CN114091539A - Multi-mode deep learning rolling bearing fault diagnosis method - Google Patents

Multi-mode deep learning rolling bearing fault diagnosis method Download PDF

Info

Publication number
CN114091539A
CN114091539A CN202111394970.0A CN202111394970A CN114091539A CN 114091539 A CN114091539 A CN 114091539A CN 202111394970 A CN202111394970 A CN 202111394970A CN 114091539 A CN114091539 A CN 114091539A
Authority
CN
China
Prior art keywords
rolling bearing
wavelet
layer
domain
time domain
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111394970.0A
Other languages
Chinese (zh)
Inventor
雷文平
薛阳
胡鑫
李永耀
闫灏
王宏超
陈磊
陈宏�
李凌均
王丽雅
韩捷
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhengzhou Enpu Technology Co ltd
Original Assignee
Zhengzhou Enpu Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhengzhou Enpu Technology Co ltd filed Critical Zhengzhou Enpu Technology Co ltd
Priority to CN202111394970.0A priority Critical patent/CN114091539A/en
Publication of CN114091539A publication Critical patent/CN114091539A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Signal Processing (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

The invention provides a multi-mode deep learning rolling bearing fault diagnosis method, and belongs to the technical field of bearing fault diagnosis. The time domain signal, the frequency domain signal and the wavelet domain signal of the rolling bearing under different working conditions are judged by constructing a rolling bearing fault diagnosis model, and the condition that the bearing state of the rolling bearing under the current working condition is normal or a certain type of fault exists is output. By adopting the method, the vibration signal can be comprehensively analyzed from three modes of a time domain, a frequency domain and a wavelet domain, the relation among different domain characteristics can be enhanced, information possibly leaked by single domain characteristics can be supplemented, the average accuracy of the method is higher than that of single time domain WDCNN, EMD + VPMCD and the traditional fusion method, the robustness is stronger, and the accuracy of fault diagnosis of the rolling bearing can be improved.

