CN104616033A - Fault diagnosis method for rolling bearing based on deep learning and SVM (Support Vector Machine) - Google Patents

Abstract

The invention provides a fault diagnosis method for a rolling bearing based on a deep learning and SVM (Support Vector Machine). The method comprises using a manure learning algorithm in a deep belief network theory to complete a characteristic extraction task needed by fault diagnosis; automatically extracting the substantive characteristics of data input independent of manual selection from simple to complicate, from low to high, and automatically digging abundant information concealed in known data; in addition, classifying and identifying a test sample by adopting an SVM classification method, seeking and finding a global minimum of a target function through an effective method previously designed, so as to solve the problem that a deep belief network may be trapped into a locally optimal solution. According to the fault diagnosis method for the rolling bearing based on the deep learning and SVM provided by the invention, the accuracy and effectiveness of the fault diagnosis method for a rolling bearing can be improved, and a new effective way can be provided to solve the accuracy and effectiveness of the fault diagnosis method, therefore the fault diagnosis method can be extensively applied complex systems in chemistry, metallurgy, electric power, aviation fields and the like.

Description

Based on degree of depth study and the Fault Diagnosis of Roller Bearings of support vector machine

Technical field

The invention belongs to mechanical fault diagnosis and Artificial technical field of intelligence, particularly relate to a kind of based on degree of depth study and the Fault Diagnosis of Roller Bearings of support vector machine.

Background technology

Rolling bearing is one of of paramount importance mechanical component in rotating machinery, is widely used in each important departments such as chemical industry, metallurgy, electric power, aviation, and it is also one of the most flimsy element simultaneously.The quality of bearing performance and operating mode directly has influence on the performance of axle associated therewith and the gear be arranged in rotating shaft and even entire machine equipment, and its defect can cause equipment to produce abnormal vibrations and noise, even causes device damage.Therefore, rolling bearing fault is diagnosed, especially for the analysis of incipient failure morning, avoid the generation of accident, particularly important in production reality.

Mechanical fault feature extracting method conventional at present mainly contains fast Fourier and becomes (Fast Fouriertransform, be called for short FFT), wavelet transformation and empirical mode decomposition (Empirical mode decomposition is called for short EMD), artificial intelligence etc.FFT method cannot take into account the overall picture of signal in time domain and frequency domain and Localization Problems simultaneously.During wavelet transformation, wavelet basis is different, and decomposition result is different, the more difficult selection of wavelet basis.Signal decomposition can be multiple IMF (Intrinsicmode function by EMD method, intrinsic mode function) component, the time-frequency distributions that Hilbert conversion can obtain signal is done to all IMF components, but in theory also there are some problems, as the mode in EMD method is obscured, owes envelope, crossed the problem such as envelope, end effect, be all among research.Based in the method for diagnosing faults of artificial intelligence, mainly utilize artificial neural network at present, by constantly learning and carry out system the feedback of information, the classification of complete Duplicate diagnostic target; But its shortcoming explanatory difference that is reasoning process, and when when diagnostic sample incomplete (data have disappearance), neural network can not carry out effective reasoning work, and the early sign of fault cannot be utilized to carry out corresponding diagnosis to bearing.

Summary of the invention

For the above-mentioned problems in the prior art, the invention provides a kind of based on degree of depth study and the Fault Diagnosis of Roller Bearings of support vector machine, first it adopt degree of depth belief network to carry out the essential characteristic of learning training sample data, support vector machine classification method is adopted to carry out Classification and Identification to test sample book afterwards, thus determine rolling bearing fault operating mode classification, realize the diagnosis to rolling bearing fault classification, to improve accuracy and the validity of rolling bearing fault diagnosis.

For achieving the above object, present invention employs following technological means:

Based on degree of depth study and the Fault Diagnosis of Roller Bearings of support vector machine, comprise the steps:

1) during rolling bearing rotation work under four kinds of different operating modes, rolling bearing is gathered under often kind of operating mode respectively at the vibration acceleration signal of different rotating speeds work by acceleration transducer, carry out noise suppression preprocessing, and add operating mode label, using through pre-service and each vibration acceleration signal data of adding under the various operating modes after operating mode label as training sample; Described four kinds of operating modes are respectively normal operation, the running of bearing inner race fault, the running of bearing roller fault, the running of bearing outer ring fault;

2) degree of depth belief network model is set up, adopt training sample to the training of degree of depth belief network model, training sample is inputted in degree of depth belief network model, adopt non-supervisory greed successively training method successively train and tuning, obtain connection weights and the offset parameter of degree of depth belief network model;

3) using the training sample under various operating mode as the input determining the degree of depth belief network model connecting weights and offset parameter, degree of depth study is carried out to training sample, adopt the degree of depth belief network model determining to connect weights and offset parameter to be reconstructed each training sample under often kind of operating mode respectively, obtain the training sample reconstruction signal that each training sample under often kind of operating mode is corresponding;

4) gather the vibration acceleration signal data of rolling bearing to be measured when rotation work by acceleration transducer, and carry out noise suppression preprocessing, as test sample book;

5) using test sample book as the input determining the degree of depth belief network model connecting weights and offset parameter, degree of depth study is carried out to test sample book, adopt the degree of depth belief network model determining to connect weights and offset parameter to be reconstructed test sample book, obtain test sample book reconstruction signal;

6) using the matching characteristic of test sample book reconstruction signal as test sample book, using training sample reconstruction signal corresponding for each training sample under often kind of operating mode as coupling benchmark, support vector machine classification method is adopted to mate with training sample test sample book, by the operating mode classification that the operating mode kind judging belonging to the training sample mated the most with test sample book is test sample book, thus obtain the fault diagnosis result of rolling bearing to be measured.

