CN115597869A

CN115597869A - Bearing fault diagnosis method based on MTF-SDAE-LightGBM

Info

Publication number: CN115597869A
Application number: CN202211173890.7A
Authority: CN
Inventors: 舒勇; 张彬桥; 雷钧
Original assignee: China Three Gorges University CTGU
Current assignee: China Three Gorges University CTGU
Priority date: 2022-09-26
Filing date: 2022-09-26
Publication date: 2023-01-13

Abstract

A bearing fault diagnosis method based on MTF-SDAE-LightGBM comprises the following steps: firstly, encoding and converting an acquired original one-dimensional bearing vibration signal into a two-dimensional image with time correlation preserved by using a Markov transfer field; initializing a stacking noise reduction self-encoder, optimizing the hyperparameter of the SDAE by using an African bald eager optimization algorithm so as to obtain an optimal SDAE structure, inputting a two-dimensional characteristic diagram into the optimized stacking noise reduction self-encoder to extract fault characteristics and carrying out unsupervised training on the SDAE, updating the weight and deviation of each layer of DAE by using a gradient descent method in the training process, and carrying out fine adjustment by using a small amount of labeled data after the training is finished; and finally, inputting the bearing fault into a classifier of the lightweight gradient elevator for bearing fault diagnosis and classification. The method improves the accuracy and robustness of fault diagnosis and shortens the running time of the fault diagnosis model.

Description

Bearing fault diagnosis method based on MTF-SDAE-LightGBM

Technical Field

The invention belongs to the technical field of bearing fault diagnosis, and particularly relates to a bearing fault diagnosis method based on MTF-SDAE-LightGBM.

Background

The traditional fault diagnosis is generally that an original signal is processed, and then characteristic design, selection and extraction are carried out, wherein the characteristic design, selection and extraction comprise Fourier transform, wavelet packet transform, hilbert-Huang transform, EMD decomposition and the like; and finally, sending the data to a classifier for fault diagnosis, wherein the classifier comprises a support vector machine, an extreme learning machine and the like. However, the traditional fault diagnosis method excessively depends on manual experience, the diagnosis process is complicated, the traditional fault diagnosis belongs to shallow learning, the model extraction feature capacity is not strong, the generalization capability is poor, the fault diagnosis accuracy is not high, and especially under the condition that the original data size is huge and complex, the defects of the shallow learning model are more prominent.

Compared with a shallow learning model, the deep learning model has strong advantages. The SDAE network is a very typical deep learning model, and currently achieves good results in the field of image classification and fault diagnosis. Although the fault diagnosis is performed by using the SDAE network, most of the fault diagnosis is performed by using one-dimensional vibration sequence signals, the time correlation of the signals is not considered, and the problem of information loss of fault signals exists.

The performance of the SDAE network is greatly influenced by the structural parameters of the SDAE network, and particularly three hyper-parameters, namely the number of hidden layer nodes of the network, sparse parameters and the random zero setting proportion of input data. At present, the SDAE super-parameter is mostly determined by an empirical enumeration method, and the group of parameters with the optimal effect is selected by comparing various different combinations. The method needs to reselect the hyper-parameters for the fault diagnosis problems in different fields, is time-consuming and labor-consuming, and has weak generalization capability of the model. Although a set of hyper-parameters with good effect can be found out according to specific situations, it is difficult to design an SDAE network structure which can be applied to all situations, so that the method for extracting the hyper-parameters of the SDAE network in a self-adaptive manner according to different tasks is significant for the development of the SDAE.

Therefore, the technical problem to be solved by the invention is as follows: under the noise environment, how to improve the accuracy and robustness of the bearing fault diagnosis model and how to accelerate the training time of the network.

Disclosure of Invention

In view of the technical problems in the background art, the bearing fault diagnosis method based on the MTF-SDAE-LightGBM provided by the invention can be used for solving the problems of difficulty in extracting bearing fault characteristics and low fault diagnosis accuracy rate in a noise environment and overcoming the defect that an SDAE network cannot automatically extract hyper-parameters and structural parameters of a model, the accuracy and robustness of fault diagnosis by the SDAE in the noise environment are improved, and the convergence speed of the model is accelerated.

In order to solve the technical problems, the invention adopts the following technical scheme to realize:

a bearing fault diagnosis method based on MTF-SDAE-LightGBM comprises the following steps:

the method comprises the following steps: the method comprises the steps of collecting original one-dimensional vibration data of different parts of a rolling bearing in a noise environment through a vibration sensor, and preprocessing the data.

Step two: converting the vibration data in the step one into a two-dimensional image with retention time correlation by using a Markov Transfer Field (MTF), and dividing a two-dimensional image sample into a training set and a test set according to a ratio of 3.