Description

Multi-mode deep learning rolling bearing fault diagnosis method
Technical Field
The invention provides a multi-mode deep learning rolling bearing fault diagnosis method, and belongs to the technical field of bearing fault diagnosis.
Background
In the operation process of mechanical equipment, the working condition of the mechanical equipment is always complex and changeable, load and noise are two important indexes for evaluating the working condition of the mechanical equipment, and the change of the load and the existence of the noise can cause the change of the vibration characteristic of the rolling bearing to a great extent, so that the judgment of the operation state of the equipment becomes more complex and difficult. Therefore, developing the state detection and fault diagnosis of the rolling bearing under variable working conditions has become one of the important development directions in the field of mechanical vibration.
Most of the existing fault feature extraction is based on one of three modal signals of a Time domain (Time domain), a Frequency domain (Frequency) and a Wavelet domain (Wavelet domain). However, when the modal signal of the frequency domain is extracted, the energy of the bearing is transferred to the middle frequency band and the high frequency band; when a modal signal of a time domain is extracted, the defect of the rolling bearing is sensitive but the amplitude and the frequency are not sensitive enough; when extracting the modal signals of the wavelet domain, the analysis can be performed at multiple resolutions and multiple scales. Therefore, analysis of single-domain signals lacks joint extraction of modal signals of other domains, and is not suitable for all fault diagnosis conditions, especially for rolling bearings under variable working conditions, complex fault features cannot be extracted quickly, and therefore timeliness and accuracy of diagnosis are affected.
Therefore, the technology based on multi-domain feature combination is an important method for solving the problem of fault diagnosis of the rolling bearing under variable load, whether the bearing has a fault or not and what kind of fault occurs can be accurately judged under the actual working condition by the method, and then important basis is provided for maintenance of mechanical equipment, and the method has important significance for guaranteeing production safety, improving production quality and preventing accidents.
Meanwhile, when the vibration signal of the rolling bearing is collected, noise, such as factory background noise, random excitation of the sensor and other components except the vibration information, is inevitably mixed. When the noise is large, the vibration amplitude and frequency of the signal can be influenced, the vibration characteristics are directly changed, and the final diagnosis result is influenced.
If a vibration with a high frequency occurs at a certain moment of the rotating machine, the generated pulse signal is strong but has a short duration, and the vibration is likely to be just in a separated time interval and be missed, so that a false judgment or a missed judgment is caused.
By combining the above analysis, the prior art has the problems of poor timeliness and low accuracy when the fault of the rolling bearing is diagnosed.
Disclosure of Invention
The invention aims to provide a rolling bearing fault diagnosis method based on multi-mode deep learning, which is used for solving the problem of low accuracy in fault diagnosis of a rolling bearing.
In order to achieve the purpose, the invention provides a rolling bearing fault diagnosis method based on multi-mode deep learning, which comprises the following steps:
1) acquiring a time domain signal of the rolling bearing to be tested vibrating under the current working condition, and processing the time domain signal to obtain a frequency domain signal and a wavelet domain signal corresponding to the current working condition;
2) inputting a time domain signal, a frequency domain signal and a wavelet domain signal corresponding to the current working condition into a pre-established fault diagnosis model of the rolling bearing, and outputting the bearing state of the rolling bearing to be tested under the current working condition; the bearing state is normal or a corresponding fault type;
the rolling bearing fault diagnosis model comprises a primary model and a secondary model, wherein the primary model comprises a time domain convolution network model, a frequency domain convolution network model and a wavelet domain convolution network model; the time domain convolution network model is used for performing feature extraction on an input time domain signal to obtain time domain features and inputting the time domain features into the secondary model, the frequency domain convolution network model is used for performing feature extraction on an input frequency domain signal to obtain frequency domain features and inputting the frequency domain features into the secondary model, and the wavelet domain convolution network model is used for performing feature extraction on an input wavelet domain signal to obtain wavelet domain features and inputting the wavelet domain features into the secondary model; the secondary model is used for performing characteristic level fusion on the input time domain characteristics, frequency domain characteristics and wavelet domain characteristics and outputting the bearing state of the rolling bearing under the corresponding working condition;
the training data of the rolling bearing fault diagnosis model comprises time domain signals, frequency domain signals and wavelet domain signals of rolling bearings in different bearing states vibrating under different working conditions.