Above-mentioned based in the Fault Diagnosis of Roller Bearings of degree of depth study and support vector machine, specifically, described step 2) in the joint distribution function of degree of depth belief network model set up be:

$E(v,h|\mathrm{\θ})=-\underset{i=1}{\overset{I}{\mathrm{\Σ}}}{a}_i{v}_i-\underset{j=1}{\overset{J}{\mathrm{\Σ}}}{b}_j{h}_j-\underset{i=1}{\overset{I}{\mathrm{\Σ}}}\underset{j=1}{\overset{J}{\mathrm{\Σ}}}{w}_{\mathrm{ij}}{v}_i{h}_j;---\left(1\right)$

Wherein, θ=(w _ij, a _i, b _j) be degree of depth belief network model parameter, w _ijrepresent visible layer i-th node v of degree of depth belief network _iwith a hidden layer jth node h _jbetween connection weights, a _iand b _jrepresent visible layer i-th node v respectively _ioffset parameter and a hidden layer jth node h _joffset parameter;

Adopt non-supervisory greed successively training method degree of depth belief network model is successively trained and tuning, concrete mode is:

21) mode of successively training is adopted to train the limited Boltzmann machine in each layer of degree of depth belief network model, the visible layer that the hidden layer of the limited Boltzmann machine of low one deck exports as the limited Boltzmann machine of last layer inputs, until obtain the output of the last one deck hidden layer of degree of depth belief network model; Be specially:

The joint probability distribution Probability p (v, h| θ) of visible layer node and hidden layer node is:

p(v,h|θ)＝e ^-E(v,h|θ)/Z(θ)； (2)

Wherein, $Z\left(\mathrm{\θ}\right)=\underset{v}{\mathrm{\Σ}}\underset{h}{\mathrm{\Σ}}{e}^{-E(v,h|\mathrm{\θ})}$ For partition function;

When the state of given visible layer node, a hidden layer jth node h _jactivation probability be:

$p(h_j=1|v,\mathrm{\θ})=\mathrm{\σ}(b_j+\underset{i=1}{\overset{I}{\mathrm{\Σ}}}{v}_i{w}_{\mathrm{ji}});---\left(3\right)$

When the state of given hidden layer node, visible layer i-th node v _iactivation probability be:

$p(v_i=1|h,\mathrm{\θ})=\mathrm{\σ}(a_j+\underset{i=1}{\overset{I}{\mathrm{\Σ}}}{h}_i{w}_{\mathrm{ji}});---\left(4\right)$

Wherein, σ (x)=1/ (1+e ^-x) be sigmoid function;

According to above-mentioned activation probability, when given training sample is inputed to visible layer node, after adopting all nodes of joint distribution function excitation hidden layer of degree of depth belief network model, then carry out the excitation of next hidden layer node, thus regain visible layer node; Then, adopt condition distribution sdpecific dispersion algorithm being calculated to visible layer data, obtain hidden layer data, again with the condition distribution calculating gained hidden layer data, calculate visible layer data, reconstruct is realized to visible layer data, utilizes gradient descent method, to adjustment and the renewal of degree of depth belief network model parameter, the renewal difference connecting weights and offset parameter is respectively:

Δw _ij＝ε(<v _ih _j> _data-<v _ih _j> _recon)； (6)

Δa _i＝ε(<v _i> _data-<v _i> _recon)； (7)

Δb _j＝ε(<h _j> _data-<h _j> _recon)； (8)

Wherein, Δ w _ijrepresent visible layer i-th node v of degree of depth belief network _iwith a hidden layer jth node h _jbetween connection weight w _ijcarry out the renewal difference upgraded, Δ a _iwith Δ b _jrepresent visible layer i-th node v respectively _ioffset parameter a _icarry out the renewal difference that upgrades and a hidden layer jth node h _joffset parameter b _jcarry out the renewal difference upgraded; ε is the learning rate of training, <> _datarepresent the mathematical expectation on the distribution that training dataset defines, <> _reconrepresent the mathematical expectation in the distribution that the degree of depth belief network model after reconstruct exports;

Successively train through above-mentioned, until obtain the output of the last one deck hidden layer of degree of depth belief network model;

22) output of the last one deck hidden layer of degree of depth belief network model of step 21 gained is carried out to the training of counterpropagation network, and the error in classification successively back-propagation of the actual classification result of the classification results that training prediction is exported and training sample, tuning is carried out to the connection weights of each layer of degree of depth belief network model; Be specially:

The training of counterpropagation network is divided into propagated forward and back-propagating two processes; In propagated forward process, successively output layer is propagated into as input by the output of the last one deck hidden layer of degree of depth belief network model of step 21 gained, obtain the class categories predicted, and the actual classification result of operating mode label determination training sample according to training sample, again the actual classification result of the classification results of prediction and training sample is compared and obtain error in classification, this error in classification is successively returned backward, thus the parameter of tuning degree of depth training network; In back-propagating process, need the value of the sensitivity δ calculating every one deck, sensitivity δ by top-down transmission with the connection weights of Corrected Depth belief network model;

For output layer, suppose that the actual output of i-th node is o _i, desired output is d _i, so the calculation expression of the sensitivity δ of i-th node is:

δ _i＝o _i(1-o _i)(d _i-o _i)； (9)

For m hidden layer, the calculation expression of the sensitivity of i-th node is:

${\mathrm{\&delta;}}_i^m=y_i^m(1-y_i^m)\underset{j}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}^m{\mathrm{\&delta;}}_j^{m+1};---\left(10\right)$

Wherein, represent the sensitivity of m hidden layer i-th node, represent the sensitivity of m+1 hidden layer i-th node, represent the output of m hidden layer i-th node, represent m hidden layer i-th node and a m+1 hidden layer jth internodal weights that are connected;

After obtaining the sensitivity δ of each node in each hidden layer thus, carry out renewal tuning by the connection weights of following formula to degree of depth belief network model:

$w_{\mathrm{ij}}^m\&LeftArrow;w_{\mathrm{ij}}^m+{\mathrm{\&epsiv;}}_{\mathrm{fine}-\mathrm{tuning}}\&times;y_i^m{\mathrm{\&delta;}}_j^{m+1};---\left(11\right)$

$b_j^m\&LeftArrow;b_j^m+{\mathrm{\&epsiv;}}_{\mathrm{fine}-\mathrm{tuning}}\&times;{\mathrm{\&delta;}}_j^{m+1};---\left(12\right)$

Wherein, ε _fine-tuningrepresent tuning learning rate, represent the offset parameter of a m hidden layer jth node;

After tuning is carried out to the connection weights of each layer of degree of depth belief network model, finally determine connection weights and the offset parameter of entire depth belief network model.

