CN106874957A - A kind of Fault Diagnosis of Roller Bearings - Google Patents

Abstract

The present invention relates to a kind of Fault Diagnosis of Roller Bearings, its feature extraction tasks needed for utilizing the theoretical learning algorithm of convolutional neural networks to complete fault diagnosis, artificial selection can be independent of, from simple to complex, by rudimentary to the senior substantive characteristics for automatically extracting input data, and energy automatic mining goes out to be hidden in the abundant information in given data；In addition, employ support vector regression method carries out Classification and Identification to test sample, support vector regression has powerful generalization ability, unknown new samples are identified with more preferable precision, Classification and Identification is carried out to sample as grader using support vector regression, the general shortcoming of the grader generalization ability that deep learning can be overcome to give tacit consent to.The present invention can improve the accuracy and validity of rolling bearing fault diagnosis, to solve the problems, such as that rolling bearing fault diagnosis provide a kind of new effective way, in can be widely applied to the complex mechanical system fault diagnosis in the fields such as chemical industry, metallurgy, electric power, aviation.

Description

A kind of Fault Diagnosis of Roller Bearings

Technical field

The invention belongs to mechanical fault diagnosis and Artificial technical field of intelligence, more particularly to one kind is based on convolution god Through network and the Fault Diagnosis of Roller Bearings of support vector regression.

Background technology

Rolling bearing is one of mostly important critical component in rotating machinery, rolling bearing be widely used in chemical industry, Each key areas such as metallurgy, electric power, aviation, but it is also often in the severe working environment such as high temperature, high speed, heavy duty simultaneously In, it is one of most flimsy element to cause rolling bearing.Bearing performance is directly influenced with the quality of operating mode and is further associated Axle and the gear in rotating shaft or even entire machine equipment performance, its defect can cause equipment produce abnormal vibrations And noise, or even device damage is caused, in fact, the probability that mechanical failure problem is attributed to bearing fault is very high.Therefore, it is right Rolling bearing fault is diagnosed, and especially for the analysis of early incipient failure, realizes that fast and accurately bearing fault monitoring is right It is significant in the normal work of plant equipment and safety in production.

Feature extraction is substantially a kind of conversion, is turned sample in different spaces by way of mapping or converting Change.Mechanical breakdown feature extracting method conventional at present mainly have Fourier transformation (Fourier Transform, abbreviation FT), Fast Fourier Transform (FFT) (Fast Fourier Transform, abbreviation FFT), wavelet transformation (Wavelet Transform, letter Claim WT) and empirical mode decomposition (Empirical Mode Decomposition, abbreviation EMD), Hilbert-Huang transform (Hilbert-Huang Transform, abbreviation HHT) etc..

Fourier transformation can clearly rapidly process signal as Linear Time-Frequency Analysis method, with certain time-frequency Resolution ratio, its flexibility and practicality are more protruded, but because Fourier transformation is expression of the signal in frequency domain, time resolution Rate is zero, and it has uncertainty to non-linear, non-stationary signal, causes its range of application to have limitation.FFT methods Overall picture and Localization Problems of the signal in time domain and frequency domain cannot simultaneously be taken into account.Wavelet transformation can carry out office to temporal frequency Portionization is analyzed, and reaches high frequency treatment time subdivision, and frequency subdivision, is adaptively analyzed to time frequency signal at low frequency, but small Ripple base is different, and decomposition result is different, the more difficult selection of wavelet basis.Signal decomposition can be multiple IMF (Intrinsic by EMD methods Mode Function, intrinsic mode function) all IMF components are Hilbert and become the when frequency division that transducing obtains signal by component Cloth, but there is also some problems in theory, such as the mode in EMD methods is obscured, deficient envelope, envelope, end effect excessively are asked Topic, is among research.It, by the EMD merogenesis of signal, is the flat culture of non-stationary signal that HHT is, it has broken away from linear peace The constraint of stability, there is high accuracy to jump signal.

The feature extracting method for being used at present is based on signal processing technology, mainly based on artificial extraction, fault diagnosis Accuracy of identification depend on the good and bad degree of feature extraction.

In view of above-mentioned defect, the design people is actively subject to research and innovation, a kind of based on convolutional neural networks to found With the Fault Diagnosis of Roller Bearings of support vector regression, to improve the accuracy and validity of rolling bearing fault diagnosis.

The content of the invention

In order to solve the above technical problems, being returned based on convolutional neural networks and supporting vector it is an object of the invention to provide one kind The Fault Diagnosis of Roller Bearings returned, it learns to extract effective spy of training sample data first using convolutional neural networks Levy, Classification and Identification is carried out to test sample using support vector regression sorting technique afterwards, so that it is determined that rolling bearing fault work Condition classification, realizes the diagnosis to rolling bearing fault classification, to improve the accuracy and validity of rolling bearing fault diagnosis.