According to the invention, a rolling bearing fault diagnosis model is constructed to judge time domain signals, frequency domain signals and wavelet domain signals of a rolling bearing under different working conditions, and the bearing state of the rolling bearing under the current working condition is normal or a certain type of fault exists; the built fault diagnosis model of the rolling bearing comprises a primary model and a secondary model, wherein the primary model comprises a time domain convolution network model, a frequency domain convolution network model and a wavelet domain convolution network model, the time domain convolution network model can extract time domain features from input time domain signals, the frequency domain convolution network model can extract frequency domain features from input frequency domain signals, the wavelet domain convolution network model can extract wavelet domain features from input wavelet domain signals, and the secondary model can perform feature fusion on the time domain features, the frequency domain features and the wavelet domain features output by the primary model, then judge the fused features and output corresponding bearing state diagnosis results. By adopting the method and the device, the vibration signal can be comprehensively analyzed from three modes of a time domain, a frequency domain and a wavelet domain, the relation among different domain characteristics can be enhanced, information which is possibly leaked by single domain characteristics can be supplemented, and the accuracy of fault diagnosis on the rolling bearing under different working conditions can be improved.
Further, in the above method, the time domain convolution network model includes a convolution layer, a maximum pooling layer, a convolution layer, a full-link layer, a maximum pooling layer, a Flatten layer, and a full-link layer, which are connected in sequence.
Aiming at the time domain convolution network model, a specific implementation mode is provided, and the time domain convolution network model is composed of a convolution layer, a maximum pooling layer, a convolution layer, a full-connection layer, a maximum pooling layer, a Flatten layer and a full-connection layer which are sequentially connected, so that the characteristics of an input time domain signal are extracted, and the time domain characteristics are output.
Further, in the above method, the frequency domain convolution network model includes a convolution layer, a maximum pooling layer, a convolution layer, a full-link layer, a maximum pooling layer, a Flatten layer, and a full-link layer, which are connected in sequence.
Aiming at the frequency domain convolution network model, a specific implementation form is provided, and the model consists of a convolution layer, a maximum pooling layer, a convolution layer, a full-link layer, a maximum pooling layer, a Flatten layer and a full-link layer which are connected in sequence and is used for extracting the characteristics of an input frequency domain signal and outputting the frequency domain characteristics.
Further, in the above method, the wavelet domain convolution network model includes a plurality of convolution modules, a scatter layer, and a fully-connected layer, which are connected in sequence, and the convolution module includes a convolution layer, a maximum pooling layer, and a BN layer, which are connected in sequence.
The convolution module is formed by sequentially connecting convolution layers, a maximum pooling layer and a BN layer and is used for processing input wavelet domain signals and outputting wavelet domain characteristics.
Further, in the above method, the number of convolution modules in the wavelet domain convolution network model is 4.
The wavelet domain convolution network model is connected with the Flatten layer and the full connection layer in sequence by adopting 4 convolution modules, and can accurately extract wavelet domain characteristics in wavelet domain signals.
Further, in the above method, the two-stage model performs feature-level fusion on the input time domain feature, frequency domain feature and wavelet domain feature by using the following formula:
Figure BDA0003369945850000041
in the formula, ZaddRepresenting the output of the two-level model, XiRepresenting the time domain characteristics of the ith node, YiRepresenting the frequency domain characteristics of the ith node, ZiFeatures of wavelet domain, W, representing the ith nodeiAnd the weight of the ith node is represented, c represents the number of output nodes of the full connecting layer in the fault diagnosis model of the rolling bearing, and i is more than or equal to 1 and less than or equal to c.
The secondary model fuses the time domain features, the frequency domain features and the wavelet domain features in a weighting combination mode, the information quantity of the fused features can be increased by adopting the formula, and meanwhile, the dimension of the model is guaranteed not to change.
Further, in the above method, the wavelet decomposition is performed on the time domain signal by using the DB4 wavelet function, so as to obtain the wavelet domain signal under the corresponding working condition.
A specific implementation mode is provided for the process of obtaining wavelet domain signals from time domain signal processing, and a DB4 wavelet function is adopted as a wavelet basis function to perform wavelet decomposition on the time domain signals, so that corresponding wavelet domain signals are obtained.
Further, in the above method, when the wavelet decomposition is performed on the time domain signal by using the DB4 wavelet function, the obtained fourth layer component is used as the wavelet domain signal under the corresponding working condition.
After the time domain signal is subjected to wavelet decomposition by using the DB4 wavelet function, the vibration characteristic component of the fourth layer is more obvious than that of other layers as can be seen from the Hilbert envelope spectrum, so the signal decomposition quantity of the fourth layer is used as the wavelet domain signal.
Further, in the above method, the failure types include an inner ring failure, an outer ring failure, and a rolling body failure.
According to the fault types of the rolling bearings appearing in the actual working conditions, the rolling bearings are classified and sorted, and are divided into three types, namely inner ring faults, outer ring faults and rolling body faults, so that the fault diagnosis requirement can be met, the workload can be reduced, and the diagnosis efficiency can be improved.
Further, in the above method, gaussian white noise is added to time domain signals of the rolling bearings of different bearing states in the training data, which vibrate under different working conditions, to simulate environmental noise.
Gaussian white noise is added into the time domain signal to simulate the noise of the rotating machine and the surrounding factory environment in the operation process, so that the data of the training model can be closer to the real environment, and the reliability and the accuracy of model judgment are improved.