Above-mentioned based in the Fault Diagnosis of Roller Bearings of degree of depth study and support vector machine, as a kind of preferred version, described step 6) the employing support vector machine classification method concrete mode of mating with training sample test sample book is:

61) in the training sample of four kinds of operating modes, regard positive class as wherein kth class training sample, { 1,2,3,4} is regarded other 3 class training sample as negative class, is obtained the categorised decision function f of kth class by two class support vector machines sorting techniques k ∈ _k(x):

$f_k\left(x\right)=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_n^k{y}_{n}K(x,x_n)+b_k;$

Wherein, for kth class categorised decision function f _kin (x) n-th training sample reconstruction signal x _ncorresponding Lagrange coefficient; b _kfor kth class categorised decision function f _kthe optimal hyperlane position parameter of (x); y _nrepresent the key words sorting that the n-th training sample is corresponding, the y when the n-th training sample belongs to positive class _n=1, the y when the n-th training sample belongs to negative class _n=-1; N ∈ 1,2 ..., N}, N are the sum of the training sample of four kinds of operating modes; K (x, x _n) presentation class decision function f _kthe training sample reconstruction signal x of input quantity x relative to n-th of (x) _ngaussian radial basis function kernel function;

Obtain the categorised decision function corresponding to each operating mode in four kinds of operating modes thus;

62) input quantity using test sample book reconstruction signal as categorised decision function corresponding to four kinds of operating modes, calculate four the categorised decision functional values of test sample book reconstruction signal as input quantity, with the operating mode classification that the operating mode kind judging corresponding to wherein maximum categorised decision functional value is test sample book, obtain the fault diagnosis result of rolling bearing to be measured.

Compared to prior art, the present invention has following beneficial effect:

1, the present invention is based on the Fault Diagnosis of Roller Bearings of degree of depth study and support vector machine, ripe learning algorithm is utilized in degree of depth belief network theory to complete feature extraction tasks needed for fault diagnosis, artificial selection can not be relied on from simple to complex, by rudimentary to the senior essential characteristic automatically extracting input data, a large amount of manpowers can be saved, and energy automatic mining goes out to be hidden in the abundant information in given data, be particularly useful for noisy, uncertain, dynamic system.

2, the present invention is based in the Fault Diagnosis of Roller Bearings of degree of depth study and support vector machine, have employed support vector machine classification method and Classification and Identification is carried out to test sample book, learning process in support vector machine classification method can be regarded as one and optimize the process finding optimum solution, therefore the effective ways designed before can adopting go the global minimum finding and find objective function, thus solve the problem that degree of depth belief network may be absorbed in locally optimal solution.

3, compared with the prior art, Fault Diagnosis of Roller Bearings of the present invention can improve accuracy and the validity of rolling bearing fault diagnosis, thering is provided a kind of new effective way for solving rolling bearing fault diagnosis problem, can be widely used in the complication system in the fields such as chemical industry, metallurgy, electric power, aviation.

Accompanying drawing explanation

Fig. 1 is the FB(flow block) of the Fault Diagnosis of Roller Bearings that the present invention is based on degree of depth study and support vector machine.

Fig. 2 is the original vibration acceleration signal time domain distribution example figure (time domain unit is ms) of rolling bearing inner ring fault running.

Fig. 3 is the original vibration acceleration signal time domain distribution example figure (time domain unit is ms) of rolling bearing rolling body fault running.

Fig. 4 is the model architecture schematic diagram of degree of depth belief network model.

Fig. 5 is the model schematic of limited Boltzmann machine.

Fig. 6 is the position relationship schematic diagram of SVM standardization optimal separating hyper plane.

Embodiment

In order to overcome the deficiencies in the prior art, the present invention is based on the Fault Diagnosis of Roller Bearings of degree of depth study and support vector machine, first degree of depth belief network is adopted to carry out the essential characteristic of learning training sample data, support vector machine classification method is adopted to carry out Classification and Identification to test sample book afterwards, thus determine rolling bearing fault operating mode classification, to improve accuracy and the validity of rolling bearing fault diagnosis.

Degree of depth belief network (Deep BeliefNetwork is called for short DBN) has powerful function representation ability, presents the good characteristic from a few sample learning data essential characteristic.Research shows that the degree of depth network structure be made up of multilayered nonlinear mapping layer is more more effective than shallow structure, represents and complicated classification has good effect and efficiency at complicated function.Degree of depth belief network has been set forth many hidden layers neural network and has been had excellent feature learning ability, its study to feature have more essential portraying to data, thus contribute to classification and visual.

Support vector machine (Support Vector Machine, being called for short SVM) core concept of sorter is, by certain Nonlinear Mapping (kernel function), input vector is mapped to a high-dimensional feature space, and construct optimal separating hyper plane, thus realize Classification and Identification.In solution small sample, non-linear and high dimensional pattern identification, there is unique advantage, study can be limited well again, be particularly suitable for the data processing of small sample set, therefore, it is possible to be applied in fault diagnosis and failure prediction.

Based on the above-mentioned advantage that degree of depth belief network and support vector machine possess, the above-mentioned advantage that degree of depth belief network and support vector machine possess by Fault Diagnosis of Roller Bearings of the present invention is integrated, degree of depth study and support vector machine is utilized to carry out the classification of rolling bearing fault operating mode, realize the identification to rolling bearing fault and diagnosis, its concrete operations flow process as shown in Figure 1, comprises the steps:

1) during rolling bearing rotation work under four kinds of different operating modes, rolling bearing is gathered under often kind of operating mode respectively at the vibration acceleration signal of different rotating speeds work by acceleration transducer, carry out noise suppression preprocessing, and add operating mode label, using through pre-service and each vibration acceleration signal data of adding under the various operating modes after operating mode label as training sample; Described four kinds of operating modes are respectively normal operation, the running of bearing inner race fault, the running of bearing roller fault, the running of bearing outer ring fault.