Fault Diagnosis of Roller Bearings of the invention, comprises the following steps：

Step 1：During rolling bearing rotation work under four kinds of different operating modes, gather every respectively by acceleration transducer The vibration acceleration signal that rolling bearing works in different rotating speeds under kind operating mode, carries out noise suppression preprocessing, and adds operating mode label, Using each vibration acceleration signal data under the various operating modes by pre-processing and after adding operating mode label as training sample；

Step 2：Convolutional neural networks model is set up, convolutional neural networks model is trained using training sample, will In training sample input convolution god's machine network model, using there is supervision, successively training method is successively trained and tuning, is obtained The connection weight and offset parameter of convolutional neural networks model；

Step 3：Using the training sample under various operating modes as the convolutional Neural for determining connection weight and offset parameter The input of network model, deep learning is carried out to training sample, using the convolutional Neural net for determining connection weight and offset parameter Network model carries out validity feature extraction to each training sample under every kind of operating mode respectively, obtains each training under every kind of operating mode The corresponding training sample characteristic information of sample, with the training sample features training support vector regression grader for being extracted；

Step 4：Vibration acceleration signal number of the rolling bearing to be measured in rotation work is gathered by acceleration transducer According to, and noise suppression preprocessing is carried out, as test sample；

Step 5：Using test sample as the convolutional neural networks model for determining connection weight and offset parameter input, it is right Test sample carries out deep learning, and test sample is entered using the convolutional neural networks model for determining connection weight and offset parameter Row feature extraction, obtains test sample characteristic signal；

Step 6：Using test feature information as test sample matching characteristic, by each training sample under every kind of operating mode Corresponding training sample characteristic information as matching benchmark, using the support vector regression grader for training to test sample with Training sample carries out matching diagnosis, and the operating mode kind judging belonging to the training sample that will the most be matched with test sample is test specimens This operating mode classification, so as to obtain the fault diagnosis result of rolling bearing to be measured.

Further, the mean square error loss function of the convolutional neural networks model set up in the step 2 is：

Wherein,It is k-th target labels value of sample m,It is corresponding k-th network output valve.

Solution makes the minimum parameter of mean square error loss function set up network, is realized by below equation：

Step 2.1：Convolutional neural networks are by several processes come solution formula (2).

The data input convolutional layer that the first step will be trained, carries out convolution algorithm.

The input of each hidden layer is the output of last layer, and computing formula is as follows：

s_i=ρ (v_i),with v_i=W_i·s_i-1+b_i. (3)

Wherein, W_iIt is the connection weight between convolutional neural networks adjacent two layers, s is the training data of input, b_iIt is convolution Offset parameter between neutral net adjacent two layers, ρ is activation primitive.

According to above-mentioned activation probability, when given training sample is input into visible node layer, using convolutional neural networks After all nodes of the distribution function excitation hidden layer of model, then the excitation of next hidden layer node is carried out, so as to regain New node layer value.

One convolutional layer can include several different convolution characteristic patterns, therefore the output of this layer can be expressed as preceding layer institute There is adding with formula shows as follows for convolution characteristic pattern：

Wherein, symbol * represents convolution algorithm, and convolution algorithm can be expressed as follows：

The second step of convolutional neural networks solution formula (2) is that the feature exported from convolutional layer is input into a down-sampling Layer (being called pond layer), the effect of down-sampling layer is polymerization.Aggregation formula is：

Wherein down () represents down-sampling formula,L layers of the multiplying property biasing of i-th node is represented,Represent l layers I-th additivity biasing of node.

Step 2.2：Last layer output to step 2.1 gained convolutional neural networks, successively training side is supervised using having Method is successively trained and tuning, and concrete mode is：

Weight and biasing are calculated using propagated forward.

Successively propagated as input by the output of last layer of hidden layer of convolutional neural networks model of step 2.1 gained To output layer, the class categories predicted.

With chain type derived function loss function to the gradient of each weight, i.e. sensitivity.Gradient calculation formula is：

In convolutional neural networks, the calculation expression of l layers of gradient (sensitivity) is：

Wherein,Represent each element multiplication.

Thus after obtaining the sensitivity δ of each node in each hidden layer, as the following formula to convolutional neural networks model Connection weight is updated tuning：

Operating mode label according to training sample determines the actual classification result of training sample, will train the classification of prediction output The actual classification result of result and training sample is compared and obtains error in classification, by error in classification successively back-propagation, so that Realize carrying out tuning, the specific formula that connection weight is updated to the connection weight parameter of each layer of convolutional neural networks model For：

Wherein, η is learning rate.

Successively trained by above-mentioned, until obtaining the output of last layer of hidden layer of convolutional neural networks model.

