2018-11-30 | Rotary machine fault diagnosis method based on convolution neural network with optimized structure |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811451754.3A CN109596326B (en) | 2018-11-30 | 2018-11-30 | Rotary machine fault diagnosis method based on convolution neural network with optimized structure |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109596326A true CN109596326A (en) | 2019-04-09 |
| CN109596326B CN109596326B (en) | 2020-06-12 |
Family
ID=65959190
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811451754.3A Active CN109596326B (en) | 2018-11-30 | 2018-11-30 | Rotary machine fault diagnosis method based on convolution neural network with optimized structure |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109596326B (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110068462A (en) * | 2019-05-14 | 2019-07-30 | 北京科技大学 | A kind of motor bearings method for diagnosing faults and device |
| CN110261116A (en) * | 2019-07-08 | 2019-09-20 | 华南理工大学 | A kind of Bearing Fault Detection Method and device |
| CN110307983A (en) * | 2019-06-19 | 2019-10-08 | 电子科技大学 | Unmanned plane Method for Bearing Fault Diagnosis based on CNN-Bagging |
| CN110307982A (en) * | 2019-06-19 | 2019-10-08 | 电子科技大学 | Bearing fault classification method based on CNN and Adaboost |
| CN110333076A (en) * | 2019-06-19 | 2019-10-15 | 电子科技大学 | Method for Bearing Fault Diagnosis based on CNN-Stacking |
| CN110346692A (en) * | 2019-08-20 | 2019-10-18 | 国网河南省电力公司电力科学研究院 | A kind of wire selection method for power distribution network single phase earthing failure based on time-frequency image information |
| CN110361176A (en) * | 2019-06-05 | 2019-10-22 | 华南理工大学 | A kind of intelligent failure diagnosis method for sharing neural network based on multitask feature |
| CN110414709A (en) * | 2019-06-18 | 2019-11-05 | 重庆金融资产交易所有限责任公司 | Debt risk intelligent Forecasting, device and computer readable storage medium |
| 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 |
| CN110672343A (en) * | 2019-09-29 | 2020-01-10 | 电子科技大学 | Rotary machine fault diagnosis method based on multi-attention convolutional neural network |
| CN110781751A (en) * | 2019-09-27 | 2020-02-11 | 杭州电子科技大学 | Emotional electroencephalogram signal classification method based on cross-connection convolutional neural network |
| CN110916623A (en) * | 2019-11-11 | 2020-03-27 | 浙江大学 | Small-sized flexible pulse diagnosis instrument capable of applying pressure actively and detection method thereof |
| CN111027641A (en) * | 2019-12-25 | 2020-04-17 | 许昌北邮万联网络技术有限公司 | Fault diagnosis method and device for rotary mechanical equipment |
| CN112577747A (en) * | 2020-12-07 | 2021-03-30 | 东南大学 | Rolling bearing fault diagnosis method based on space pooling network |
| CN112633132A (en) * | 2020-12-18 | 2021-04-09 | 合肥工业大学 | Bearing fault diagnosis method and system based on two-dimensional vibration image enhancement |
| CN112683533A (en) * | 2020-12-18 | 2021-04-20 | 合肥工业大学 | Signal enhancement method and system for bearing fault diagnosis |
| CN114325352A (en) * | 2022-01-04 | 2022-04-12 | 电子科技大学 | Analog filter circuit fault diagnosis method based on empirical wavelet transform |
| CN114444622A (en) * | 2022-04-11 | 2022-05-06 | 中国科学院微电子研究所 | Fruit detection system and method based on neural network model |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160117587A1 (en) * | 2014-10-27 | 2016-04-28 | Zhicheng Yan | Hierarchical deep convolutional neural network for image classification |
| CN108010016A (en) * | 2017-11-20 | 2018-05-08 | 华中科技大学 | A kind of data-driven method for diagnosing faults based on convolutional neural networks |
| CN108319928A (en) * | 2018-02-28 | 2018-07-24 | 天津大学 | A kind of deep learning model and application based on Multi-objective PSO optimization |
| CN108426713A (en) * | 2018-02-26 | 2018-08-21 | 成都昊铭科技有限公司 | Rolling bearing Weak fault diagnostic method based on wavelet transformation and deep learning |