Step three: an initial SDAE network structure is established, the network layer number of the SDAE is set to be 3, namely the SDAE is formed by stacking 3 DAEs. The initial stacked noise reduction self-encoder (SDAE) is then optimized using the African Condore optimization algorithm (AVOA): the method is characterized in that a hyper-parameter of a stacking noise reduction self-encoder is adaptively selected by utilizing a African bald eagle optimization algorithm, wherein the hyper-parameter comprises the number of hidden layer nodes, a sparse parameter and an input data random zero-setting proportion parameter, so that an optimal stacking noise reduction self-encoder structure is obtained.

The African bald iriry optimization algorithm (AVOA) was proposed by Abdollahzadeh et al in 2021, and compared with other metaheuristic algorithms based on group intelligence, the AVOA exploration mechanism and development mechanism are more advanced and comprehensive. The AVOA enhances the exploration capability in the exploration mechanism by optimizing the random strategy, and meanwhile, the exploration capability is also improved in the exploration mechanism. Compared with other optimization algorithms, the AVOA overcomes the defect that the traditional optimization algorithm falls into the local optimal solution and has higher convergence speed. The invention utilizes AVOA to train the SDAE network, the training process is an optimizing process, and the hyperparameter of the SDAE network is determined by the minimum value of the error classification rate, thereby improving the parameter adjusting efficiency and enhancing the generalization capability of the model.

Step four: inputting the training set into an optimized SDAE to extract deep features, carrying out unsupervised training on an initial stacking noise reduction self-encoder, updating the weight and deviation of each layer of DAE by using a gradient descent method, and carrying out fine adjustment on the network by using a small amount of labeled data after the training is finished.

Step five: and establishing a lightweight gradient hoist (LightGBM) fault classifier model, inputting the two-dimensional image depth features extracted by the stacking noise reduction self-encoder into the LightGBM, and training the LightGBM classifier model.

The lightweight gradient hoist (LightGBM) algorithm was proposed by microsoft in 2017, belongs to ensemble learning, and is one of the most elegant hoisting methods at present. The LightGBM algorithm improves the training speed of the model and is suitable for the condition of large data processing amount. According to the method, the deep features extracted by the SDAE are input into the LightGBM for fault classification, so that the model training time is shortened, the fault classification accuracy is improved, and the robustness is stronger.

Step six: and inputting the test set into an optimized MTF-SDAE-LightGBM model to obtain a rolling bearing fault diagnosis classification result.

Further, the detailed process of acquiring the original vibration signal of the bearing by using the vibration sensor in the first step is that firstly, the most common 4 state data of the bearing, namely the vibration data of the outer ring fault, the inner ring fault, the rolling element fault and the normal state, are acquired, wherein the diameters of each fault of the outer ring fault, the inner ring fault and the rolling element fault are divided into 3 types, and then the data are preprocessed to obtain 10 bearing vibration sequence data with the length of n and different states.

Further, the specific method for converting the bearing vibration sequence data into the two-dimensional image in the second step is as follows:

s1: firstly, 10 kinds of sequence data (each with the length of n) are divided into Q bins (similar to quantile) according to the value range, and each data point i belongs to a unique Q _i ,i∈(1,2,...,Q)。

S2: constructing a Markov transfer matrix W, wherein the size of the matrix is as follows: [ Q, Q ]]Wherein W [ i, j]By q _i The frequency of the data in (1) is determined, and the calculation formula is as follows:

wherein, w _ij Representing quantiles q _i At quantile q _j The latter probability, in particular w _ij ＝P(x _t ∈q _i |x _t-1 ∈q _j )。

S3: constructing a Markov transition field M, wherein the matrix size is as follows: [ n, n ]],M[i,j]Has a value of W [ q ] _i ,q _j ]

Wherein m is _ij Representing transition probabilities, i.e. quantiles q _i Transfer to quantile q _j In particular m, of _ij ＝P(q _i →q _j )。

Further, an initial SDAE network structure is established in step three, specifically:

s1, setting the hidden layer of the SDAE to be 3 layers, setting the activation function of each DAE to be Sigmoid, setting the learning rate to be 0.01 and setting the batch processing size to be 100. Changing the size of the random zero-setting proportion parameter by using a dropout technique to enable input data of an input layer to be damaged, namely adding noise into the input data to form a denoising autoencoder; applying a sparsity limit to the DAE of each layer through the hidden layer sparsity coefficient rho even if most of neurons of the DAE of each layer are in an inactive state, so as to form a sparse self-encoder, and implementing the limit is to add an additional penalty term in the loss function, wherein the KL divergence is selected.

The KL divergence is defined as follows:

where ρ represents a sparse coefficient (usually, a value close to 0), and ρ is _j Representing the average activation value of the jth node of the SDAE hidden layer.

And S2, inputting the damaged sample into an input layer to train a first-layer DAE, taking an implicit layer of the damaged sample as the input of a 2 nd DAE after the training is finished, and repeating the steps until the 3 rd DAE training is finished.

And S3, stacking the 3-layer DAEs trained in the step S2 to form an SDAE, and then adding an output layer on the top layer of the SDAE network.