There is certain difference each other in the vibration acceleration signal of rolling bearing rotation work under four kinds of different operating modes, such as, Fig. 2 and Fig. 3 respectively illustrates the original vibration acceleration signal time domain distribution plan (time domain unit be ms) of rolling bearing under the running of inner ring fault, rolling body fault Operation Conditions, and its signal difference is comparatively obvious.Therefore based on the vibration acceleration signal data of rolling bearing under different operating mode, its failure condition can be identified.

2) degree of depth belief network model is set up, adopt training sample to the training of degree of depth belief network model, training sample is inputted in degree of depth belief network model, adopt non-supervisory greed successively training method successively train and tuning, obtain connection weights and the offset parameter of degree of depth belief network model.

The model architecture schematic diagram of degree of depth belief network model as shown in Figure 4, from structure, degree of depth belief network model is by the unsupervised limited Boltzmann machine of multilayer (Restricted Boltzmann machine, RBM) network and one deck have backpropagation (back-propagation, the BP) network of supervision to form.

The training package of degree of depth training network model is containing " pre-training (pre-training) " and " tuning (fine-tuning) " 2 steps.Pre-training stage degree of depth training network adopts the mode that successively (layerwise) trains to train the RBM in each layer, and the visible layer that the hidden layer of low one deck RBM exports as the RBM of last layer inputs.The fine-tuning stage adopts the BP network of supervised learning mode to last one deck to train, and reality is exported and expect the error successively back-propagation exported, and carries out tuning to the weights of entire depth training network.In fact the training process of RBM network can regard the initialization to deep layer BP network weight as, makes degree of depth training network overcome the training time length that BP network causes because of random initializtion weighting parameter and the shortcoming being easily absorbed in locally optimal solution.

1. limited Boltzmann machine.

Limited Boltzmann machine is a kind of typically based on model (the energy-based model of energy, EBM), it is made up of a visible layer (visible layer) and a hidden layer (hidden layer), and its model schematic as shown in Figure 5.Wherein, v and h represents visible layer and hidden layer respectively, W represent two-layer between connection weights.For visible layer and hidden layer, its annexation is that interlayer neuron connects entirely, and does not have neuron to be connected in layer.

Suppose that visible layer and hidden layer are two-valued variable, the neuron number of visible layer and hidden layer is respectively I and J, v _iand h _jrepresent that i-th visible layer god is by the state of unit with a jth hidden layer neuron respectively.For one group specific (v, h), the joint distribution function of degree of depth belief network model is:

$E(v,h|\mathrm{\&theta;})=-\underset{i=1}{\overset{I}{\mathrm{\&Sigma;}}}{a}_i{v}_i-\underset{j=1}{\overset{J}{\mathrm{\&Sigma;}}}{b}_j{h}_j-\underset{i=1}{\overset{I}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{J}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ij}}{v}_i{h}_j;---\left(1\right)$

Wherein, θ=(w _ij, a _i, b _j) be degree of depth belief network model parameter, w _ijrepresent visible layer i-th node v of degree of depth belief network _iwith a hidden layer jth node h _jbetween connection weights, a _iand b _jrepresent visible layer i-th node v respectively _ioffset parameter and a hidden layer jth node h _joffset parameter;

The joint distribution function of degree of depth belief network model, can obtain the joint probability distribution of (v, h):

p(v,h|θ)＝e ^-E(v,h|θ)/Z(θ)； (2)

Wherein, for partition function.In practical problems, we are it is of concern that the distribution about observational variable (likelihood function) the i.e. p (v| θ) that defines of RBM, and it is the marginal distribution of joint probability p (v, h| θ).

Due to the special construction (in layer, neuron is without connection) of RBM, when the state of given visible layer node, be separate between the state of activation of each hidden layer node.Now, a hidden layer jth node h _jactivation probability be:

$p(h_j=1|v,\mathrm{\&theta;})=\mathrm{\&sigma;}(b_j+\underset{i=1}{\overset{I}{\mathrm{\&Sigma;}}}{v}_i{w}_{\mathrm{ji}});---\left(3\right)$

Wherein, σ (x)=1/ (1+e ^-x) be sigmoid function.Similar, when the state of given hidden layer node, visible layer i-th node v _iactivation probability be:

$p(v_i=1|h,\mathrm{\&theta;})=\mathrm{\&sigma;}(a_j+\underset{i=1}{\overset{I}{\mathrm{\&Sigma;}}}{h}_i{w}_{\mathrm{ji}});---\left(4\right)$

RBM adopts the mode of iteration to train, and the target of training is to learn out parameter θ=(w _ij, a _i, b _j) value, with the training data that matching is given.Parameter θ can obtain by asking the very big log-likelihood function on training sample set (setting the total quantity of training sample as T), that is:

${\mathrm{\&theta;}}^{*}=\arg \underset{\mathrm{\&theta;}}{\max }L\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}\ln p\left(v^{\left(t\right)}\right|\mathrm{\&theta;});---\left(5\right)$

According to above-mentioned activation probability, when training sample being inputed to visible layer node, after adopting all nodes of joint distribution function excitation hidden layer of degree of depth belief network model, then carrying out the excitation of next hidden layer node, thus regaining visible layer node; Then, adopt condition distribution sdpecific dispersion algorithm being calculated to visible layer data, obtain hidden layer data, again with the condition distribution calculating gained hidden layer data, calculate visible layer data, reconstruct is realized to visible layer data, utilizes gradient descent method, to adjustment and the renewal of degree of depth belief network model parameter.Adopt Hinton to propose to sdpecific dispersion (contrastivedivergence, CD) algorithm, the renewal difference that can connect weights and offset parameter is respectively as follows:

Δw _ij＝ε(<v _ih _j> _data-<v _ih _j> _recon) (6)；

Δa _i＝ε(<v _i> _data-<v _i> _recon) (7)；

Δb _j＝ε(<h _j> _data-<h _j> _recon) (8)；

Wherein, Δ w _ijrepresent visible layer i-th node v of degree of depth belief network _iwith a hidden layer jth node h _jbetween connection weight w _ijcarry out the renewal difference upgraded, Δ a _iwith Δ b _jrepresent visible layer i-th node v respectively _ioffset parameter a _icarry out the renewal difference that upgrades and a hidden layer jth node h _joffset parameter b _jcarry out the renewal difference upgraded; ε is the learning rate of training, <> _datarepresent the mathematical expectation on the distribution that training dataset defines, <> _reconrepresent the mathematical expectation in the distribution that the degree of depth belief network model after reconstruct exports.

Successively train through above-mentioned, until obtain the output of the last one deck hidden layer of degree of depth belief network model.

2. the output of the last one deck hidden layer of degree of depth belief network model is carried out to the training of counterpropagation network, and the error in classification successively back-propagation of the actual classification result of the classification results that training prediction is exported and training sample, tuning is carried out to the connection weights of each layer of degree of depth belief network model.

Counterpropagation network is a kind of sorter having supervision, classifies to the proper vector that RBM is obtained by pre-training, and plays the effect of tuning entire depth training network parameter.The training of counterpropagation network is divided into propagated forward and back-propagating two processes; In propagated forward process, successively output layer is propagated into as input feature value by the output of RBM, obtain the class categories predicted, and the actual classification result of operating mode label determination training sample according to training sample, again the actual classification result of the classification results of prediction and training sample is compared and obtain error in classification, this error in classification is successively returned backward, thus the parameter of tuning degree of depth training network; In back-propagating process, need the value of the sensitivity δ calculating every one deck, sensitivity δ by top-down transmission with the connection weights of Corrected Depth belief network model.

For output layer, suppose that the actual output of i-th node is o _i, desired output is d _i, so the calculation expression of the sensitivity δ of i-th node is:

δ _i＝o _i(1-o _i)(d _i-o _i)； (9)

For m hidden layer, the calculation expression of the sensitivity of i-th node is:

${\mathrm{\&delta;}}_i^m=y_i^m(1-y_i^m)\underset{j}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}^m{\mathrm{\&delta;}}_j^{m+1};---\left(10\right)$

Wherein, represent the sensitivity of m hidden layer i-th node, represent the sensitivity of m+1 hidden layer i-th node, represent the output of m layer i-th node, represent i-th node and lower one deck jth internodal weights of m layer;

After obtaining the sensitivity δ of each node in each hidden layer thus, carry out renewal tuning by the connection weights of following formula to degree of depth belief network model:

$w_{\mathrm{ij}}^m\&LeftArrow;w_{\mathrm{ij}}^m+{\mathrm{\&epsiv;}}_{\mathrm{fine}-\mathrm{tuning}}\&times;y_i^m{\mathrm{\&delta;}}_j^{m+1};---\left(11\right)$

$b_j^m\&LeftArrow;b_j^m+{\mathrm{\&epsiv;}}_{\mathrm{fine}-\mathrm{tuning}}\&times;{\mathrm{\&delta;}}_j^{m+1};---\left(12\right)$

Wherein, ε _fine-tuningrepresent tuning learning rate, represent a m layer jth node;

After tuning is carried out to the connection weights of each layer of degree of depth belief network model, finally determine connection weights and the offset parameter of entire depth belief network model.

3) using the training sample under various operating mode as the input determining the degree of depth belief network model connecting weights and offset parameter, degree of depth study is carried out to training sample, adopt the degree of depth belief network model determining to connect weights and offset parameter to be reconstructed each training sample under often kind of operating mode respectively, obtain the training sample reconstruction signal that each training sample under often kind of operating mode is corresponding.

According to the characteristic of degree of depth belief network, the degree of depth belief network model connecting weights and offset parameter is determined after utilizing training, tuning, the former data sample with lower error can be reconstructed, the data sample reconstruction signal that such reconstruct obtains can embody and feature the essential characteristic of former data sample, from the fixed reference feature that these essential characteristic can be utilized as Classification and Identification.Therefore, this step utilizes the degree of depth belief network model determining to connect weights and offset parameter to be reconstructed each training sample under often kind of operating mode, thus utilizes training sample reconstruction signal corresponding to each training sample under often kind of operating mode test sample book to be carried out to the coupling benchmark of Classification and Identification as the later stage.

4) gather the vibration acceleration signal data of rolling bearing to be measured when rotation work by acceleration transducer, and carry out noise suppression preprocessing, as test sample book.

5) using test sample book as the input determining the degree of depth belief network model connecting weights and offset parameter, degree of depth study is carried out to test sample book, adopt the degree of depth belief network model determining to connect weights and offset parameter to be reconstructed test sample book, obtain test sample book reconstruction signal.

In like manner, this step utilizes the degree of depth belief network model determining to connect weights and offset parameter to be reconstructed test sample book, the essential characteristic comprised in the vibration acceleration signal number of rolling bearing to be measured is portrayed by the test sample book reconstruction signal obtained, for mating with the essential characteristic that the training sample reconstruction signal under various operating mode embodies, to realize the identification to fault condition classification belonging to rolling bearing to be measured.

6) using the matching characteristic of test sample book reconstruction signal as test sample book, using training sample reconstruction signal corresponding for each training sample under often kind of operating mode as coupling benchmark, support vector machine classification method is adopted to mate with training sample test sample book, by the operating mode classification that the operating mode kind judging belonging to the training sample mated the most with test sample book is test sample book, thus obtain the fault diagnosis result of rolling bearing to be measured.