Drawings
FIG. 1 is a block flow diagram of a rolling bearing fault diagnosis method in an embodiment of the present invention;
FIG. 2 is a schematic diagram of a WTF-MCNN network according to an embodiment of the present invention;
FIG. 3 is a schematic structural diagram of a bearing test bed according to an embodiment of the present invention;
FIG. 4 is a schematic diagram illustrating a time-domain signal data interception method according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of the Hilbert envelope spectrum of the first to fourth layer components of the wavelet signal corresponding to the inner ring fault IR07 in an embodiment of the present invention;
FIG. 6 is a block diagram illustrating a process for training a WTF-MCNN network according to an embodiment of the invention;
FIG. 7 is a schematic diagram of the distribution of the diagnosis accuracy of the variable load diagnosis experiment performed by the conventional method, the deep learning method and the method of the present invention under the noise with the signal-to-noise ratio of 15 in the embodiment of the present invention;
FIG. 8 is a schematic diagram of the distribution of the diagnosis accuracy of the variable load diagnosis experiment performed by the conventional method, the deep learning method and the method of the present invention under the noise with the signal-to-noise ratio of 10 in the embodiment of the present invention;
FIG. 9 is a schematic diagram of the distribution of the diagnosis accuracy of the variable load diagnosis experiment performed by the conventional method, the deep learning method and the method of the present invention under the noise with the signal-to-noise ratio of 5 in the embodiment of the present invention;
FIG. 10 is a diagnostic experiment result confusion matrix chart using the WTF-MCNN algorithm under the noise with the signal-to-noise ratio of 5 according to the embodiment of the present invention.
In the figure, 1 is a fan end, 2 is an acceleration sensor, 3 is a driving end, and 4 is a coupling torque sensor/encoder.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments.
The method comprises the following steps:
the method respectively extracts features from time domain, frequency domain and wavelet-based time-frequency domain data by using a deep learning method, fuses the features of the three fields, constructs WTF-MCNN with time domain, frequency domain and wavelet domain three-mode fusion to strengthen the connection among different domain features, and simultaneously supplements information possibly leaked by single domain features, thereby improving the accuracy of judging the rolling bearing fault in a complex environment.
As shown in fig. 1, in the rolling bearing fault diagnosis method of the multi-modal deep learning (hereinafter, referred to as the method of the present invention) of the present invention, a multi-state recognition model, i.e., a WTF-MCNN network, is constructed in a model training stage, and a vibration signal of a rolling bearing to be recognized under different loads is determined by using the model in a fault detection stage, so that it is obtained that the bearing state of the rolling bearing is normal or faulty, and if the bearing state is faulty, a fault type is output.
Specifically, as shown in fig. 2, the WTF-MCNN network constructed in the model training stage includes a first-level model and a second-level model which are connected in sequence, the first-level model includes a wavelet domain neural convolution network on the left side in fig. 2, a frequency domain neural convolution network on the right side in fig. 2, and a middle time domain neural convolution network, the second-level model includes an Add layer and a full connection layer which are connected in sequence, and the wavelet domain neural convolution network, the frequency domain neural convolution network, and the time domain neural convolution network are respectively connected to the Add layer of the second-level model.
The time domain convolution neural network is used for extracting the characteristics of an input time domain signal, and the network structure of the time domain convolution neural network comprises a convolution layer, a maximum pooling layer, a convolution layer, a full-connection layer, a maximum pooling layer, a Flatten layer and a full-connection layer which are connected in sequence. And (3) adopting a larger convolution kernel, performing maximum pooling dimension reduction on the features obtained by the first convolution layer through a maximum pooling layer, inputting the features obtained by the second convolution layer into the first full-connection layer, and reducing feature parameters through the second maximum pooling layer. The size of each data sample input by the network is 2048, in order to improve the receptive field of the filter and obtain the characteristics more comprehensively, the number of channels of a network convolution kernel is respectively set to 16 and 64, the expansion value is set to same, the activation function is ReLU, maxporoling is adopted, the size of the maximum pooling layer is 3, the number of output nodes of a full connection layer is 100, Softmax is adopted as the activation function of the output layer, a BN layer is used for carrying out batch normalization operation on the samples, and in combination with a cross entropy loss function and an Adam algorithm, each parameter is adjusted according to the network back propagation result, so that the network operates towards the direction with fast gradient decline.
The frequency domain convolution neural network is used for carrying out feature extraction on an input frequency domain signal, and the network structure of the frequency domain convolution neural network comprises a convolution layer, a maximum pooling layer, a convolution layer, a full-connection layer, a maximum pooling layer, a Flatten layer and a full-connection layer which are sequentially connected. The size of each data sample input by the network is 800, a larger convolution kernel is also adopted, the number of channels is respectively set to be 16 and 32, the activation function is ReLU, maxporoling is adopted, a spreading value (padding) is same, the number of output nodes of the full connection layer is 100, and the rest settings are consistent with the time domain convolution neural network.