Further, in the third step, self-adaptive selection is carried out on the hyper-parameters of the stacked noise reduction self-encoder by using an African bald eagle optimization algorithm, and the specific steps are as follows:

s1, known from the step three, the SDAE needs 7 hyper-parameters which need to be extracted in a self-adaptive manner, and all baldor populations are set to be a 7-dimensional vector [ X [) _i1 ,X _i2 ,X _i3 ,X _i4 ,X _i5 ,X _i6 ,X _i7 ]Each population representing a hyper-parameter to be optimized, wherein X _i1 ,X _i2 ,X _i3 Representing the number of SDAE hidden layer nodes, X _i4 ,X _i5 ,X _i6 Representing the SDAE hidden layer sparse coefficient, X _i7 Representing a random zeroing scale parameter of the input data.

S2, setting parameters of AVOA, including bald eagle group size n and search stage selection parameter p ₁ Developing a first stage selection parameter p ₂ Developing a second stage selection parameter p ₃ The maximum iteration times T and the classification error rate threshold C are calculated, and the stacked noise reduction self-encoder is trained layer by layer.

S3, calculating bald irises group, and finding out the optimal bald irises at the positions: initializing the bald eagle population and calculating the fitness (classification error rate) of all the bald eagles, selecting the optimal feasible solution as the optimal bald eagle of the first team in the result, the sub-optimal feasible solution as the optimal bald eagle of the second team, and the ith bald eagle at the tth iteration moving toward the optimal bald eagle and the sub-optimal bald eagle position by the formula (7).

In the formula R _i (t) indicating the ith bald-only position at the tth iteration; bestV ₁ Indicating an optimum Condor positionPlacing; bestV ₂ Representing a sub-optimal bald eagle location; l is ₁ And L ₂ Is between [0,1]And L is a parameter of ₁ +L ₂ ＝1；p _i Expressing an optimal bald irising probability; f. of _i Indicating the degree of adaptability of other baldness irises.

S4: degree of hunger with baldness: if the bald irises are in an unburnt state, the bald irises are full of physical strength, which can cause the bald irises to forage for food at a long distance; on the contrary, if the bald irises are in the hungry state, there is no strength, and the bald irises can be only brought close to the bald irises having foods, so that they become extremely offensive. According to this behavior of the dorsalis, it is possible to divide into two stages, namely an exploration stage and a development stage, which are distinguished by the degree of hunger.

Wherein F represents the starvation degree of the ith bald only in the tth iteration, T represents the maximum iteration number, and z is [ -1,1]H is [ -2,2 ] to]A random number in between. rand ₁ Represents [0,1 ]]A random value in between.

And (3) an exploration phase: in the AVOA algorithm, baldric explores the surrounding environment first and then forages, there are two different search strategies in total, with parameter P ₁ To decide which strategy to select. Parameter P ₁ Must be given in advance, and has a range of [0,1 ]]The following formula illustrates the way of exploring a baldness vulture.

In the formula P _i (t) indicating the ith bald-only position in the tth iteration; p is _i (t + 1) represents the ith bald only position at the t +1 th iteration; x represents the bald random position, X =2 × rand, rand ∈ [0,1 ]]，P ₁ Representing a selection parameter; rand ₂ ∈[0,1]，rand ₃ ∈[0,1]，rand _p1 ∈[0,1]；u _b ,l _b Representing the search area space upper and lower limits, respectively.

And (3) in a development stage: the development stage is divided into two processes, namely a first development stage and a second development stage; if | F | ≦ 0.5 ≦ 1, bald irises are in the first stage of development; if the absolute value of F is less than 0.5 and more than or equal to 0, bald is in the second stage of development;

the first stage of development: when the temperature is higher than the set temperature

When, bald irises select food competition behaviors; when the temperature is higher than the set temperature

When the bald spirit selects the hovering flying behavior, the position updating formula is as follows:

formula of middle rand ₄ ∈[0,1],rand ₅ ∈[0,1],rand ₆ ∈[0,1],rand _p2 ∈[0,1]；P ₂ The first stage selection parameters are developed, and other parameters are as above.

And a second development stage:

the bald spirit selects the gathering behavior;

the bald spirit selects the attack behavior, and the position updating formula is as follows:

in the formula BestV ₁ (t)，BestV ₂ (t) respectively representing an optimal position and a sub-optimal position of the bald eagle at the tth iteration; rand _p3 ∈[0,1]；P ₃ Representing a second stage of development selection parameter; levy (d) represents a random walk with a step size of heavy tail distribution, bald iri being completely random and isotropic in each step direction.

S5: and continuously updating the position of the bald irises in the exploration stage and the development stage of the bald irises in the step S4, continuously iterating, calculating the fitness, and determining the optimal bald irises position.

