CN111982514B - Bearing fault diagnosis method based on semi-supervised deep belief network - Google Patents

Abstract

The application provides a bearing fault diagnosis method based on a semi-supervised deep belief network, which includes the following steps: collecting vibration signal data; performing wavelet transform noise reduction on the vibration signal data and completing reconstruction; performing reconstruction on the reconstructed signal data according to the label Classify and extract the corresponding time-domain features; filter the classified signal data according to the set rules, and divide the filtered data into training data sets and test data sets; input the training data sets into the semi-supervised deep belief network, and The network is trained in depth; the test data set is input into the semi-supervised deep belief network, and the fault classification is performed on the data in the test data set through the model after deep training. Beneficial effects of the present application: inputting the working data can directly output the judged bearing fault, which has a high degree of automation; this method is more suitable for bearing fault diagnosis under multiple working conditions, has stronger compatibility and universality, and improves the accuracy of bearing faults. diagnostic accuracy.

Description

A bearing fault diagnosis method based on semi-supervised deep belief network

Technical Field

The present disclosure relates to the field of semi-supervised deep belief network learning, and in particular to a bearing fault diagnosis method based on a semi-supervised deep belief network.

Background Art

For bearing fault diagnosis, the current traditional bearing fault diagnosis methods are unable to automatically extract features and require manual feature extraction and rely on expert knowledge for judgment; intelligent bearing fault diagnosis methods mostly focus on single load and single speed fault diagnosis models, which can automatically classify bearing fault types without manual feature extraction.

Traditional bearing fault diagnosis methods require manual feature extraction, which is time-consuming and labor-intensive. The results after feature extraction cannot be automatically classified and require expert knowledge for judgment, further increasing labor costs and low efficiency. Intelligent bearing fault diagnosis algorithms train models for single load and single speed, which makes it impossible to adapt to bearing fault diagnosis under multiple working conditions and has low diagnostic accuracy. In addition, such methods also require a large amount of data with fault type labels, and the cost of manufacturing labeled data is expensive.

Summary of the invention

The purpose of this application is to provide a bearing fault diagnosis method based on a semi-supervised deep belief network to address the above problems.

In a first aspect, the present application provides a bearing fault diagnosis method based on a semi-supervised deep belief network, comprising the following steps:

Collect vibration signal data;

Perform wavelet transform on vibration signal data to reduce noise and complete reconstruction;

Classify the reconstructed signal data according to the labels and extract the corresponding time domain features;

The classified signal data is screened according to the set rules, and the screened data is divided into a training data set and a test data set;

Input the training data set into the semi-supervised deep belief network to perform deep training on the network;

The test data set is input into the semi-supervised deep belief network, and the data in the test data set is fault classified and judged by the deeply trained model.

According to the technical solution provided in the embodiment of the present application, the wavelet transform is performed on the vibration signal data to reduce noise and complete reconstruction, specifically including:

The vibration signal data x(t) is regarded as space V ₀ in the subspace of the function space. Space V ₀ can be decomposed into V ₁ and W ₁ , that is, V ₀ =V ₁ +W ₁ , V ₁ can be decomposed into V ₂ and W ₂ , and so on. V _j-1 can be decomposed into V _j and W _j , where j is the resolution, j = 0 or 1;

The signal decomposition and denoising process of the vibration signal data x(t) is:

其中

为分辨率为0时的线性组合权重，P₀x(t)称为x(t)在V₀上的平滑逼近，也就是x(t)在分辨率为0情况下的概貌，φ_0k(t)为尺度函数；

in

is the linear combination weight when the resolution is 0, P ₀ x(t) is called the smooth approximation of x(t) on V ₀ , that is, the overview of x(t) when the resolution is 0, φ _0k (t) is the scale function;

其中

为分辨率为1时的线性组合权重，P₁x(t)称为x(t)在V₁上的平滑逼近，也就是x(t)在分辨率为1情况下的概貌，φ_1k(t)为尺度函数；

in

is the linear combination weight when the resolution is 1, P ₁ x(t) is called the smooth approximation of x(t) on V ₁ , that is, the overview of x(t) when the resolution is 1, φ _1k (t) is the scaling function;