After carrying out tuning to the connection weight of each layer of convolutional neural networks model, whole convolutional neural networks mould is finally determined The connection weight and offset parameter of type.

Further, the step 6 is carried out using support vector regression sorting technique to test sample and training sample The concrete mode matched somebody with somebody is：

Step 6.1：Find an optimal hyperlane for support vector regression.Support vector regression function is defined as follows：

Wherein, x_iIt is the sample characteristics of input,And α_iIt is Lagrange multiplier, b is biasing, and K () is kernel function.

The present invention selects gaussian radial basis function (RBF) kernel function：

Wherein：σ is the parameter of RBF kernel functions.

The optimal problem of support vector regression is：

Wherein, | | w | | is 2 norms of weight, and C is the regularization factor, ξ_iWithIt is slack variable, ε is the limit of error.

Construct following Lagrangian：

Wherein, μ_iIt is on ξ_iLagrange multiplier.

Formula (14) is zero to the partial derivative of w, b and ξ, is obtained：

Formula (15) is updated in formula (14), and the convex optimization problem of its antithesis is converted into by object function is minimized, Obtain convex optimization aim：

Step 6.2：In four kinds of training samples of operating mode, the corresponding label of every kind of operating mode is y, y ∈ { 0,1,2,3 }, is led to Cross the categorised decision function that support vector regression sorting technique obtains M class problems：

Wherein, α_iWithIt is the Lagrange coefficient in categorised decision function；B is the optimal hyperlane of categorised decision function Position parameter；N is four kinds of sums of the training sample of operating mode；K(x_i, x) represent gaussian radial basis function；

Thus the categorised decision function under four kinds of operating modes is obtained；

Step 6.3：Using test sample feature as the input quantity of the corresponding categorised decision function of four kinds of operating modes, survey is calculated Sample eigen as input quantity support vector regression Decision Classes decision function value, i.e., the operating mode kind judging corresponding to it is The operating mode classification of test sample, obtains the fault diagnosis result of rolling bearing to be measured.

Further, four kinds of operating modes are respectively normal operation, the operating of bearing inner race failure, bearing roller failure fortune Turn, bearing outer ring failure is operated.

By such scheme, the present invention at least has advantages below：

1st, Fault Diagnosis of Roller Bearings of the present invention based on convolutional neural networks and support vector regression, using convolution Neural network theory learning algorithm adaptively complete fault diagnosis needed for feature extraction, automatic mining goes out to be hidden in datum Abundant information in, has broken away from the dependence to a large amount of signal transacting knowledge and diagnosis engineering experience, save labour cost and Time, and there is very big advantage in terms of monitoring, diagnosing ability and generalization ability；

2nd, in Fault Diagnosis of Roller Bearings of the present invention based on convolutional neural networks and support vector regression, employ Support vector regression sorting technique carries out Classification and Identification to test sample, and support vector regression sorting technique can be directly to multiclass Failure is classified, and its learning process can be regarded as the process that optimal solution is found in an optimization, and support vector regression has Unknown new samples are identified with more preferable precision, using support vector regression as classification by powerful generalization ability Device carries out Classification and Identification to sample, the general shortcoming of the grader generalization ability that deep learning can be overcome to give tacit consent to；

3rd, compared with the prior art, Fault Diagnosis of Roller Bearings of the invention can improve rolling bearing fault diagnosis Accuracy and validity, to solve the problems, such as that rolling bearing fault diagnosis provide a kind of new effective way, can be widely applied to In the complication system in the fields such as machinery, chemical industry, metallurgy, electric power, aviation.

Described above is only the general introduction of technical solution of the present invention, in order to better understand technological means of the invention, And can be practiced according to the content of specification, below with presently preferred embodiments of the present invention and coordinate accompanying drawing describe in detail as after.

Brief description of the drawings

Fig. 1 is the flow of Fault Diagnosis of Roller Bearings of the present invention based on convolutional neural networks and support vector regression Figure；

Fig. 2 is that Fault Diagnosis of Roller Bearings of the present invention based on convolutional neural networks and support vector regression is corresponding Theory diagram；

Fig. 3 is the original vibration acceleration signal time domain distribution map (time domain unit is s) of rolling bearing health status operating；

Fig. 4 is original vibration acceleration signal time domain distribution map (the time domain unit of rolling bearing inner ring malfunction operating For s)；

Fig. 5 is original vibration acceleration signal time domain distribution map (the time domain list of rolling bearing rolling element malfunction operating Position is for s)；

(time domain unit is the original vibration acceleration signal time domain distribution map of Fig. 6 housing washers malfunction operating s)；

Fig. 7 is back-propagation algorithm flow chart；

Fig. 8 is the model architecture schematic diagram of convolutional neural networks model；

Fig. 9 is training sample classification results figure；

Figure 10 is test sample classification results figure.