S6: if the iteration times reach the set maximum value or the classification error rate is smaller than the set threshold value, the training is finished, and the optimal hyper-parameter of the SDAE is obtained; otherwise, returning to the step3 until the judgment condition is met.

Further, in the fourth step, the weight and deviation of the DAE of each layer are updated by using a gradient descent method in the process of unsupervised training of the stacked noise reduction self-encoder, and the specific steps are as follows:

s1: calculating the output layer l _nl The formula is as follows:

s2: calculating the residual error of the hidden layer l, and the expression is shown as follows:

wherein, i represents the ith node in the hidden layer l, and j represents the jth node in the hidden layer l + 1.

S3: partial derivatives are calculated, and the formula is as follows:

where C (W, b; x, y) represents the error function between input and output, W _ij Representing a weight matrix, b _i Indicating a hidden layer threshold.

S4: updating the network weight value, wherein the formula is as follows:

where η represents the learning rate of the network update.

Further, a lightweight gradient hoist (LightGBM) fault classifier model is established in the fifth step, specifically, the LightGBM classifier is accessed behind a feature representation layer of the stacked noise reduction self-encoder SDAE, deep features extracted by the SDAE are input into the LightGBM to train the classifier so that parameters of the classifier are optimized, and the fault diagnosis effect is optimal.

This patent can reach following beneficial effect:

1. according to the invention, an original one-dimensional sequence signal is coded into a two-dimensional characteristic image by using a Markov Transfer Field (MTF), and compared with one-dimensional sequence data, the two-dimensional characteristic image can fully exert the advantage of deep learning in image characteristic extraction, thereby enhancing the SDAE characteristic extraction capability and improving the fault diagnosis accuracy; compared with other encoding modes, the MTF can keep the time dependence of the original data.

2. Compared with the existing optimization algorithm, such as a particle swarm algorithm, a grey wolf algorithm, a genetic algorithm and the like, the African bald eager optimization algorithm has more advanced and comprehensive exploration mechanism and development mechanism, overcomes the defect that the traditional optimization algorithm is easy to fall into a local optimal solution, and has higher convergence speed. Compared with other optimization algorithms, the SDAE fault diagnosis optimized by AVOA has higher accuracy, faster model training and stronger generalization capability.

3. According to the method, the deep features extracted by the SDAE are input into the LightGBM for fault classification, and compared with other classifiers such as a support vector machine, a random forest, a multi-layer perceptron, xgboost, a deep confidence network and the like, the LightGBM is high in training speed, high in accuracy and strong in model robustness.

Drawings

The invention is further illustrated with reference to the following figures and examples:

FIG. 1 is a flowchart of the bearing fault diagnosis method based on MTF-SDAE-LightGBM according to the present invention;

FIG. 2 is a two-dimensional image diagram encoded by signals of 10 different working conditions of the bearing according to the present invention;

FIG. 3 is a diagram of a self-encoder of the present invention;

FIG. 4 is a diagram of an AVOA-SDAE network architecture according to the present invention;

FIG. 5 is a flowchart of AVOA-SDAE network training in accordance with the present invention;

Detailed Description

Example 1:

a preferred embodiment is shown in fig. 1 to 5, and a bearing fault diagnosis method based on MTF-SDAE-LightGBM includes the following steps:

step1: the method comprises the steps of acquiring original one-dimensional vibration data of different parts of a rolling bearing in a noise environment through a vibration sensor, acquiring 4 kinds of state data of inner ring faults, outer ring faults, rolling body faults and normal work, wherein each fault diameter of 3 fault states of the inner ring faults, the outer ring faults and the rolling body faults is 0.007in, 0.014in and 0.021in, and normalizing the data to obtain 10 original one-dimensional sequence samples with different state characteristics.

And 2, step: the method comprises the following steps of encoding original one-dimensional signals of 10 different states into two-dimensional characteristic images with time correlation by using a Markov Transfer Field (MTF), wherein the method specifically comprises the following steps:

2.1: firstly, dividing sequence data (with the length of n) into Q bins (similar to quantile) according to the value range, wherein each data point i belongs to a unique Q _i ,i∈(1,2,...,Q)。

2.2: constructing a Markov transfer matrix W, wherein the size of the matrix is as follows: [ Q, Q ]]Wherein W [ i, j ]]By q _i The frequency of the data in (1) is determined, and the calculation formula is as follows:

wherein, w _ij Representing quantiles q _i In quantile q _j The latter probability, in particular w _ij ＝P(x _t ∈q _i |x _t-1 ∈q _j )。

2.3: constructing a Markov transition field M, wherein the matrix size is as follows: [ n, n ]],M[i,j]Has a value of W [ q ] _i ,q _j ]

And 3, step3: after the coding is completed, dividing the two-dimensional image sample into a training set and a testing set according to the proportion of 3.