| CN108510153A (en) * | 2018-02-08 | 2018-09-07 | 同济大学 | A kind of multi-state rotary machinery fault diagnosis method |
| CN108647786A (en) * | 2018-07-10 | 2018-10-12 | 电子科技大学 | The rotating machinery on-line fault monitoring method of neural network is fought based on depth convolution |
-
2018
- 2018-11-30 CN CN201811451754.3A patent/CN109596326B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160117587A1 (en) * | 2014-10-27 | 2016-04-28 | Zhicheng Yan | Hierarchical deep convolutional neural network for image classification |
| CN108010016A (en) * | 2017-11-20 | 2018-05-08 | 华中科技大学 | A kind of data-driven method for diagnosing faults based on convolutional neural networks |
| CN108510153A (en) * | 2018-02-08 | 2018-09-07 | 同济大学 | A kind of multi-state rotary machinery fault diagnosis method |
| CN108426713A (en) * | 2018-02-26 | 2018-08-21 | 成都昊铭科技有限公司 | Rolling bearing Weak fault diagnostic method based on wavelet transformation and deep learning |
| CN108319928A (en) * | 2018-02-28 | 2018-07-24 | 天津大学 | A kind of deep learning model and application based on Multi-objective PSO optimization |
| CN108647786A (en) * | 2018-07-10 | 2018-10-12 | 电子科技大学 | The rotating machinery on-line fault monitoring method of neural network is fought based on depth convolution |
Cited By (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110068462A (en) * | 2019-05-14 | 2019-07-30 | 北京科技大学 | A kind of motor bearings method for diagnosing faults and device |
| CN110361176B (en) * | 2019-06-05 | 2021-11-19 | 华南理工大学 | Intelligent fault diagnosis method based on multitask feature sharing neural network |
| CN110361176A (en) * | 2019-06-05 | 2019-10-22 | 华南理工大学 | A kind of intelligent failure diagnosis method for sharing neural network based on multitask feature |
| CN110414709A (en) * | 2019-06-18 | 2019-11-05 | 重庆金融资产交易所有限责任公司 | Debt risk intelligent Forecasting, device and computer readable storage medium |
| CN110307982A (en) * | 2019-06-19 | 2019-10-08 | 电子科技大学 | Bearing fault classification method based on CNN and Adaboost |
| CN110307982B (en) * | 2019-06-19 | 2021-01-26 | 电子科技大学 | Bearing fault classification method based on CNN and Adaboost |
| CN110333076A (en) * | 2019-06-19 | 2019-10-15 | 电子科技大学 | Method for Bearing Fault Diagnosis based on CNN-Stacking |
| CN110333076B (en) * | 2019-06-19 | 2021-01-26 | 电子科技大学 | Bearing fault diagnosis method based on CNN-Stacking |
| CN110307983A (en) * | 2019-06-19 | 2019-10-08 | 电子科技大学 | Unmanned plane Method for Bearing Fault Diagnosis based on CNN-Bagging |
| CN110261116A (en) * | 2019-07-08 | 2019-09-20 | 华南理工大学 | A kind of Bearing Fault Detection Method and device |
| CN110346692A (en) * | 2019-08-20 | 2019-10-18 | 国网河南省电力公司电力科学研究院 | A kind of wire selection method for power distribution network single phase earthing failure based on time-frequency image information |
| 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 |
| 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 |
| CN110781751A (en) * | 2019-09-27 | 2020-02-11 | 杭州电子科技大学 | Emotional electroencephalogram signal classification method based on cross-connection convolutional neural network |
| CN110672343A (en) * | 2019-09-29 | 2020-01-10 | 电子科技大学 | Rotary machine fault diagnosis method based on multi-attention convolutional neural network |
| CN110916623A (en) * | 2019-11-11 | 2020-03-27 | 浙江大学 | Small-sized flexible pulse diagnosis instrument capable of applying pressure actively and detection method thereof |
| CN111027641A (en) * | 2019-12-25 | 2020-04-17 | 许昌北邮万联网络技术有限公司 | Fault diagnosis method and device for rotary mechanical equipment |
| CN111027641B (en) * | 2019-12-25 | 2024-02-13 | 许昌北邮万联网络技术有限公司 | Fault diagnosis method and device for rotary mechanical equipment |
| CN112577747A (en) * | 2020-12-07 | 2021-03-30 | 东南大学 | Rolling bearing fault diagnosis method based on space pooling network |