For the wavelet domain neural convolution network, input data is a wavelet domain image, each image is 128 × 128 pixels after being processed by a computer, the dimension of each image is 128 × 128 × 3 due to the fact that the image is a color image and color represents signal amplitude, four convolution layers, four pooling layers and four BN (Batch Normalization) layers are arranged, the convolution layers, the pooling layers and the Batch Normalization layers are alternately arranged and then connected with a Flatten layer and a full connection layer. The size of the convolution kernel is 3 multiplied by 3, the number of channels is respectively set to be 8, 36, 64 and 128, maxpouling is adopted, the padding value is valid, the number of channels of each Batch _ size is equal to that of the convolution kernel of the corresponding layer, the number of output nodes of the full connection layer is 100, and the rest settings are consistent with the time domain convolution neural network.
The two-stage model can fuse the output characteristics of each neural convolution network by adopting a multi-mode fusion technology, and the multi-mode fusion technology comprises data-level fusion, decision-level fusion and characteristic-level fusion. In this embodiment, a feature level fusion mode is adopted to perform feature fusion on the output of each network, and an add mode is used to realize weighted combination of three features of a time domain, a frequency domain and a wavelet domain to form a WTF-MCNN three-mode fusion network, which can increase the information content of the fusion features and keep the dimensionality of the network unchanged.
For the input time domain feature XiFrequency domain characteristic YiSum wavelet domain feature ZiFusing with a mathematical expression as shown in equation (4):
Figure BDA0003369945850000081
in the formula, ZaddRepresenting the output of the two-level model, XiRepresenting the time domain characteristics of the ith node, YiRepresenting the frequency domain characteristics of the ith node, ZiFeatures of wavelet domain, W, representing the ith nodeiAnd the weight of the ith node is represented, c represents the number of output nodes of the full connecting layer in the fault diagnosis model of the rolling bearing, and i is more than or equal to 1 and less than or equal to c.
And after a WTF-MCNN three-mode fusion network is obtained, establishing a data set to train and test the WTF-MCNN three-mode fusion network.
In this embodiment, 4 sets of data sets are constructed, and each set of data set includes a time domain signal, a frequency domain signal, and a wavelet domain signal. The data set is structured as shown in table 1 below and is divided into a training set containing data under one load and a test set containing data under three other loads, for example, data set D1/023 represents the training set of data under 1HP load and the test set of data under 0HP, 2HP and 3HP loads.
TABLE 1 training/testing data set composition table for multimodal identification models
Figure BDA0003369945850000082
Figure BDA0003369945850000091
The data in the data set is obtained by the following steps:
(1) the test rig is set up to acquire the input signals required for the test. In this embodiment, adopt the electrical engineering laboratory bearing laboratory of the university of Kaiser west to test with the laboratory data set, consequently build the test bench as shown in figure 3, including motor and the load of being connected with motor drive, the one end of motor is fan end 1, and the other end is drive end 3, and the motor passes through the transmission shaft of drive end 3 to be connected with the load, still installs two acceleration sensor 2 on the motor, installs coupling torque sensor/encoder 4 on the transmission shaft. The driving end 3 selects a deep groove ball bearing with the model number of SKF6205 for testing, and when the method is actually adopted, the self-selection can be carried out according to the model number of the rolling bearing to be tested.
And testing the rolling bearings in a normal state, an inner ring fault, an outer ring fault and a rolling element fault under four loads of 0 HP-3 HP (HP), and collecting corresponding vibration data to form a variable load data set. In the present embodiment, the sampling frequency of the vibration signal is set to 12 kHZ.
The construction of the variable load data set requires interception of vibration data, and the sample interception mode shown in fig. 4 is adopted to intercept vibration signals, so that a large number of data samples are obtained. The sampling length of each group of samples is 2048, the step length is 128, 300 groups of samples are intercepted in total, rolling bearing state data with the data sample number of 6000 are obtained, the rolling bearing state data comprise 10 different types and degrees of bearing state data of a rolling bearing in a normal state, an inner ring fault, an outer ring fault and a rolling body fault, and the number of the test data of the rolling bearings in different states under different loads is shown in table 2.
TABLE 2 Rolling bearing state data table under different loads
Figure BDA0003369945850000092
Figure BDA0003369945850000101
In table 2, N represents the normal state of the bearing; IR07, IRl4 and IR21 indicated inner ring failure with lesion diameters of 7mils, 14mils and 21mils (milli-mils), respectively; b07, B14 and B21 indicate different degrees of rolling element failure; OR07, ORl4, and OR21 indicate different degrees of outer ring failure. Under the working conditions that the loads are 0HP, 1HP, 2HP and 3HP respectively, 300 vibration data of various bearing states are obtained, and 3000 vibration data are obtained in total.
Because the acquired vibration signal is a time domain signal acquired in a laboratory, in order to ensure that the finally established multi-modal identification model can reliably identify the vibration signal of the rolling bearing to be detected in the factory environment, Gaussian white noise is added into the time domain signal in the test data to simulate the noise of the rotating machine and the surrounding factory environment in the operation process. In this embodiment, gaussian white noise is added according to the following formula (1) and formula (2):
R＝10 log(x/xn) (1)
x′＝x+xn (2)