其中

为分辨率为1时离散细节，D₁x(t)为x(t)在W₁上的投影，ψ_1k(t)为小波函数；

in

is the discrete detail when the resolution is 1, D ₁ x(t) is the projection of x(t) on W ₁ , and ψ _1k (t) is the wavelet function;

P ₀ x(t)=P ₁ x(t)+D ₁ x(t);

The signals after decomposition and noise reduction are:

When j = 0,

When j = 1,

The reconstructed signal is:

According to the technical solution provided in the embodiment of the present application, the label includes the fault location information and damage size information of the bearing, the fault location information includes inner ring fault, rolling element fault and outer ring fault, and the damage size information includes 0.178mm, 0.356mm and 0.533mm; the label includes ten categories, nine of which are labels containing both fault location information and damage size information, and the tenth category of labels is a label for the health status of the bearing.

According to the technical solution provided in the embodiment of the present application, the extraction of corresponding time domain features specifically includes: extracting the mean, mean square value, root mean square value, average amplitude, kurtosis value, peak value, peak-to-peak value, standard deviation, variance, skewness value, pulse factor, skewness coefficient, waveform factor, peak coefficient, margin coefficient, kurtosis factor, and waveform entropy from the reconstructed signal data to form a total data set.

According to the technical solution provided in the embodiment of the present application, the classified signal data is screened according to the set rules, specifically including:

The maximum mean difference (MMD) algorithm is used to screen the total data set. Given two distributions s and t, the MMD between them is defined as:

其中E为对分配的期望，

为将原始数据映射到再生核希尔伯特空间(RKHS)的函数，当s与t分布相同时，MMD²(s,t)＝0，与此映射关联的内核函数为k(x^s,x^t)＝<φ(x^s),φ(x^t)>；

Where E is the expectation of the distribution,

is a function that maps the original data to the reproducing kernel Hilbert space (RKHS). When s and t have the same distribution, MMD ² (s, t) = 0, and the kernel function associated with this mapping is k(x ^s , x ^t ) = <φ(x ^s ),φ(x ^t )>;

与

分别表示分布s与分布t，MMD的经验估计如下：

and

Denote distribution s and distribution t respectively, and the empirical estimate of MMD is as follows:

When the data meet the condition of 0≤L _M (D _s ,D _t )＜0.16, they are selected as the screened data, and when they do not meet the condition, they are eliminated.

According to the technical solution provided in the embodiment of the present application, 80% of the screened data is selected as a training set and 20% as a test set, and 10% of the data in the training set is set with labels, and the remaining 90% of the data in the training set is unlabeled data, and the semi-supervised deep belief network predicts its labels and performs identification.

According to the technical solution provided in the embodiment of the present application, the inputting of the training data set into the semi-supervised deep belief network to perform deep training on the network specifically includes:

The semi-supervised deep belief network (SSDBN) is composed of a stack of semi-supervised restricted Boltzmann machines (SSRBM). SSDBN consists of an input layer, multiple hidden layers, and an output layer. Each layer has a certain number of neurons. Neurons have only two states: activated and unactivated. Neurons in each layer are not connected to each other. Neurons in each layer are fully connected, and activation functions are used to transmit information, update weights and biases. SSRBM consists of a visual layer, a hidden layer, and a supervisory layer. The energy function of SSRBM is defined as:

其中v为可视层，h为隐层，u为监督层，ψ＝(w_ij,p_kj,a_i,c_k,b_j)，a为可视层偏置值，b为隐层的偏置值，c为监督层的偏置值，λ为控制有监督学习与无监督学习比例的权重参数，w为可视层与隐层之间的权重，p为监督层间的权重，ψ＝(w_ij，p_kj，a_i，c_k，b_j)，可以更新为：