Specific embodiment

With reference to the accompanying drawings and examples, specific embodiment of the invention is described in further detail.Hereinafter implement Example is not limited to the scope of the present invention for illustrating the present invention.

In order to overcome the deficiencies in the prior art, rolling bearing of the present invention based on convolutional neural networks and support vector regression Method for diagnosing faults, first using convolutional neural networks come the substantive characteristics of learning training sample data, afterwards using support to Amount returns sorting technique and carries out Classification and Identification to test sample, so that it is determined that rolling bearing fault operating mode classification, to improve rolling The accuracy and validity of bearing failure diagnosis.

Convolutional neural networks (Convolutional Neural Network, abbreviation CNN) possess powerful function representation Ability, possesses the good characteristic from initial data learning data substantive characteristics.Research shows by multilayered nonlinear mapping layer group Into depth network structure it is more more efficient than shallow structure, complicated function represent and complicated classification on have good effect and effect Rate.

The core concept of support vector regression (Support Vector Regression, abbreviation SVR) grader is to pass through Certain Nonlinear Mapping (kernel function), a high-dimensional feature space is mapped to by input vector, constructs optimal separating hyper plane, from And realize Classification and Identification.It solves traditional SVMs and needs to use multiple graders when many classification problems are processed, Cause the shortcoming that computation complexity increases, the training time is long.Support vector regression effectively can be identified to malfunction, in event Barrier diagnosis and failure predication aspect are applied.

Based on the above-mentioned advantage that convolutional neural networks and support vector regression possess, rolling bearing fault diagnosis of the invention Method is integrated the above-mentioned advantage that convolutional neural networks and support vector regression possess, using deep learning and supporting vector Recurrence carries out the classification of rolling bearing fault operating mode, realizes identification and diagnosis to rolling bearing fault, its concrete operations flow As depicted in figs. 1 and 2, comprise the following steps：

Step 1：During rolling bearing rotation work under four kinds of different operating modes, gather every respectively by acceleration transducer The vibration acceleration signal that rolling bearing works in different rotating speeds under kind operating mode, carries out noise suppression preprocessing, and adds operating mode label, Using each vibration acceleration signal data under the various operating modes by pre-processing and after adding operating mode label as training sample； Four kinds of operating modes are respectively normal operation, the operating of bearing inner race failure, the operating of bearing roller failure, bearing outer ring failure fortune Turn.

There is certain difference each other in rolling bearing vibration acceleration signal of rotation work under four kinds of different operating modes Different, Fig. 3 to Fig. 6 respectively illustrates rolling bearing in health status operating, the operating of inner ring failure, the operating of rolling element failure and outer ring Original vibration acceleration signal time-domain diagram (time domain unit is s) under failure Operation Conditions, signal has notable difference, but also not Bearing health status can clearly be separated by time-domain signal figure.Therefore can be added based on vibration of the rolling bearing under different operating modes Rate signal data, are identified to its failure situation.

Step 2：Convolutional neural networks model is set up, convolutional neural networks model is trained using training sample, will In training sample input convolutional neural networks model, using there is supervision, successively training method is successively trained and tuning, is obtained The connection weight and offset parameter of convolutional neural networks model.

The model architecture schematic diagram of convolutional neural networks model as shown in fig. 7, from structure, convolutional neural networks model If being made up of dried layer convolutional layer and pond layer.

The training method of convolutional neural networks is back-propagation algorithm, algorithm flow schematic diagram as shown in figure 8, the original of algorithm Reason is, to the gradient of each weight, full weight renewal to be carried out according to gradient descent algorithm using chain type derived function loss function.

The cost function that solution convolutional neural networks model is used is mean square error loss function, and its formula is：

Wherein,It is k-th target labels value of sample m,It is corresponding k-th network output valve.

Solution makes the minimum parameter of mean square error loss function set up network, is realized by below equation：

Convolutional layer is exactly feature extraction layer.

In convolutional layer, the input of each unit is connected with the regional area of preceding layer, and extracts the local feature.Use The weights of the characteristic pattern of same convolution kernel are identicals, i.e., weights are shared.Local connection and weights are shared and can be greatly reduced The quantity of parameter.

Step 2.1：Convolutional neural networks are by several processes come solution formula (2).The data that the first step will be trained are defeated Enter convolutional layer, carry out convolution algorithm.The input of each hidden layer is the output of last layer, and computing formula is as follows：

s_i=ρ (v_i),with v_i=W_i·s_i-1+b_i. (3)

Wherein, W_iIt is the connection weight between convolutional neural networks adjacent two layers, s is the training data of input, b_iIt is convolution Offset parameter between neutral net adjacent two layers, ρ is activation primitive.