And 4, step4: establishing a 3-layer SDAE network initial structure, and carrying out self-adaptive selection on hyper-parameters of the stacking noise reduction self-encoder by utilizing an African bald eagle optimization algorithm, wherein the hyper-parameters comprise the number of nodes of a hidden layer, sparse parameters and input data random zero proportion parameters, so as to obtain an optimal stacking noise reduction self-encoder structure.

And 5: inputting the training set samples into the optimal SDAE in the step4 to extract deep features, carrying out unsupervised training on the initial stacking noise reduction self-encoder, updating the weight and the offset of the network by using a gradient descent method, and carrying out fine adjustment on the network by using a small amount of labeled samples after the training is finished.

Step 6: and establishing a lightweight gradient hoist (LightGBM) fault classifier model, inputting the two-dimensional image depth features extracted by the stacking noise reduction self-encoder into the LightGBM, and training the LightGBM classifier model.

And 7: and inputting the test set into an optimized MTF-SDAE-LightGBM model to obtain a rolling bearing fault diagnosis classification result.

With reference to fig. 3, the process of extracting the bearing fault features by continuous training of dae is as follows:

1. the single-layer stacking noise reduction self-encoder comprises an encoding network and a decoding network. The original data coded by MTF is converted into a two-dimensional characteristic image x, noise is added into the x, namely, input random zero setting is carried out to obtain a damage sample x ', the encoder maps the x' to a hidden layer for training, a decoder reconstructs the damage sample to obtain a clean output reconstruction sample y which is very similar to the input x, and the characteristics with stronger robustness can be extracted through the coding and decoding processes.

2. Randomly zeroing a part of the two-dimensional characteristic image x by random distribution to obtain a damaged sample x', wherein the random distribution is as follows:

χ～q _D (χ|X)

3. the encoding process is as follows:

h＝f(x′)＝s _f (W ₁ x′+b ₁ )

wherein s is _f As an activation function of the encoder, W ₁ As a weight matrix between the input layer to the hidden layer, b ₁ The layer-to-hidden layer bias vectors are input.

4. The decoding process is as follows:

y＝g(h)＝s _g (W ₂ h+b ₂ )

wherein s is _g As an activation function of the encoder, W ₂ As a weight matrix between the input layer to the hidden layer, b ₂ The layer-to-hidden layer bias vectors are input.

5. The activation function is chosen as Sigmoid function, defined as follows:

f(x)＝1/(1+e ^-x )

6. in order to quantitatively judge the 'good or bad' of the DAE network training, the following loss function is constructed:

wherein J (W, b) is a reconstruction error,

is a penalty term, beta is a sparse penalty factor, let beta =3,s ₂ Representing the total node number of the hidden layer;

is a mean square error term;

is L ₂ The regularization term, the purpose of which is to prevent the DAE network from overfitting, represents the weight decay factor, let λ =3e-3.

With reference to fig. 4, the steps of constructing, pre-training, and fine-tuning of sdae are as follows:

SDAE is constructed by stacking multiple DAEs, with individual DAEs belonging to a shallow network and DAEs stacked together to form a deep network. The output of the first layer DAE serves as the input of the second layer DAE, the output of the second layer DAE serves as the input of the third layer DAE, and so on.

Greedy training is carried out on the SDAE layer by layer, initial parameters of the SDAE obtained after layered pre-training comprise network weight and classification layer weight, the parameters are close to global optimal parameters, the optimization range of the SDAE parameters is narrowed, and the capability of extracting deep features is improved.

3. The SDAE network is finely adjusted by adopting a back propagation algorithm, and the updating of the weight is realized by utilizing a gradient descent algorithm, which specifically comprises the following steps:

s1: calculating the output layer l _nl The formula is as follows:

wherein i represents the ith node in the hidden layer l, and j represents the jth node in the hidden layer l + 1.

S3: partial derivatives are calculated, the formula is as follows:

where C (W, b; x, y) represents the error function between input and output, W _ij Representing a weight matrix, b _i Indicating the hidden layer threshold.

S4: updating the network weight value, wherein the formula is as follows:

where η represents the learning rate of the network update.

The following steps of extracting the SDAE network hyper-parameters by AVOA adaptation are specifically described with reference to FIG. 5:

SDAE requires 7 hyper-parameters to be extracted adaptively, and all the bald eagle groups are set as a 7-dimensional vector [ X ] _i1 ,X _i2 ,X _i3 ,X _i4 ,X _i5 ,X _i6 ,X _i7 ]Each population representing a hyper-parameter to be optimized, wherein X _i1 ,X _i2 ,X _i3 Representing the number of SDAE hidden layer nodes, X _i4 ,X _i5 ,X _i6 Representing the SDAE hidden layer sparse coefficient, X _i7 Representing the input data random zero-setting scale parameter.

2. Setting parameters of AVOA, including baleage population size n, search stage selection parameter p ₁ Developing a first stage selection parameter p ₂ Developing a second stage selection parameter p ₃ And training the stacked noise reduction self-encoder layer by layer according to the current iteration time T, the highest iteration time T and a classification error rate threshold C.