Support vector machine (SupportVector Machines, be called for short SVM) in 1963 in AT & TBell laboratory by propositions such as Vapnik, it be theoretical with the VC dimension in statistics and Structural risk minization principle for theoretical foundation, between the complexity (i.e. the study precision of specific training sample) and learning ability (i.e. the correct ability identifying arbitrary sample) of model, optimal compromise is sought, to obtain best Generalization Ability according to limited sample information.SVM by DUAL PROBLEMS OF VECTOR MAPPING in the space of a more higher-dimension, a maximum separation lineoid is set up in higher dimensional space, and two lineoid parallel to each other are set up on the both sides that can separate data lineoid, separating hyperplane makes the distance of two parallel lineoid maximize, its distance is larger, and the error of classification results is less.

Fig. 6 is the position relationship schematic diagram of the standardization optimal hyperlane in two-dimentional two class situations, and H is separating hyperplane H1, H2 is two lineoid parallel to each other, and H1, H2 are class interval d=2/||w||.For ensureing the linearization of data, need data-mapping to kernel function space; Meanwhile, for effectively two classes separately, ensureing that two classes are correctly separately made class interval maximum, objective function is namely had to be:

$\underset{w,b}{\min }\frac{1}{2}{\left|\right|w\left|\right|}^2+C\left(\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&epsiv;}}_i\right)=\frac{1}{2}(w\&CenterDot;w)+C\left(\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&epsiv;}}_i\right);---\left(13\right)$

Have lineoid H1, H2:

Formula (14) can be equivalent to:

Wherein: represent w with between inner product, table is mapped to kernel function space xi; B and C is constant; ε _i> 0 is slack variable, represent training sample mistake point degree, its value larger expression mistake point sample is more.

Lagrange multiplier method is used to formula (13) and (15), obtains:

Wherein: α _i> 0, β _i> 0 is Lagrange coefficient, and L (w, b, α) is Lagrangian function.

Formula (16) is to w, ε _ibe zero with the partial derivative of b, obtain:

Wushu (17) substitutes into formula (16), and the solution of optimal hyperlane is equivalent to the solution of following dual problem.

$\begin{array}{c}\max Q\left(\mathrm{\&alpha;}\right)=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i-\frac{1}{2}\underset{i,j=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j{y}_i{y}_{j}K(x_i\&CenterDot;x_j)\\ s.t.\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{y}_i{\mathrm{\&alpha;}}_i=0,C{>\mathrm{\&alpha;}}_i>0\end{array};---\left(18\right)$

Wherein:

Use Lagrange multiplier method, obtaining solution is:

The classifying rules function being obtained optimal classification surface by formula (19) is:

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

$K(x\&CenterDot;x_j)=\exp (-\frac{{\left|\right|x-x_i\left|\right|}^2}{{\mathrm{\&sigma;}}^2});---\left(21\right)$

Wherein: σ is the parameter of RBF kernel function.

SVM algorithm is the optimal classification surface sought based on statistics between data, by nonlinear data has been mapped to kernel function space, has made its linearization, and then simplifies computation complexity, had good classifying quality.

Step 6) in, the concrete mode classification of the support vector machine classification method that can use has a lot, such as one-against-one (OVO-SVM), one-to-many classification (also referred to as more than a pair classification, OVR-SVM), directed acyclic graph classification (DAG-SVMS), decision tree classification, error correcting output codes classification etc.But consider in the inventive method and only relate to normal operation, bearing inner race fault operates, bearing roller fault operates, bearing outer ring fault operates the fault condition Classification and Identification of these four kinds of operating modes, identification classification is few, consider the factor of recognition efficiency and accuracy, adopt the svm classifier method of one-to-many comparatively applicable, because the SVM classifier decision function of computing only has four (often kinds fault condition classification corresponding one) to adopt one-to-many sorting technique to need to set up and carry out to identify, and the vibration acceleration signal data of rolling bearing under these four kinds of different operating modes are after degree of depth belief network model reconstruction, the difference of its essential characteristic has been enough to identify, can ensure preferably to identify accuracy.

In the methods of the invention, step 6) preferably adopt the concrete mode that one-to-many support vector machine classification method mates with training sample test sample book to be:

61) in the training sample of four kinds of operating modes, regard positive class as wherein kth class training sample, { 1,2,3,4} is regarded other 3 class training sample as negative class, is obtained the categorised decision function f of kth class by two class support vector machines sorting techniques k ∈ _k(x):

$f_k\left(x\right)=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_n^k{y}_{n}K(x,x_n)+b_k;$

Wherein, for kth class categorised decision function f _kin (x) n-th training sample reconstruction signal x _ncorresponding Lagrange coefficient; b _kfor kth class categorised decision function f _kthe optimal hyperlane position parameter of (x); y _nrepresent the key words sorting that the n-th training sample is corresponding, the y when the n-th training sample belongs to positive class _n=1, the y when the n-th training sample belongs to negative class _n=-1; N ∈ 1,2 ..., N}, N are the sum of the training sample of four kinds of operating modes; K (x, x _n) presentation class decision function f _kthe training sample reconstruction signal x of input quantity x relative to n-th of (x) _ngaussian radial basis function kernel function;

Obtain the categorised decision function corresponding to each operating mode in four kinds of operating modes thus;

62) input quantity using test sample book reconstruction signal as categorised decision function corresponding to four kinds of operating modes, calculate four the categorised decision functional values of test sample book reconstruction signal as input quantity, with the operating mode classification that the operating mode kind judging corresponding to wherein maximum categorised decision functional value is test sample book, obtain the fault diagnosis result of rolling bearing to be measured.

Data verification by experiment, the Fault Diagnosis of Roller Bearings based on degree of depth study and support vector machine of the present invention is adopted to carry out fault diagnosis by above-mentioned flow process, under the condition of 600 training samples (often kind of operating mode 150 training samples), 800 rolling bearing fault diagnosis identification is carried out in random sampling, its recognition accuracy reaches 94.3% (in industry, mean failure rate recognition accuracy is only about 83%), can meet practical application request completely.