其中τ为迭代次数，η为学习率；

Where v is the visible layer, h is the hidden layer, u is the supervised layer, ψ = ( _wij , _pkj , _ai , _ck , _bj ), a is the bias value of the visible layer, b is the bias value of the hidden layer, c is the bias value of the supervised layer, λ is the weight parameter that controls the ratio of supervised learning to unsupervised learning, w is the weight between the visible layer and the hidden layer, p is the weight between the supervised layers, ψ = ( _wij , _pkj , _ai , _ck , _bj ), which can be updated as:

Where τ is the number of iterations and η is the learning rate;

The activation function connecting each layer of SSRBM is the Isigmoid function:

其中a为阈值，α为斜率，且满足：

其中α_min为使Isigmoid函数工作的最小斜率。

Where a is the threshold, α is the slope, and it satisfies:

Where α _min is the minimum slope for the Isigmoid function to work.

According to the technical solution provided in the embodiment of the present application, the test data set is input into the semi-supervised deep belief network, and the data in the test data set is subjected to fault classification and judgment through the deeply trained model, specifically including: the semi-supervised deep belief network retains the trained weights and bias matrix, the supervision layer refers to the labeled data to assist in predicting the label of the unlabeled data, and after the test data is input, it is passed from the input layer to the output layer according to the SSDBN, and finally classified.

Beneficial effects of the present invention: The present application provides a bearing fault diagnosis method based on a semi-supervised deep belief network, which can directly output the judged bearing fault by inputting working data, and has a high degree of automation; the method is more suitable for bearing fault diagnosis under multiple working conditions, has stronger compatibility and universality, and improves the accuracy of bearing fault diagnosis; and the semi-supervised deep belief network requires a small amount of labeled data during model training, which reduces training costs.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a first embodiment of the present application;

FIG2 is a diagram of the structure of a semi-supervised deep belief network in the first embodiment of the present application.

DETAILED DESCRIPTION

In order to enable those skilled in the art to better understand the technical solution of the present invention, the present application is described in detail below in conjunction with the accompanying drawings. The description in this section is only exemplary and explanatory and should not have any limiting effect on the scope of protection of the present application.

FIG1 is a flow chart of a first embodiment of the present application, comprising the following steps:

S1. Collect vibration signal data.

In this embodiment, the acceleration sensor is arranged on the outer ring part of the rotating mechanical bearing. When the vibration amplitude of the bearing is large, a magnetic adsorption acceleration sensor can be selected, and when the vibration amplitude of the bearing is small, an adhesive-attached acceleration sensor can be selected.

S2. Perform wavelet transform on the vibration signal data to reduce noise and complete reconstruction.

This step specifically includes: considering the vibration signal data x(t) as a space V ₀ in a subspace of the function space, the space V ₀ can be decomposed into V ₁ and W ₁ , that is, V ₀ =V ₁ +W ₁ , V ₁ can be decomposed into V ₂ and W ₂ , and so on, V _j-1 can be decomposed into V _j and W _j , j is the resolution, j=0 or 1;

The signal decomposition and denoising process of the vibration signal data x(t) is:

为分辨率为0时的线性组合权重，P₀x(t)称为x(t)在V₀上的平滑逼近，也就是x(t)在分辨率为0情况下的概貌，φ_0k(t)为尺度函数；in

is the linear combination weight when the resolution is 0, P ₀ x(t) is called the smooth approximation of x(t) on V ₀ , that is, the overview of x(t) when the resolution is 0, φ _0k (t) is the scale function;

为分辨率为1时的线性组合权重，P₁x(t)称为x(t)在V₁上的平滑逼近，也就是x(t)在分辨率为1情况下的概貌，φ_1k(t)为尺度函数；in

is the linear combination weight when the resolution is 1, P ₁ x(t) is called the smooth approximation of x(t) on V ₁ , that is, the overview of x(t) when the resolution is 1, φ _1k (t) is the scaling function;

为分辨率为1时离散细节，D₁x(t)为x(t)在W₁上的投影，ψ_1k(t)为小波函数；in

is the discrete detail when the resolution is 1, D ₁ x(t) is the projection of x(t) on W ₁ , and ψ _1k (t) is the wavelet function;

P ₀ x(t)＝P ₁ x(t)+D ₁ x(t)

The signals after decomposition and noise reduction are:

When j = 0,

When j = 1,

The reconstructed signal is:

In this step, after the original vibration signal data x(t) is subjected to wavelet transformation, the noise interference of the vibration signal is reduced and the data is normalized.