wherein R represents the signal-to-noise ratio, x represents the vibration signal of the rolling bearing under different loads, and xnRepresenting white gaussian noise, and x' representing the vibration signal of the rolling bearing under strong noise and varying load.
(2) The resulting time domain signal is processed to convert to a frequency domain signal and a wavelet domain signal.
In this embodiment, the time domain signal X is transformed by fast fourier transformnIs processed to obtain a corresponding frequency domain signal Fn. Using wavelet transformation to transform timeDomain signal XnConverted into wavelet domain signal DnSelecting DB4 wavelet function as wavelet basis function, and applying original time domain data XnWavelet decomposition is performed to obtain wavelet components of multiple layers. The continuous wavelet transform expression is:
Figure BDA0003369945850000102
in the formula, #a,b(t) is a wavelet family subfunction dependent on parameters a and b, a is a scale factor, and the sub-wavelet psi can be controlleda,b(t) the size of the image, b is a translation factor, which controls the center position of the wavelet subfunction.
In order to select the appropriate wavelet components, comparisons are made by plotting the Hilbert envelope spectra of the wavelet components of the respective layers. For example, the hubert envelope spectrums of the wavelet components of the first to fourth layers obtained after wavelet decomposition of the time domain signal of the rolling bearing with the fault type of the inner ring fault IR07 are shown in fig. 5, the hubert envelope spectrum of the wavelet component of the first layer, the hubert envelope spectrum of the wavelet component of the second layer, the hubert envelope spectrum of the wavelet component of the third layer and the hubert envelope spectrum of the wavelet component of the fourth layer are sequentially arranged from top to bottom, and the vibration characteristic of the wavelet component of the fourth layer is most obvious as seen from the figure, so that the wavelet domain signal after time domain signal conversion is selected as the wavelet domain component of the DB4 wavelet decomposition.
After obtaining the time domain signal, the frequency domain signal and the wavelet domain signal, inputting the corresponding time domain convolutional neural network, frequency domain convolutional neural network and wavelet domain convolutional neural network for testing, wherein the testing process is shown in fig. 6, and the method comprises the following steps:
and S1, preprocessing data. And obtaining a frequency domain signal and a time-frequency domain signal through data preprocessing, forming a training data set with the original time domain signal, and labeling the data set by utilizing a one-hot coding technology to finally obtain a time domain + frequency domain + wavelet domain data set.
S2, setting proper batch-size. Different modal data are different in size, different batch-sizes need to be divided into the different modal data to form n batches of modal data, and the different batches of modal data are input into corresponding neural convolution networks respectively.
And S3, initializing parameters. And (4) randomizing the initial weight W and the bias b of the network, and determining hyper-parameters such as training times, convolution kernel size, pooling size, learning rate epsilon and the like.
S4, carrying out convolution and sampling on the modal data of the ith batch, reducing neurons layer by layer, solving partial derivatives of the output results of each layer in the previous step, updating the weight W and the bias b of each layer according to the result of the derivation, setting hyper-parameters such as the learning rate epsilon and the like, determining the number of output nodes of each output layer according to the cross entropy function result and the number of layers of the network, and combining the hidden layer and the output layer in the network to form a network training model.
S5, comparing the prediction result of the network training model with the label of the data set, calculating the Loss function Loss of the objective function and the convolutional neural network, calculating the partial derivative of the parameter according to the Loss function, and adjusting the parameter of the whole network according to a back propagation algorithm to enable the network to operate towards the direction of fast gradient decline.
And S6, repeating the steps S4 and S5, and finishing the training of the modal data of each batch until the model indexes meet the requirements.
In the fault detection stage, when the rolling bearing under the actual working condition is subjected to fault diagnosis by adopting the invention, time domain signals of the vibration of the rolling bearing to be identified under different loads are obtained, the time domain signals are processed to obtain corresponding frequency domain signals and wavelet domain signals, a multi-mode identification model obtained in the model training stage is applied, the time domain signals, the frequency domain signals and the wavelet domain signals are respectively input into a time domain convolution neural network, a frequency domain convolution neural network and a wavelet domain convolution neural network, corresponding time domain characteristics, frequency domain characteristics and wavelet domain characteristics are output, then the time domain characteristics, the frequency domain characteristics and the wavelet domain characteristics are input into the multi-mode identification model, the fused result is judged through characteristic fusion, the state of the rolling bearing to be identified is output, and the diagnosis process is completed.
In order to verify the diagnostic performance of the WTF-MCNN three-mode fusion network in the multi-mode recognition model, a special control group is set for carrying out experiment comparison analysis.