According to above-mentioned activation probability, when given training sample is input into visible node layer, using convolutional neural networks After all nodes of the distribution function excitation hidden layer of model, then the excitation of next hidden layer node is carried out, so as to regain New node layer value.

One convolutional layer can include several different convolution characteristic patterns, therefore the output of this layer can be expressed as preceding layer institute There is adding with formula shows as follows for convolution characteristic pattern：

Wherein, symbol * represents convolution algorithm, and convolution algorithm can be expressed as follows：

Pond layer (polymer layer) is Feature Mapping layer.

Pond layer plays a part of Further Feature Extraction, and it is that feature to being come in and gone out from convolutional layer carries out aggregate statistics, this A little statistical natures not only have much lower dimension, while can also improve result.

The second step of convolutional neural networks solution formula (2) is that the feature exported from convolutional layer is input into pond layer.Institute It is with formula：

Wherein down () represents down-sampling formula,L layers of the multiplying property biasing of i-th node is represented,Represent l layers I-th additivity biasing of node.

Step 2.2：Last layer output to step 2.1 gained convolutional neural networks, successively training side is supervised using having Method is successively trained and tuning, and concrete mode is：

Weight and biasing are calculated using propagated forward.Last layer of the convolutional neural networks model by step 2.1 gained is hidden Output containing layer is successively traveled to output layer as input, the class categories predicted.Letter is lost with chain type derived function Several gradients to each weight, i.e. sensitivity.Gradient calculation formula is：

In convolutional neural networks, the calculation expression of l layers of gradient (sensitivity) is：

Wherein,Represent each element multiplication.

Operating mode label according to training sample determines the actual classification result of training sample, will train the classification of prediction output The actual classification result of result and training sample is compared and obtains error in classification, by error in classification successively back-propagation, so that Realize carrying out tuning, the specific formula that connection weight is updated to the connection weight parameter of each layer of convolutional neural networks model For：

Wherein, η is learning rate.

Successively trained by above-mentioned, until obtaining the output of last layer of hidden layer of convolutional neural networks model.

After carrying out tuning to the connection weight of each layer of convolutional neural networks model, whole convolutional neural networks mould is finally determined The connection weight and offset parameter of type.

Step 3：Using the training sample under various operating modes as the convolutional Neural for determining connection weight and offset parameter The input of network model, deep learning is carried out to training sample, using the convolutional Neural net for determining connection weight and offset parameter Network model carries out validity feature extraction to each training sample under every kind of operating mode respectively, obtains each training under every kind of operating mode The corresponding training sample characteristic information of sample.

According to the characteristic of convolutional neural networks, using the convolution god that connection weight and offset parameter are determined after training, tuning Through network model, acquisition can represent the feature of primary signal essential information, such that it is able to using these substantive characteristics as dividing The input of class identification.

With these training sample features training support vector regression graders for being extracted, support vector regression classification is obtained Device model.

Step 4：Vibration acceleration signal number of the rolling bearing to be measured in rotation work is gathered by acceleration transducer According to, and noise suppression preprocessing is carried out, as test sample.

Step 5：Using test sample as the convolutional neural networks model for determining connection weight and offset parameter input, it is right Test sample carries out deep learning, and test sample is entered using the convolutional neural networks model for determining connection weight and offset parameter Row feature extraction, obtains the feature of test sample.

Similarly, the step is utilized and determines the convolutional neural networks model of optimal connection weight and offset parameter to test specimens Originally feature extraction is carried out, in the vibration acceleration signal number by the rolling bearing to be measured included in the test sample feature that obtains Comprising substantive characteristics matched with the substantive characteristics that the training sample reconstruction signal under various operating modes is embodied it is right to realize The identification of the affiliated fault condition classification of rolling bearing to be measured.

Step 6：Using test feature information as test sample matching characteristic, by each training sample under every kind of operating mode Corresponding training sample characteristic information as matching benchmark, using the support vector regression grader for training to test sample with Training sample is matched, and the operating mode kind judging belonging to the training sample that will the most be matched with test sample is test sample Operating mode classification, so as to obtain the fault diagnosis result of rolling bearing to be measured.

Support vector regression (Support Vector Regression, abbreviation SVR) is based on SVMs (Support Vector Machine, abbreviation SVM) put forward for multi-class classification a kind of method.SVM is in 1963 Year is invented by Vapnik etc. and proposed,, with structural risk minimization principle as theoretical foundation, it maps vector from lower dimensional space for it To in a more higher dimensional space, a maximum separation hyperplane (dimension fewer than higher dimensional space dimension one is set up in higher dimensional space Dimension), sorted data into by optimal hyperlane.Support vector regression is the extension of SVM, and support vector regression is by many classification problems Develop into regression problem, can directly carry out multi-class classification.

The purpose of support vector regression is to find an optimal hyperlane, and the learning strategy of this optimal hyperlane is interval Maximize, i.e., the interval between it enables to supporting vector takes maximum.