S3. Classify the reconstructed signal data according to the labels and extract the corresponding time domain features.

The fault data of the bearing is generally divided into three categories, namely, fault location, damage size and load speed. The label data set in this embodiment includes both fault location and damage size information. The fault location information includes inner ring fault, rolling element fault and outer ring fault. The damage size information includes 0.178mm, 0.356mm and 0.533mm. In this embodiment, the label includes ten categories of label data, nine of which are labels containing both fault location information and damage size information, and the tenth label is a label for the health status of the bearing.

In this step, the corresponding time domain features are extracted by extracting the mean, mean square value, root mean square value, average amplitude, kurtosis value, peak value, peak-to-peak value, standard deviation, variance, skewness value, pulse factor, skewness coefficient, waveform factor, peak coefficient, margin coefficient, kurtosis factor, and waveform entropy from the reconstructed signal data to form a total data set.

S4. The classified signal data is screened according to the set rules, and the screened data is divided into a training data set and a test data set.

In this embodiment, the maximum mean difference (MMD) algorithm is used to screen the total data set to select data sets with smaller feature distribution differences. Given two distributions s and t, the MMD between them is defined as:

为将原始数据映射到再生核希尔伯特空间(RKHS)的函数，当s与t分布相同时，MMD²(s,t)＝0，与此映射关联的内核函数为k(x^s,x^t)＝<φ(x^s),φ(x^t)>；Where E is the expectation of the distribution,

is a function that maps the original data to the reproducing kernel Hilbert space (RKHS). When s and t have the same distribution, MMD ² (s, t) = 0, and the kernel function associated with this mapping is k(x ^s , x ^t ) = <φ(x ^s ),φ(x ^t )>;

与

分别表示分布s与分布t，MMD的经验估计如下：

and

Denote distribution s and distribution t respectively, and the empirical estimate of MMD is as follows:

When the data meet the condition of 0≤L _M (D _s ,D _t )＜0.16, they are selected as the screened data, and when they do not meet the condition, they are eliminated.

80% of the screened data is selected as a training set and 20% as a test set, and 10% of the data in the training set is set to be labeled, and the remaining 90% of the data in the training set is unlabeled data, and the semi-supervised deep belief network predicts its labels and identifies it. In this embodiment, since the semi-supervised deep belief network greatly reduces the dependence on labels, it only needs to ensure that 10% of the data in the training set is labeled for training, which also reduces the various costs of making labeled data.

S5. Input the training data set into the semi-supervised deep belief network and perform deep training on the network.

In this embodiment, the semi-supervised deep belief network (SSDBN) is formed by stacking semi-supervised restricted Boltzmann machines (SSRBM). The SSDBN consists of an input layer, multiple hidden layers, and an output layer. As shown in FIG2 , the SSDBN in this embodiment stacks 3 SSRBMs, a total of 4 layers, each layer has a certain number of neurons, and the neurons have only two states: activated and unactivated. In this embodiment, the number of neurons in the output layer is 10, and the neurons in each layer are not connected to each other. The neurons in each layer are fully connected, and the activation function is used to transmit information, update weights and biases. The SSRBM consists of a visual layer, a hidden layer, and a supervisory layer. The energy function of the SSRBM is defined as:

其中τ为迭代次数，η为学习率；Where v is the visible layer, h is the hidden layer, u is the supervised layer, ψ = ( _wij , _pkj , _ai , _ck , _bj ), a is the bias value of the visible layer, b is the bias value of the hidden layer, c is the bias value of the supervised layer, λ is the weight parameter that controls the ratio of supervised learning to unsupervised learning, w is the weight between the visible layer and the hidden layer, p is the weight between the supervised layers, ψ = ( _wij , _pkj , _ai , _ck , _bj ) can be updated as:

Where τ is the number of iterations and η is the learning rate;

The activation function connecting each layer of SSRBM is generally a sigmoid function. Since it is easy to cause gradient vanishing and gradient exploding problems, it is modified to an Isigmoid function:

其中α_min为使Isigmoid函数工作的最小斜率。Where a is the threshold, α is the slope, and it satisfies:

Where α _min is the minimum slope for the Isigmoid function to work.