The method is characterized in that an experiment combining EMD-VPMCD with time domain and frequency domain features is selected by adopting a traditional method, the specific process is that EMD decomposition is carried out on a vibration signal under load and noise to obtain a correlation coefficient of an intrinsic mode function set (IMF) and the signal, the first 4 layers of IMF components with the maximum correlation coefficient are selected, 8 features including 4 time domain features and 4 frequency domain features are selected to be combined with VPMCD, and the fault diagnosis and classification experiment of the rolling bearing under the noise variable load based on the traditional information fusion method is realized. Meanwhile, a deep learning method WDCNN is adopted to carry out a fault diagnosis classification experiment of the variable load rolling bearing, and EMD-VPMCD and WDCNN which are traditional methods and deep learning methods are compared and analyzed with the method, so that the diagnosis performance of the method is verified.
In order to avoid the contingency of experimental results, under the noise of signal-to-noise ratios of 15, 10 and 5 respectively, 10 repeated experiments are carried out on 4 load-changing working conditions to obtain an average value, wherein the comparison result under the condition of the signal-to-noise ratio of 15 is shown in fig. 7, and it can be seen that under the weak noise of the signal-to-noise ratio of 15, the diagnosis accuracy of the WDCNN method is reduced to 75.67%, wherein the accuracy of a D1/023 data set is higher to 86.74%, but the accuracy of D0/123 and D3/012 data sets with severe load changes is only about 71%; the mean accuracy of the EMD + VPMCD method was 92.03%, the highest diagnostic accuracy was 95.02% in the D2/013 dataset and the lowest accuracy was 88.74% in the D3/012 dataset. The WTF-MCNN tri-modal fusion network provided by the invention has the average accuracy of 95.36%, 19.69% higher than that of a WDCNN method, 3.33% higher than that of an EMD + VPMCD method, and higher accuracy than that of the other two methods under four load changes, wherein the accuracy of a D2/013 data set is 95.49%, and the diagnosis accuracy of the EMD + VPMCN method under the load changes is 0.47% lower than that of the invention, so that the method is a working condition with the closest accuracy of the two methods. Therefore, under the light-noise working condition, when the rolling bearing is tested under each load change, the diagnosis accuracy rate reaches above 93.28%, and the diagnosis effect is the best.
As shown in fig. 8, comparing and analyzing the diagnostic results of the methods under the condition of the signal-to-noise ratio of 10, it can be seen that the average accuracy of the WDCNN method is reduced to 12.20% at a moderate noise with the signal-to-noise ratio of 10, and the WDCNN method has no diagnostic capability. The average accuracy of the EMD + VPMCD method is 86.24%, wherein the accuracy of the data set tests of D1/023 and D2/013 under relatively smooth load changes is 87.69% and 88.26%, respectively, the accuracy of the data set tests of D1/023 and D2/013 under relatively smooth load changes is reduced by 6.56% and 6.76% respectively relative to the signal-to-noise ratio of 15, the reduction range is large, the accuracy of the data set tests of D0/123 and D3/012 under severe load changes is 84.61% and 84.40%, respectively, and relatively speaking, the diagnosis result of the EMD + VPD method under the noise condition on relatively smooth load changes can be adopted, and the diagnosis accuracy of the working condition on severe load changes is lower than 85%, so that the diagnosis result is not adopted. The WTF-MCNN method has the average accuracy of 93.54 percent, is reduced by 1.82 percent compared with the signal-to-noise ratio of 15, has the lowest accuracy of 91.22 percent in a D0/123 data set, and has the highest accuracy of 95.01 percent in a D1/023 data set, and the result shows that the method can still finish the accurate diagnosis of the rolling bearing fault under the variable-load working condition under the condition of moderate noise.
As shown in fig. 9, comparing and analyzing the diagnostic results of the methods under the condition of the signal-to-noise ratio 5, it can be seen that, under the strong noise of the signal-to-noise ratio 5, the accuracy of the WDCNN method is about 10% under all load changes, and is equal to the probability of random classification of fault types, so that the WDCNN method is considered to have no diagnostic capability under this condition. The average accuracy of the EMD + VPMCD method is reduced to 69.74%, and the diagnosis effect is poor. At the moment, the average accuracy of the WTF-MCNN method is reduced to 90.62%, which is 2.92% lower than that of the signal-to-noise ratio 10, the diagnosis accuracy of the D3/012 data set is the lowest, 88.05%, which is 22.93% higher than that of the EMD + VPMCD method, and the method has stronger robustness.
By combining the comparison results, it can be clearly seen that, compared with the deep learning method WDCNN and the traditional method EMD + VPMCD, the WTF-MCNN network provided by the invention has higher diagnosis accuracy for the rolling bearing faults under 4 load changes under the conditions of weak noise, moderate noise and strong noise, and although the accuracy is reduced along with the enhancement of the noise, the method of the invention has higher diagnosis accuracy, stronger robustness and better anti-noise performance from figures 7, 8 and 9.
In order to further clarify the detailed condition of various fault diagnoses of the rolling bearing under variable load under the condition of strong noise. Taking the D3/012 data set test result of a drastic load change with a signal-to-noise ratio of 5 as an example, the test result is made into a confusion matrix, as shown in fig. 10, under the working conditions of strong noise and a drastic load change, 31.6% of B021 and 30.6% of IR014 faults are misdiagnosed as B007, 33% of B021 and 18.9% of IR014 faults are misdiagnosed as B014, 2.9% of IR014 faults are misdiagnosed as IR007, and the rest of the categories are all 100% in diagnostic status. It can be seen that faults of the rolling bearing in different degrees are easily mistakenly distinguished from each other under the working condition, and in addition, faults of the inner ring and faults of the rolling body are also easily mistakenly distinguished, so that misdiagnosis is caused by confusion of the faults of the rolling body and the faults of the inner ring due to the fact that the faults of the rolling body are difficult to identify and the influence of variable load and strong noise. In the 10 types of failure, the accuracy of diagnosis of 8 types of failure is 100% except for classification error of 2 types of failure, i.e., B021 and IR 014.