3. Calculating a bald eagle population, finding a location optimal bald eagle: initializing the bald eagle group, and calculating the adaptability (classification error rate) of all bald eagles, selecting the optimum feasible solution as the optimum bald eagle of the first team, the sub-optimum feasible solution as the optimum bald eagle of the second team in the result, and the ith bald eagle at the tth iteration moving toward the optimum bald eagle and the sub-optimum bald eagle position by the formula (7).

4. The optimal bald position is determined by updating the bald eagle position through the exploration stage and development stage of the bald eagle, i.e., equations (8), (9), (10), (11), and recalculating the fitness through continuous iteration.

5. If the iteration times T reach the set maximum value T or the classification error rate is smaller than the set threshold value C, finishing the training to obtain the optimal hyperparameter of the SDAE; otherwise, returning to the step3, wherein t = t +1, until the discrimination condition is met.

And inputting the two-dimensional image depth features extracted by the optimized stacking denoising self-encoder into the LightGBM, training a LightGBM classifier model, and obtaining a fault classification result. The fault classification algorithm flow of the LightGBM is as follows:

1. initializing LightGBM, whose output is:

wherein, L (y) _i C) a loss function for the ith sample, n the total number of samples

2. Carry out iteration

Calculating the negative gradient r _ti

Calculating the best fit value c _tj

Wherein T represents the current iteration number, T represents the maximum iteration number, R _tj J =1,2,3.. J, J representing the total number of leaf nodes.

3. Continuously updating the classifier LightGBM, wherein the updated output is:

4. judging, and if the judgment condition is met, obtaining the optimal output f (x); otherwise, returning to the step2

And finally, inputting the test set into an optimized MTF-SDAE-LightGBM model to obtain a fault diagnosis classification result of the rolling bearing.

The advantages of the method are demonstrated below with reference to a specific example:

step1, selecting an American CWRU public data set, wherein the model of a bearing is SKF6205, the sampling frequency is 12KHZ, the rotating speed is 1772r/min, the fault diameters of three fault states (inner ring fault, rolling body fault and outer ring fault) are 0.007in, 0.014in and 0.021in (1in =25.4 mm), and the original one-dimensional data of 10 different states are obtained in total by adding a normal state, as shown in Table 1.

TABLE 1CWRU data set Rolling bearing failure types

Step2, converting original one-dimensional vibration signals of 10 different states into two-dimensional characteristic images of 64 multiplied by 64 sizes by using a Markov Transfer Field (MTF), selecting 8000 samples by using an overlapping sampling technology, and respectively dividing the two-dimensional image samples into a training set and a test set according to the ratio of 3, namely 6000 samples of the training set and 2000 samples of the test set.

And Step3, establishing an SDAE network with 3 layers, initializing SDAE network parameters and AVOA parameters as shown in a table 2, and optimizing network hyper-parameters of the SDAE by utilizing the AVOA, wherein the network hyper-parameters comprise three hidden layer node numbers, three hidden layer sparse coefficients and random zero setting proportion parameters, and 7 hyper-parameters in total are shown in a table 3.

Table 2: algorithm parameters for AVOA-SDAE

Table 3: SDAE optimal hyper-parameter extracted by AVOA self-adaption

And Step4, training the optimized SDAE, adding noise into a training set to form a damaged sample, inputting the damaged sample into the SDAE, taking the output of the first layer as the input of the second layer and the output of the second layer as the input of the third layer, performing greedy training layer by layer, and updating the weight and the deviation by using a gradient descent method in order to minimize the error between the input data and the reconstructed data.

Step5, 8 different networks are used for testing, each network is tested for 10 times, the final results are averaged, and the comparison results of the different networks are shown in table 4. The table shows that the average failure diagnosis accuracy of the training set of the optimized MTF-SDAE-LightGBM model reaches 99.72%, the average failure diagnosis accuracy of the testing set reaches 99.32%, and the effectiveness and reliability of the failure diagnosis performance of the model are proved; under the same condition, the fault diagnosis accuracy of the SADE model optimized by the AVOA is higher than that of SDAE optimized by PSO, GA and GWO and SDAE not optimized, and the SADE optimized by the AVOA is proved to have higher feature extraction capability and stronger generalization capability; under the same condition, the LightGBM is used for comparing the time used for classifying the support vector machine, the softmax and the deep confidence network, and the LightGBM is proved to have higher accuracy, higher training speed and stronger model robustness.

Table 4: different network fault diagnosis result comparison

The above-described embodiments are merely preferred embodiments of the present invention, and should not be construed as limiting the present invention, and the scope of the present invention is defined by the claims, and equivalents including technical features described in the claims. I.e., equivalent alterations and modifications within the scope hereof, are also intended to be within the scope of the invention.