Concrete operations that support vector regression sorting technique is matched to test sample with training sample are according to being：

Step 6.1：Support vector regression function is defined as follows：

Wherein, x_iIt is the sample characteristics of input,And α_iIt is Lagrange multiplier, b is biasing, and K () is kernel function.

The present invention selects gaussian radial basis function (RBF) kernel function：

Wherein：σ is the parameter of RBF kernel functions.

The optimal problem of support vector regression is：

Wherein, | | w | | is 2 norms of weight, and C is the regularization factor, ξ_iWithIt is slack variable, ε is the limit of error.

Construct following Lagrangian：

Wherein, μ_iIt is on ξ_iLagrange multiplier.

Formula (14) is zero to the partial derivative of w, b and ξ, is obtained：

Formula (15) is updated in formula (14), and the convex optimization problem of its antithesis is converted into by object function is minimized, Obtain convex optimization aim：

Step 6.2：In four kinds of training samples of operating mode, the corresponding label of every kind of operating mode is y, y ∈ { 0,1,2,3 }, is led to Cross the categorised decision function that support vector regression sorting technique obtains M class problems：

Wherein, α_iWithIt is the Lagrange coefficient in categorised decision function；B is the optimal hyperlane of categorised decision function Position parameter；N is four kinds of sums of the training sample of operating mode；K(x_i, x) represent gaussian radial basis function.

Thus the categorised decision function under four kinds of operating modes is obtained.

Step 6.3：Using test sample feature as the input quantity of the corresponding categorised decision function of four kinds of operating modes, survey is calculated Sample eigen as input quantity support vector regression Decision Classes decision function value, i.e., the operating mode kind judging corresponding to it is The operating mode classification of test sample, obtains the fault diagnosis result of rolling bearing to be measured.

Verified by experimental data, using the rolling bearing based on convolutional neural networks and support vector regression of the invention Method for diagnosing faults carries out fault diagnosis by above-mentioned flow, in 250 training samples and the data qualification of 250 test samples Under, this method can reach 99.6% to the recognition accuracy of training sample, as shown in figure 9, can be reached to the accuracy rate of test sample To 98%, as shown in Figure 10, this nicety of grading disclosure satisfy that practical application request.

In sum, Fault Diagnosis of Roller Bearings of the present invention based on convolutional neural networks and support vector regression, Feature extraction needed for adaptively completing fault diagnosis using convolutional neural networks theory study algorithm, automatic mining goes out hiding Abundant information in given data, has broken away from the dependence to a large amount of signal transacting knowledge and diagnosis engineering experience, saves labor Dynamic cost and time, and there is very big advantage in terms of monitoring, diagnosing ability and generalization ability.Because employ support to Amount returns sorting technique and Classification and Identification is carried out to test sample, and support vector regression sorting technique directly can be entered to multiclass failure Row classification, its learning process can be regarded as the process that optimal solution is found in an optimization, have efficacious prescriptions using what is designed before Method looks for and finds the global minimum of object function, and method is relatively stable and accurate.Compared with the prior art, it is of the invention Fault Diagnosis of Roller Bearings can improve the accuracy and validity of rolling bearing fault diagnosis, to solve rolling bearing event Barrier diagnosis problem provides a kind of new effective way, can be widely applied to answering for the fields such as machinery, chemical industry, metallurgy, electric power, aviation In miscellaneous system.

The above is only the preferred embodiment of the present invention, is not intended to limit the invention, it is noted that for this skill For the those of ordinary skill in art field, on the premise of the technology of the present invention principle is not departed from, can also make it is some improvement and Modification, these are improved and modification also should be regarded as protection scope of the present invention.

Claims (4)

1. a kind of Fault Diagnosis of Roller Bearings, it is characterised in that including step：

Step 1：When rolling bearing rotation work under four kinds of different operating modes, every kind of work is gathered respectively by acceleration transducer The vibration acceleration signal that rolling bearing works in different rotating speeds under condition, carries out noise suppression preprocessing, and adds operating mode label, will be through Each vibration acceleration signal data under the various operating modes crossed after pre-processing and adding operating mode label are used as training sample；

Step 2：Convolutional neural networks model is set up, convolutional neural networks model is trained using training sample, will trained In sample input convolutional neural networks model, using there is supervision, successively training method is successively trained and tuning, obtains convolution Connection weight and offset parameter of neural network model etc.；

Step 3：Using the training sample under various operating modes as the convolutional neural networks for determining connection weight and offset parameter The input of model, deep learning is carried out to training sample, using the convolutional neural networks mould for determining connection weight and offset parameter Type carries out validity feature extraction to each training sample under every kind of operating mode respectively, and with the training sample features training extracted Support vector regression grader, obtains support vector regression sorter model；