In sum, the present invention is based on the Fault Diagnosis of Roller Bearings of degree of depth study and support vector machine, ripe learning algorithm is utilized in degree of depth belief network theory to complete feature extraction tasks needed for fault diagnosis, artificial selection can not be relied on from simple to complex, by rudimentary to the senior essential characteristic automatically extracting input data, a large amount of manpowers can be saved, and energy automatic mining goes out to be hidden in the abundant information in given data, be particularly useful for noisy, uncertain, dynamic system; In addition, owing to have employed support vector machine classification method, Classification and Identification is carried out to test sample book, learning process in support vector machine classification method can be regarded as one and optimize the process finding optimum solution, therefore the effective ways designed before can adopting go the global minimum finding and find objective function, thus solve the problem that degree of depth belief network may be absorbed in locally optimal solution; Compared with prior art, Fault Diagnosis of Roller Bearings of the present invention can improve accuracy and the validity of rolling bearing fault diagnosis, thering is provided a kind of new effective way for solving rolling bearing fault diagnosis problem, can be widely used in the complication system in the fields such as chemical industry, metallurgy, electric power, aviation.

What finally illustrate is, above embodiment is only in order to illustrate technical scheme of the present invention and unrestricted, although with reference to preferred embodiment to invention has been detailed description, those of ordinary skill in the art is to be understood that, can modify to technical scheme of the present invention or equivalent replacement, and not departing from aim and the scope of technical solution of the present invention, it all should be encompassed in the middle of right of the present invention.

Claims (3)

1., based on degree of depth study and the Fault Diagnosis of Roller Bearings of support vector machine, it is characterized in that, comprise the steps:

1) during rolling bearing rotation work under four kinds of different operating modes, rolling bearing is gathered under often kind of operating mode respectively at the vibration acceleration signal of different rotating speeds work by acceleration transducer, carry out noise suppression preprocessing, and add operating mode label, using through pre-service and each vibration acceleration signal data of adding under the various operating modes after operating mode label as training sample; Described four kinds of operating modes are respectively normal operation, the running of bearing inner race fault, the running of bearing roller fault, the running of bearing outer ring fault;

2) degree of depth belief network model is set up, adopt training sample to the training of degree of depth belief network model, training sample is inputted in degree of depth belief network model, adopt non-supervisory greed successively training method successively train and tuning, obtain connection weights and the offset parameter of degree of depth belief network model;

3) using the training sample under various operating mode as the input determining the degree of depth belief network model connecting weights and offset parameter, degree of depth study is carried out to training sample, adopt the degree of depth belief network model determining to connect weights and offset parameter to be reconstructed each training sample under often kind of operating mode respectively, obtain the training sample reconstruction signal that each training sample under often kind of operating mode is corresponding;

4) gather the vibration acceleration signal data of rolling bearing to be measured when rotation work by acceleration transducer, and carry out noise suppression preprocessing, as test sample book;

5) using test sample book as the input determining the degree of depth belief network model connecting weights and offset parameter, degree of depth study is carried out to test sample book, adopt the degree of depth belief network model determining to connect weights and offset parameter to be reconstructed test sample book, obtain test sample book reconstruction signal;

6) using the matching characteristic of test sample book reconstruction signal as test sample book, using training sample reconstruction signal corresponding for each training sample under often kind of operating mode as coupling benchmark, support vector machine classification method is adopted to mate with training sample test sample book, by the operating mode classification that the operating mode kind judging belonging to the training sample mated the most with test sample book is test sample book, thus obtain the fault diagnosis result of rolling bearing to be measured.

2., according to claim 1 based on degree of depth study and the Fault Diagnosis of Roller Bearings of support vector machine, it is characterized in that, described step 2) in the joint distribution function of degree of depth belief network model set up be:

$E(v,h|\mathrm{\&theta;})=-\underset{i=1}{\overset{I}{\mathrm{\&Sigma;}}}{a}_i{v}_i-\underset{j=1}{\overset{J}{\mathrm{\&Sigma;}}}{b}_j{h}_j-\underset{i=1}{\overset{I}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{J}{\mathrm{\&Sigma;}}}{w}_{\mathrm{ij}}{v}_i{h}_j;---\left(1\right)$

Wherein, θ=(w _ij, a _i, b _j) be degree of depth belief network model parameter, w _ijrepresent visible layer i-th node v of degree of depth belief network _iwith a hidden layer jth node h _jbetween connection weights, a _iand b _jrepresent visible layer i-th node v respectively _ioffset parameter and a hidden layer jth node h _joffset parameter;

Adopt non-supervisory greed successively training method degree of depth belief network model is successively trained and tuning, concrete mode is:

21) mode of successively training is adopted to train the limited Boltzmann machine in each layer of degree of depth belief network model, the visible layer that the hidden layer of the limited Boltzmann machine of low one deck exports as the limited Boltzmann machine of last layer inputs, until obtain the output of the last one deck hidden layer of degree of depth belief network model; Be specially:

The joint probability distribution Probability p (v, h| θ) of visible layer node and hidden layer node is:

p(v,h|θ)＝e ^-E(v,h|θ)/Z(θ)；(2)

Wherein, $Z\left(\mathrm{\&theta;}\right)=\underset{v}{\mathrm{\&Sigma;}}\underset{h}{\mathrm{\&Sigma;}}{e}^{-E(v,h|\mathrm{\&theta;})}$ For partition function;

When the state of given visible layer node, a hidden layer jth node h _jactivation probability be:

$p(h_j=1|v,\mathrm{\&theta;})=\mathrm{\&sigma;}(b_j+\underset{i=1}{\overset{I}{\mathrm{\&Sigma;}}}{v}_i{w}_{\mathrm{ji}});---\left(3\right)$

When the state of given hidden layer node, visible layer i-th node v _iactivation probability be:

$p(v_i=1|h,\mathrm{\&theta;})=\mathrm{\&sigma;}(a_i+\underset{i=1}{\overset{I}{\mathrm{\&Sigma;}}}{h}_j{w}_{\mathrm{ji}});---\left(4\right)$