| CN112577747B (en) * | 2020-12-07 | 2023-08-08 | 东南大学 | Rolling bearing fault diagnosis method based on space pooling network |
| CN112683533A (en) * | 2020-12-18 | 2021-04-20 | 合肥工业大学 | Signal enhancement method and system for bearing fault diagnosis |
| CN112633132A (en) * | 2020-12-18 | 2021-04-09 | 合肥工业大学 | Bearing fault diagnosis method and system based on two-dimensional vibration image enhancement |
| CN114325352A (en) * | 2022-01-04 | 2022-04-12 | 电子科技大学 | Analog filter circuit fault diagnosis method based on empirical wavelet transform |
| CN114325352B (en) * | 2022-01-04 | 2023-04-18 | 电子科技大学 | Analog filter circuit fault diagnosis method based on empirical wavelet transform |
| CN114444622A (en) * | 2022-04-11 | 2022-05-06 | 中国科学院微电子研究所 | Fruit detection system and method based on neural network model |
| CN114444622B (en) * | 2022-04-11 | 2022-06-17 | 中国科学院微电子研究所 | Fruit detection system and method based on neural network model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109596326B (en) | 2020-06-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109596326A (en) | Rotary machinery fault diagnosis method based on optimization structure convolutional neural networks | |
| Ding et al. | Energy-fluctuated multiscale feature learning with deep convnet for intelligent spindle bearing fault diagnosis | |
| CN110361176B (en) | Intelligent fault diagnosis method based on multitask feature sharing neural network | |
| CN111797567B (en) | Bearing fault classification method and system based on deep learning network | |
| CN107526853B (en) | Rolling bearing fault mode identification method and device based on stacking convolutional network | |
| CN109781411B (en) | Bearing fault diagnosis method combining improved sparse filter and KELM | |
| CN110334764A (en) | Rotating machinery intelligent failure diagnosis method based on integrated depth self-encoding encoder | |
| CN109782603A (en) | The detection method and monitoring system of rotating machinery coupling fault | |
| CN111351665B (en) | Rolling bearing fault diagnosis method based on EMD and residual error neural network | |
| CN110031227A (en) | A kind of Rolling Bearing Status diagnostic method based on binary channels convolutional neural networks | |
| CN110608884B (en) | Rolling bearing state diagnosis method based on self-attention neural network | |
| CN109190464A (en) | Mechanical breakdown intelligent diagnosing method based on transfer learning under a kind of variable working condition | |
| CN113177937B (en) | Improved YOLOv 4-tiny-based cloth defect detection method | |
| CN107451340A (en) | Rotating machinery fault quantitative Diagnosis method based on more attribute convolutional neural networks | |
| CN106769051B (en) | A kind of rolling bearing remaining life prediction technique based on MCEA-KPCA and combination S VR | |
| CN113505655A (en) | Bearing fault intelligent diagnosis method for digital twin system | |
| Deng et al. | Rolling bearing fault diagnosis based on Deep Boltzmann machines | |
| CN113916535B (en) | Bearing diagnosis method, system, equipment and medium based on time frequency and CNN | |
| Neupane et al. | Deep learning-based bearing fault detection using 2-D illustration of time sequence | |
| Guo et al. | A method of rolling bearing fault diagnose based on double sparse dictionary and deep belief network | |
| CN113221946B (en) | Method for diagnosing fault types of mechanical equipment | |
| Euldji et al. | Decision-making based on decision tree for ball bearing monitoring | |
| CN112146880B (en) | Intelligent diagnosis method for internal structure faults of rolling bearing at different rotating speeds | |
| Chen et al. | An adversarial learning framework for zero-shot fault recognition of mechanical systems | |
| CN112766200A (en) | Planetary gear box small sample fault diagnosis method based on one-dimensional convolutional neural 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 |