Step 4：Vibration acceleration signal data of the rolling bearing to be measured in rotation work are gathered by acceleration transducer, and Noise suppression preprocessing is carried out, as test sample；

Step 5：Using test sample as the input of the convolutional neural networks model for training, depth is carried out to test sample Practise, feature extraction is carried out to test sample using the convolutional neural networks model that connection weight and offset parameter is determined, obtain Test sample characteristic signal；

Step 6：Using test feature information as the matching characteristic of test sample, each training sample under every kind of operating mode is corresponded to Training sample characteristic information as matching benchmark, using the support vector regression grader for training to test sample with training Sample carries out classification and matching, and the operating mode kind judging belonging to the training sample that will the most be matched with test sample is test sample Operating mode classification, so as to obtain the fault diagnosis result of rolling bearing to be measured.

2. Fault Diagnosis of Roller Bearings according to claim 1, it is characterised in that：The volume set up in the step 2 Product neural network model mean square error loss function be：

$E^n=\frac{1}{2}\underset{k=1}{\overset{m}{\Σ}}{(y_k^n-t_k^n)}^2.---\left(1\right)$

Wherein,It is k-th target labels value of sample m,It is corresponding k-th network output valve,

Solution makes the minimum parameter of mean square error loss function set up network, is realized by below equation：

$\hat{f}:\underset{f\⋐F}{\mathrm{argmin}}E\[L(Y,f(S)\].---\left(2\right)$

Step 2.1：Convolutional neural networks by two steps come solution formula (2), the data input convolutional layer that the first step will be trained, Convolution algorithm is carried out, the input of each hidden layer is the output of last layer, and computing formula is as follows：

s_i=ρ (v_i),with v_i=W_i·s_i-1+b_i. (3)

Wherein, W_iIt is the connection weight between convolutional neural networks adjacent two layers, s is the training data of input, b_iIt is convolutional Neural Offset parameter between network adjacent two layers, ρ is activation primitive；

This layer of output can be expressed as all convolution characteristic results of preceding layer plus and, formula shows as follows：

$s_j(u,k_j)=\mathrm{\ρ}\left(\underset{k}{\mathrm{\Σ}}\right(s_{j-1}\left(.,k\right)*W_{j,k_j}\left(.,k\right)+b_j\left(.,k\right)\left)\right(u).---\left(4\right)$

Wherein, symbol * represents convolution algorithm, and convolution algorithm can be expressed as follows：

$(f*g)\left(s\right)=\underset{u=-\mathrm{\∞}}{\overset{\mathrm{\∞}}{\mathrm{\Σ}}}f\left(u\right)g(s-u).---\left(5\right)$

The second step of convolutional neural networks solution formula (2) is that the feature exported from convolutional layer is input into a down-sampling layer, For being polymerized, aggregation formula is down-sampling layer：

$s_i^l=\mathrm{\ρ}\left({\mathrm{\β}}_i^{l}down\right(s_i^{l-1})+b_i^l).---\left(6\right)$

Wherein down () represents down-sampling formula,L layers of the multiplying property biasing of i-th node is represented,Represent l layers i-th The additivity biasing of individual node.

Step 2.2：Last layer output to step 2.1 gained convolutional neural networks, using there is supervision, successively training method is entered Row is successively trained and tuning, and concrete mode is：

Weight and biasing are calculated using propagated forward：By last layer of hidden layer of convolutional neural networks model of step 2.1 gained Output as input successively traveled to output layer, the class categories predicted；With chain type derived function loss function pair The gradient of each weight, gradient calculation formula is：

$\frac{\∂E}{\∂b}=\frac{\∂E}{\∂v}\frac{\∂v}{\∂b}=\mathrm{\δ}.---\left(7\right)$

In convolutional neural networks, the calculation expression of l layers of gradient (sensitivity) is：

Wherein,Represent each element multiplication；

Operating mode label according to training sample determines the actual classification result of training sample, will train the classification results of prediction output It is compared with the actual classification result of training sample and obtains error in classification, by error in classification successively back-propagation, so as to realizes Connection weight parameter to each layer of convolutional neural networks model carries out tuning, and the specific formula that connection weight is updated is：

$W^{l+1}=W^l-\mathrm{\η}\frac{\∂E}{\∂W^l}.---\left(9\right)$

$\frac{\∂E}{\∂W^l}=s^{l-1}{\left({\mathrm{\δ}}^l\right)}^T.---\left(10\right)$

Wherein, η is learning rate；

Successively train, until obtaining the output of last layer of hidden layer of convolutional neural networks model；

After carrying out tuning to the connection weight of each layer of convolutional neural networks model, whole convolutional neural networks model is finally determined Connection weight and offset parameter.