Wherein, σ (x)=1/ (1+e ^-x) be sigmoid function;

According to above-mentioned activation probability, when given training sample is inputed to visible layer node, after adopting all nodes of joint distribution function excitation hidden layer of degree of depth belief network model, then carry out the excitation of next hidden layer node, thus regain visible layer node; Then, adopt condition distribution sdpecific dispersion algorithm being calculated to visible layer data, obtain hidden layer data, again with the condition distribution calculating gained hidden layer data, calculate visible layer data, reconstruct is realized to visible layer data, utilizes gradient descent method, to adjustment and the renewal of degree of depth belief network model parameter, the renewal difference connecting weights and offset parameter is respectively:

Δw _ij＝ε(<v _ih _j> _data-<v _ih _j> _recon)；(6)

Δa _i＝ε(<v _i> _data-<v _i> _recon)；(7)

Δb _j＝ε(<h _j> _data-<h _j> _recon)；(8)

Wherein, Δ w _ijrepresent visible layer i-th node v of degree of depth belief network _iwith a hidden layer jth node h _jbetween connection weight w _ijcarry out the renewal difference upgraded, Δ a _iwith Δ b _jrepresent visible layer i-th node v respectively _ioffset parameter a _icarry out the renewal difference that upgrades and a hidden layer jth node h _joffset parameter b _jcarry out the renewal difference upgraded; ε is the learning rate of training, <> _datarepresent the mathematical expectation on the distribution that training dataset defines, <> _reconrepresent the mathematical expectation in the distribution that the degree of depth belief network model after reconstruct exports;

Successively train through above-mentioned, until obtain the output of the last one deck hidden layer of degree of depth belief network model;

22) output of the last one deck hidden layer of degree of depth belief network model of step 21 gained is carried out to the training of counterpropagation network, and the error in classification successively back-propagation of the actual classification result of the classification results that training prediction is exported and training sample, tuning is carried out to the connection weights of each layer of degree of depth belief network model; Be specially:

The training of counterpropagation network is divided into propagated forward and back-propagating two processes; In propagated forward process, successively output layer is propagated into as input by the output of the last one deck hidden layer of degree of depth belief network model of step 21 gained, obtain the class categories predicted, and the actual classification result of operating mode label determination training sample according to training sample, again the actual classification result of the classification results of prediction and training sample is compared and obtain error in classification, this error in classification is successively returned backward, thus the parameter of tuning degree of depth training network; In back-propagating process, need the value of the sensitivity δ calculating every one deck, sensitivity δ by top-down transmission with the connection weights of Corrected Depth belief network model;

For output layer, suppose that the actual output of i-th node is o _i, desired output is d _i, so the calculation expression of the sensitivity δ of i-th node is:

δ _i＝o _i(1-o _i)(d _i-o _i)；(9)

For m hidden layer, the calculation expression of the sensitivity of i-th node is:

${\mathrm{\&delta;}}_i^m=y_i^m(1-y_i^m)\underset{j}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}^m{\mathrm{\&delta;}}_j^{m+1};---\left(10\right)$

Wherein, represent the sensitivity of m hidden layer i-th node, represent the sensitivity of m+1 hidden layer i-th node, represent the output of m hidden layer i-th node, represent m hidden layer i-th node and a m+1 hidden layer jth internodal weights that are connected;

After obtaining the sensitivity δ of each node in each hidden layer thus, carry out renewal tuning by the connection weights of following formula to degree of depth belief network model:

$w_{\mathrm{ij}}^m\&LeftArrow;w_{\mathrm{ij}}^m+{\mathrm{\&epsiv;}}_{\mathrm{fine}-\mathrm{tuning}}\&times;y_i^m{\mathrm{\&delta;}}_j^{m+1};---\left(11\right)$

$b_j^m\&LeftArrow;b_j^m+{\mathrm{\&epsiv;}}_{\mathrm{fine}-\mathrm{tuning}}\&times;{\mathrm{\&delta;}}_j^{m+1};---\left(12\right)$

Wherein, ε _fine-tuningrepresent tuning learning rate, represent the offset parameter of a m hidden layer jth node;

After tuning is carried out to the connection weights of each layer of degree of depth belief network model, finally determine connection weights and the offset parameter of entire depth belief network model.

3., according to claim 1 based on degree of depth study and the Fault Diagnosis of Roller Bearings of support vector machine, it is characterized in that, described step 6) the employing support vector machine classification method concrete mode of mating with training sample test sample book is:

61) in the training sample of four kinds of operating modes, regard positive class as wherein kth class training sample, { 1,2,3,4} is regarded other 3 class training sample as negative class, is obtained the categorised decision function f of kth class by two class support vector machines sorting techniques k ∈ _k(x):

$f_k\left(x\right)=\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_n^k{y}_{n}K(x,x_n)+b_k;$

Wherein, for kth class categorised decision function f _kin (x) n-th training sample reconstruction signal x _ncorresponding Lagrange coefficient; b _kfor kth class categorised decision function f _kthe optimal hyperlane position parameter of (x); y _nrepresent the key words sorting that the n-th training sample is corresponding, the y when the n-th training sample belongs to positive class _n=1, the y when the n-th training sample belongs to negative class _n=-1; N ∈ 1,2 ..., N}, N are the sum of the training sample of four kinds of operating modes; K (x, x _n) presentation class decision function f _kthe training sample reconstruction signal x of input quantity x relative to n-th of (x) _ngaussian radial basis function kernel function;

Obtain the categorised decision function corresponding to each operating mode in four kinds of operating modes thus;

62) input quantity using test sample book reconstruction signal as categorised decision function corresponding to four kinds of operating modes, calculate four the categorised decision functional values of test sample book reconstruction signal as input quantity, with the operating mode classification that the operating mode kind judging corresponding to wherein maximum categorised decision functional value is test sample book, obtain the fault diagnosis result of rolling bearing to be measured.