Claims

1. A bearing fault diagnosis method based on MTF-SDAE-LightGBM is characterized by comprising the following steps:

the method comprises the following steps: acquiring original one-dimensional vibration data of different parts of a rolling bearing in a noise environment through a vibration sensor, and preprocessing the data;

step two: converting the vibration data in the step one into a two-dimensional image with retention time correlation by using a Markov transfer field, and respectively dividing a two-dimensional image sample into a training set and a test set according to a proportion of 3;

step three: establishing an initial SDAE network structure, and setting the network layer number of the SDAE to be 3, namely the SDAE is formed by stacking 3 DAEs; then optimizing the initial stacked noise reduction self-encoder by utilizing a African bald eagle optimization algorithm: the method comprises the steps that firstly, a hyper-parameter of a stacking noise reduction self-encoder is selected in a self-adaptive mode by utilizing a African bald eagle optimization algorithm, wherein the hyper-parameter comprises the number of hidden layer nodes, a sparse parameter and an input data random zero setting proportion parameter, and therefore the optimal stacking noise reduction self-encoder structure is obtained;

step four: inputting the training set into an optimized SDAE (software development association algorithm) to extract deep features, carrying out unsupervised training on an initial stacking noise reduction self-encoder, updating the weight and deviation of each layer of DAE (data access index) by using a gradient descent method, and adjusting a network by using labeled data after training is finished;

step five: establishing a fault classifier model of the lightweight gradient elevator, inputting the two-dimensional image depth features extracted by the stacking noise reduction self-encoder into the LightGBM, and training the LightGBM classifier model;

2. The MTF-SDAE-LightGBM-based bearing fault diagnosis method of claim 1, wherein: in the first step, the detailed process of acquiring the original vibration signal of the bearing by using the vibration sensor comprises the following steps: firstly, 4 most common state data of the bearing, namely outer ring fault data, inner ring fault data, rolling element fault data and normal state vibration data are collected, wherein the diameters of each fault of the outer ring fault data, the inner ring fault data and the rolling element fault data are divided into 3 types, and then normalization processing is carried out on the data to obtain 10 bearing vibration sequence data with the length of n and different states.

3. The MTF-SDAE-LightGBM-based bearing fault diagnosis method of claim 1, wherein: in the second step, the method for converting the bearing vibration sequence data into the two-dimensional image comprises the following steps:

s1: firstly, 10 sequence data are divided into Q bins according to the value range, and each data point i belongs to a unique Q bin _i ,i∈(1,2,...,Q)；

S2: constructing a Markov transfer matrix W, wherein the size of the matrix is as follows: [ Q, Q ]]Wherein W [ i, j ]]By q _i The calculation formula of the frequency determination of the data in (1) is as follows:

wherein, w _ij Representing quantiles q _i At quantile q _j Probability of the latter, w _ij ＝P(x _t ∈q _i |x _t-1 ∈q _j )；

Wherein m is _ij Representing transition probabilities, i.e. quantiles q _i Transfer to quantile q _j Probability of (m) _ij ＝P(q _i →q _j )。

4. The MTF-SDAE-LightGBM-based bearing fault diagnosis method of claim 1, wherein: in the third step, an initial SDAE network structure is established, and the method comprises the following steps:

s1, setting a hidden layer of SDAE to be 3 layers, wherein an activation function of each DAE is Sigmoid, the learning rate is 0.01, and the batch processing size is 100; changing the size of the random zero-setting proportion parameter by using a dropout technique to damage input data of an input layer, namely adding noise into the input data to form a denoising autoencoder; applying sparse limitation to the DAE of each layer through an implied layer sparse coefficient, namely applying sparse limitation to the DAE of each layer even if most neurons of the DAE of each layer are in an inactive state, so as to form a sparse self-encoder, wherein the limitation is realized by adding an additional penalty term in a loss function, namely selecting KL divergence;

the KL divergence is defined as follows:

in the formula, ρ represents a sparse coefficient (value is usually close to 0), and ρ is _j Represents the average activation value of the jth node of the SDAE hidden layer;

s2, inputting the damaged sample into an input layer to train a first DAE layer, taking a hidden layer of the damaged sample as the input of a 2 nd DAE layer after the training is finished, and repeating the steps until the 3 rd DAE layer training is finished;

5. The MTF-SDAE-LightGBM-based bearing fault diagnosis method of claim 4, wherein in the third step, the hyper-parameters of the stacked noise reduction self-encoder are adaptively selected by African Condor optimization algorithm, and the steps are as follows:

s1, according to the third step, the SDAE needs 7 hyper-parameters which are extracted in a self-adaptive way, and all baldore groups are set as a 7-dimensional vector [ X _i1 ,X _i2 ,X _i3 ,X _i4 ,X _i5 ,X _i6 ,X _i7 ]Each population representing a hyper-parameter to be optimized, wherein X _i1 ,X _i2 ,X _i3 Representing the number of SDAE hidden layer nodes, X _i4 ,X _i5 ,X _i6 Representing the SDAE hidden layer sparse coefficient, X _i7 Representing a random zero-setting proportion parameter of input data;