3. Fault Diagnosis of Roller Bearings according to claim 1, it is characterised in that：The step 6 using support to Amount returns sorting technique：

Step 6.1：An optimal hyperlane for support vector regression is found, support vector regression function is defined as follows：

$f\left(x\right)=\underset{i=1}{\overset{n}{\Σ}}({\mathrm{\α}}_i^{*}-{\mathrm{\α}}_i)K(x_i\·x)+b.---\left(11\right)$

Wherein, x_iIt is the sample characteristics of input,And α_iIt is Lagrange multiplier, b is biasing, and K () is kernel function；

From gaussian radial basis function (RBF) kernel function：

$K(x_i\·x)=\exp (-\frac{\left|\right|x_i-x|{|}^2}{{\mathrm{\σ}}^2}).---\left(12\right)$

Wherein：σ is the parameter of RBF kernel functions.

The optimal problem of support vector regression is：

$\begin{array}{c}\underset{w,b}{\min }\frac{1}{2}\left|\right|w|{|}^2+C\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left({\mathrm{\ξ}}_i+{\mathrm{\ξ}}_i^{*}\right)\\ \begin{array}{cc}s.t.& y_i-w\·x_i-b\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i\\ & w\·x_i+b-y_i\≤\mathrm{\ϵ}+{\mathrm{\ξ}}_i^{*}\\ & {\mathrm{\ξ}}_i,{\mathrm{\ξ}}_i^{*}\≥0\end{array}\end{array}.---\left(13\right)$

Wherein, | | w | | is 2 norms of weight, and C is the regularization factor, ξ_iWithIt is slack variable, ε is the limit of error；

Construct following Lagrangian：

$L(w,b,\mathrm{\ξ},\mathrm{\α},\mathrm{\μ})=\frac{1}{2}\left|\right|w|{|}^2+C\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\ξ}}_i-\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i\left(y_i\right(w\·x_i+b)-1+{\mathrm{\ξ}}_i)-\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\μ}}_i{\mathrm{\ξ}}_i.---\left(14\right)$

Wherein, μ_iIt is on ξ_iLagrange multiplier；

Formula (14) is zero to the partial derivative of w, b and ξ, is obtained：

$\left\{\begin{array}{c}w=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i{x}_i\\ \underset{i=1}{\overset{n}{\Σ}}{\mathrm{\α}}_i{y}_i=0\\ C-{\mathrm{\α}}_i-{\mathrm{\μ}}_i=0\end{array}\right..---\left(15\right)$

Formula (15) is updated in formula (14), and the convex optimization problem of its antithesis is converted into by object function is minimized, obtained Convex optimization aim：

$\begin{array}{c}\underset{\mathrm{\α},{\mathrm{\α}}^{*}}{\max }-\frac{1}{2}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\left({\mathrm{\α}}_i^{*}-{\mathrm{\α}}_i\right)\left({\mathrm{\α}}_j^{*}-{\mathrm{\α}}_j\right)K\left(x_i\·x_j\right)+\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}_i\left({\mathrm{\α}}_i^{*}-{\mathrm{\α}}_i\right)-\mathrm{\ϵ}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left({\mathrm{\α}}_i^{*}+{\mathrm{\α}}_i\right)\\ \begin{array}{cc}s.t.& \underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left({\mathrm{\α}}_i^{*}-{\mathrm{\α}}_i\right)=0\\ & 0\≤{\mathrm{\α}}_i,{\mathrm{\α}}_i^{*}\≤C,i=1,2,\mathrm{...},n\end{array}\end{array}---\left(16\right)$

Step 6.2：In four kinds of training samples of operating mode, the corresponding label of every kind of operating mode is y, y ∈ { 0,1,2,3 }, by branch Hold the categorised decision function that vector regression sorting technique obtains M class problems：

$\arg \underset{m=1,2,...,M}{\min }|m-\left(\underset{i=1}{\overset{n}{\Σ}}\right({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*}\left)K\right(x_i,x)+b)|.---\left(17\right)$

Wherein, α_iWithIt is the Lagrange coefficient in categorised decision function；B is the optimal hyperlane position of categorised decision function Coefficient；N is four kinds of sums of the training sample of operating mode；K(x_i, x) represent gaussian radial basis function；

Thus the categorised decision function under four kinds of operating modes is obtained；

Step 6.3：Using test sample feature as the input of the corresponding categorised decision function of four kinds of operating modes, test sample is calculated Feature as input quantity support vector regression Decision Classes decision function value, i.e., operating mode kind judging corresponding to it is test specimens This operating mode classification, obtains the fault diagnosis result of rolling bearing to be measured.

4. the Fault Diagnosis of Roller Bearings according to claim any one of 1-3, it is characterised in that：Four kinds of operating modes Respectively run well, bearing inner race failure is operated, bearing roller failure is operated, the operating of bearing outer ring failure.