In this embodiment, the use of the Isigmoid activation function reduces the probability of gradient explosion and gradient vanishing in the model itself, making the model more stable.

S6. Input the test data set into the semi-supervised deep belief network, and use the deep trained model to perform fault classification on the data in the test data set.

In this embodiment, the semi-supervised deep belief network retains the trained weights and bias matrix, and the supervision layer refers to the labeled data to assist in predicting the labels of the unlabeled data. After the test data is input, it is passed from the input layer to the output layer according to the SSDBN and is finally classified.

In this embodiment, the bearing fault can be directly output by inputting the working data, with a high degree of automation, no need for manual feature selection, no need to rely on expert knowledge, and a fully "end-to-end" learning mode. This method is more suitable for bearing fault diagnosis under multiple working conditions, with stronger compatibility and universality, and improves the accuracy of bearing fault diagnosis. Moreover, the semi-supervised deep belief network requires a small amount of labeled data during model training, which reduces the training cost. In bearing fault diagnosis, effective data can be extracted from strong noise; there is no need to manually extract features or rely on expert experience and knowledge, and automatic and intelligent bearing fault diagnosis results can be achieved; less labeled data is used, which reduces the cost of making labeled data and improves the accuracy of bearing fault diagnosis under multiple working conditions.

This article uses specific examples to illustrate the principles and implementation methods of this application. The above examples are only used to help understand the method and its core ideas of this application. The above is only the preferred implementation method of this application. It should be pointed out that due to the limitations of textual expression and the objective existence of infinite specific structures, ordinary technicians in this technical field can make several improvements, modifications or changes without departing from the principles of this application, and can also combine the above technical features in an appropriate manner; these improvements, modifications, changes or combinations, or the direct application of the concept and technical solution of the application to other occasions without improvement, should be regarded as the scope of protection of this application.

Claims (5)

1. A bearing fault diagnosis method based on a semi-supervised deep belief network is characterized by comprising the following steps:

collecting vibration signal data;

performing wavelet transformation noise reduction on the vibration signal data and completing reconstruction;

classifying the reconstructed signal data according to the labels, and extracting corresponding time domain characteristics;

screening the classified signal data according to a set rule, and dividing the screened data into a training data set and a test data set;

inputting a training data set into a semi-supervised deep belief network, and carrying out deep training on the network;

inputting the test data set into a semi-supervised deep belief network, and carrying out fault classification judgment on the data in the test data set through a deeply trained model;

the method comprises the following steps of inputting a training data set into a semi-supervised deep belief network, and performing deep training on the network, wherein the deep training specifically comprises the following steps:

the semi-supervised deep belief network (SSDBN) is formed by stacking semi-supervised restricted Boltzmann machines (SSRBMs), the SSDBN is composed of an input layer, a plurality of hidden layers and an output layer, each layer is provided with a plurality of neurons, the neurons only have two states of activation and non-activation, the neurons of each layer are not connected with each other, the neurons in each layer are fully connected, information is transmitted by using an activation function, weight values are updated, and weight values are weighted, the SSRBMs are composed of a visible layer, a hidden layer and a supervision layer, and the energy function of the SSRBMs is defined as:

where v is the visible layer, h is the hidden layer, u is the supervisory layer, ψ = (w) _ij ,p _kj ,a _i ,c _k ,b _j ) A is the bias value of the visual layer, b is the bias value of the hidden layer, c is the bias value of the monitoring layer, lambda is the weight parameter controlling the ratio of the supervised learning to the unsupervised learning, w is the weight between the visual layer and the hidden layer, and p is the bias value between the monitoring layersWeight,/= (w) _ij ,p _kj ,a _i ,c _k ,b _j ) Can be updated as:

wherein tau is iteration times, eta is learning rate;