s2, setting parameters of AVOA, including baldric population size n and search stage selection parameter p ₁ Developing a first stage selection parameter p ₂ Developing a second stage selection parameter p ₃ Training the stacked noise reduction self-encoder layer by layer;

s3, calculating bald irises group, and finding out the optimal bald irises at the positions: initializing a bald spirit group and calculating the fitness of all bald spirits, selecting an optimal feasible solution among results as an optimal bald spirit of a first team, a sub-optimal feasible solution as an optimal bald spirit of a second team, and an i-th bald spirit at the tth iteration moving in the direction of the optimal bald spirit and the sub-optimal bald spirit by a formula (7);

in the formula R _i (t) represents the ith bald only eagle position at the tth iteration; bestV ₁ Representing an optimal bald eagle position; bestV ₂ Representing a sub-optimal bald eagle position; l is a radical of an alcohol ₁ And L ₂ Is between [0,1]And L is a parameter of ₁ +L ₂ ＝1；p _i Expressing an optimal bald irising probability; f. of _i Representing the degree of adaptability of other baldness;

s4: degree of hunger of bald Condor: if the bald irises are in an unburnt state, the bald irises are full of physical strength, which can cause the bald irises to forage for food at a long distance; on the contrary, if the bald irises are in the hungry state, there is no free physical strength, and the bald irises can be only brought close to the bald irises having foods, so that they become extremely offensive; according to the behavior of the dorsalis, the behavior can be divided into two stages, namely an exploration stage and a development stage, and the two stages are distinguished by the hunger degree;

wherein F represents the starvation degree of the ith bald only eagle at the T-th iteration, T represents the maximum number of iterations, and z is [ -1,1]H is [ -2,2 ] to]A random number in between; rand ₁ Represents [0,1 ]]A random value in between;

and (3) an exploration phase: in the AVOA algorithm, baldric explores the surrounding environment first and then forages, there are two different search strategies in total, with parameter P ₁ To decide which strategy to select; parameter P ₁ Must be given in advance, and has a range of [0,1 ]]The following formula illustrates the exploration mode of baldness;

in the formula P _i (t) indicating the ith bald-only position in the tth iteration; p _i (t + 1) represents the location of the ith bald eagle at the time of the t +1 th iteration; x represents a bald Condor random position, X =2 × rand, rand ∈ [0,1 ∈]，P ₁ Representing a selection parameter; rand ₂ ∈[0,1]，rand ₃ ∈[0,1]，rand _p1 ∈[0,1]；u _b ,l _b Respectively representing the upper limit and the lower limit of the search area space;

and (3) in a development stage: the development stage is divided into two processes, namely a first development stage and a second development stage; if | F | ≦ 0.5 ≦ 1, bald irises are in the first stage of development; if the absolute value of F is less than or equal to 0 and less than 0.5, bald eagle is in the second stage of development;

the first stage of development: when in use

formula of medium rand ₄ ∈[0,1],rand ₅ ∈[0,1],rand ₆ ∈[0,1],rand _p2 ∈[0,1]；P ₂ Representing the development of a first stage selection parameter;

and a second development stage:

bald Queen selected Convergence behavior;

the bald irises select the attack behavior, and the position updating formula is as follows:

in the formula BestV ₁ (t)，BestV ₂ (t) respectively representing an optimal position and a sub-optimal position of the bald eagle at the tth iteration; rand _p3 ∈[0,1]；P ₃ Representing a second stage of development selection parameter; levy (d) represents random walk with step size of heavy tail distribution, bald eagle in each step direction completely random and isotropic;

s5: continuously updating the position of the bald irises through the exploration stage and the development stage of the bald irises in the step S4, continuously iterating, calculating the fitness, and determining the optimal bald irises position;

6. The MTF-SDAE-LightGBM-based bearing failure diagnosis method of claim 1, wherein in step four, the weights and deviations of DAE in each layer are updated by gradient descent method during unsupervised training of the stacked noise reduction auto-encoder, and the steps are as follows:

s1: calculating the output layer l _nl The formula is as follows:

wherein i represents the ith node in the hidden layer l, and j represents the jth node in the hidden layer l + 1;

s3: partial derivatives are calculated, and the formula is as follows:

where C (W, b; x, y) represents the error function between input and output, W _ij Representing a weight matrix, b _i Representing a hidden layer threshold;

s4: updating the network weight value, wherein the formula is as follows:

where η represents the learning rate of the network update.

7. The MTF-SDAE-LightGBM-based bearing fault diagnosis method of claim 1, wherein in step five, a lightweight gradient elevator fault classifier model is established: a LightGBM classifier is accessed behind a feature representation layer of the SDAE, deep features extracted by the SDAE are input into the LightGBM to train the classifier so that parameters of the classifier are optimal, and the fault diagnosis effect is optimal.