the activation function connecting the layers of the SSRBM is the Isigmoid function:

wherein a is a threshold, α is a slope, and satisfies:

wherein alpha is _min Minimum slope to work for the Isigmoid function;

the screening of the classified signal data according to a set rule specifically includes:

the total data set was screened using the maximum mean-difference (MMD) algorithm, given two distributions s and t, the MMD between them is defined as:

where E is the expectation for the allocation,

to map raw data to a function of the Regenerated Kernel Hilbert Space (RKHS), when s and t distributions are the same, MMD ² (s, t) =0, and the kernel function associated with this map is k (x) ^s ,x ^t )＝<φ(x ^s ),φ(x ^t )>；

And

respectively representEmpirical estimates of distribution s and distribution t, MMD are as follows:

when the data is consistent with 0 being less than or equal to L _M (D _s ,D _t ) If the condition is less than 0.16, selecting the data as screened data, and if the condition is not met, rejecting the data;

and selecting 80% of the screened data as a training set, 20% of the screened data as a test set, setting 10% of the data in the training set with labels, and the rest 90% of the data in the training set as unlabeled data, and predicting and identifying the labels by using a semi-supervised deep belief network.

2. The bearing fault diagnosis method based on the semi-supervised deep belief network as claimed in claim 1, wherein the wavelet transform denoising and reconstructing are performed on the vibration signal data, and specifically comprises:

considering the vibration signal data x (t) as a space V in a subspace of the function space ₀ Space V ₀ Can be decomposed into V ₁ And W ₁ I.e. V ₀ ＝V ₁ +W ₁ ，V ₁ Can be decomposed into V ₂ And W ₂ By analogy, V _j-1 Can be decomposed into V _j And W _j J is resolution, j =0 or 1;

the signal decomposition and noise reduction process of the vibration signal data x (t) is as follows:

wherein

Is a linear combination weight with a resolution of 0, P ₀ x (t) is x (t) at V ₀ The smooth approximation of (c), i.e. the profile of x (t) at a resolution of 0, phi _0k (t) is a scale function;

wherein

For linear combining weights at resolution 1, P ₁ x (t) is x (t) at V ₁ A smooth approximation of (i.e. a profile of x (t) at a resolution of 1, phi _1k (t) is a scale function;

wherein

Discrete details at resolution 1, D ₁ x (t) is x (t) in W ₁ Projection of _1k (t) is a wavelet function;

P ₀ x(t)＝P ₁ x(t)+D ₁ x(t)；

the decomposed and denoised signals are respectively:

when j =0, the signal is transmitted,

when j =1, the signal is transmitted,

the reconstructed signal is:

3. the bearing fault diagnosis method based on the semi-supervised deep belief network of claim 1, wherein the label comprises fault location information and damage size information of the bearing, the fault location information comprises inner ring faults, rolling body faults and outer ring faults, and the damage size information comprises 0.178mm, 0.356mm and 0.533mm; the labels comprise ten types, wherein nine types are labels simultaneously containing fault position information and damage size information, and the tenth type is labels of the bearing health state.

4. The bearing fault diagnosis method based on the semi-supervised deep belief network as claimed in claim 1, wherein the extracting of the corresponding time domain features specifically comprises: and extracting a mean value, a mean square value, a root mean square value, an average amplitude value, a kurtosis value, a peak-peak value, a standard deviation, a variance, a skewness value, a pulse factor, a skewness factor, a waveform factor, a kurtosis coefficient, a margin coefficient, a kurtosis factor and a waveform entropy from the reconstructed signal data to form a total data set.

5. The bearing fault diagnosis method based on the semi-supervised deep belief network as claimed in claim 1, wherein the step of inputting the test data set into the semi-supervised deep belief network and performing fault classification and judgment on the data in the test data set through the model after deep training specifically comprises the steps of: the semi-supervised deep belief network reserves the trained weight and bias matrix, the supervision layer refers to the label data to assist in predicting the label of the unlabelled data, the test data are transmitted from the input layer to the output layer according to the SSDBN after being input, and finally the test data are classified.

Images