Claims (10)

1. A multi-mode deep learning rolling bearing fault diagnosis method is characterized by comprising the following steps:
1) acquiring a time domain signal of the rolling bearing to be tested vibrating under the current working condition, and processing the time domain signal to obtain a frequency domain signal and a wavelet domain signal corresponding to the current working condition;
2) inputting a time domain signal, a frequency domain signal and a wavelet domain signal corresponding to the current working condition into a pre-established fault diagnosis model of the rolling bearing, and outputting the bearing state of the rolling bearing to be tested under the current working condition; the bearing state is normal or a corresponding fault type;
the rolling bearing fault diagnosis model comprises a primary model and a secondary model, wherein the primary model comprises a time domain convolution network model, a frequency domain convolution network model and a wavelet domain convolution network model; the time domain convolution network model is used for performing feature extraction on an input time domain signal to obtain time domain features and inputting the time domain features into the secondary model, the frequency domain convolution network model is used for performing feature extraction on an input frequency domain signal to obtain frequency domain features and inputting the frequency domain features into the secondary model, and the wavelet domain convolution network model is used for performing feature extraction on an input wavelet domain signal to obtain wavelet domain features and inputting the wavelet domain features into the secondary model; the secondary model is used for performing characteristic level fusion on the input time domain characteristics, frequency domain characteristics and wavelet domain characteristics and outputting the bearing state of the rolling bearing under the corresponding working condition;
the training data of the rolling bearing fault diagnosis model comprises time domain signals, frequency domain signals and wavelet domain signals of rolling bearings in different bearing states vibrating under different working conditions.
2. The rolling bearing fault diagnosis method based on the multi-modal deep learning of claim 1, wherein the time domain convolution network model comprises a convolution layer, a maximum pooling layer, a convolution layer, a full-connection layer, a maximum pooling layer, a Flatten layer and a full-connection layer which are connected in sequence.
3. The rolling bearing fault diagnosis method based on the multi-modal deep learning of claim 1, wherein the frequency domain convolution network model comprises a convolution layer, a maximum pooling layer, a convolution layer, a full-connection layer, a maximum pooling layer, a Flatten layer and a full-connection layer which are connected in sequence.
4. The rolling bearing fault diagnosis method based on the multi-modal deep learning as claimed in claim 1, wherein the wavelet domain convolution network model comprises a plurality of convolution modules, a Flatten layer and a full connection layer which are connected in sequence, and the convolution modules comprise a convolution layer, a maximum pooling layer and a BN layer which are connected in sequence.
5. The multi-modal deep learning rolling bearing fault diagnosis method according to claim 4, wherein the number of convolution modules in the wavelet domain convolution network model is 4.
6. The rolling bearing fault diagnosis method based on multi-modal deep learning according to claim 1, wherein the two-stage model adopts the following formula when performing feature-level fusion on the input time domain features, frequency domain features and wavelet domain features:
Figure FDA0003369945840000021
in the formula, ZaddRepresenting the output of the two-level model, XiRepresenting the time-domain characteristics of the ith node, YiRepresenting the frequency domain characteristics of the ith node, ZiFeatures of wavelet domain, W, representing the ith nodeiAnd the weight of the ith node is represented, c represents the number of output nodes of the full connecting layer in the fault diagnosis model of the rolling bearing, and i is more than or equal to 1 and less than or equal to c.
7. The rolling bearing fault diagnosis method of multi-modal deep learning according to claim 1, wherein wavelet decomposition is performed on the time domain signal by using a DB4 wavelet function to obtain a wavelet domain signal under a corresponding working condition.
8. The rolling bearing fault diagnosis method based on multi-modal deep learning according to claim 7, wherein when wavelet decomposition is performed on time domain signals by adopting a DB4 wavelet function, the obtained fourth layer component is used as a wavelet domain signal under a corresponding working condition.
9. The multi-modal deep learning rolling bearing failure diagnosis method according to claim 1, wherein the failure types include an inner ring failure, an outer ring failure, and a rolling body failure.
10. The multi-modal deep learning rolling bearing fault diagnosis method according to claim 1, wherein gaussian white noise is added to time domain signals of rolling bearings of different bearing states in the training data that vibrate under different working conditions to simulate environmental noise.
CN202111394970.0A 2021-11-23 2021-11-23 Multi-mode deep learning rolling bearing fault diagnosis method Pending CN114091539A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111394970.0A CN114091539A (en) 2021-11-23 2021-11-23 Multi-mode deep learning rolling bearing fault diagnosis method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111394970.0A CN114091539A (en) 2021-11-23 2021-11-23 Multi-mode deep learning rolling bearing fault diagnosis method

Publications (1)

Publication Number Publication Date
CN114091539A true CN114091539A (en) 2022-02-25

Family

ID=80303468

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111394970.0A Pending CN114091539A (en) 2021-11-23 2021-11-23 Multi-mode deep learning rolling bearing fault diagnosis method

Country Status (1)

Country Link
CN (1) CN114091539A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115618214A (en) * 2022-10-14 2023-01-17 南京天洑软件有限公司 Fault diagnosis method and device for rotating equipment
CN115841082A (en) * 2023-02-22 2023-03-24 天津佰焰科技股份有限公司 Gas station abnormity diagnosis system and method
CN117556344A (en) * 2024-01-08 2024-02-13 浙江大学 Fault diagnosis method and system for ball mill transmission system based on multi-source information fusion

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115618214A (en) * 2022-10-14 2023-01-17 南京天洑软件有限公司 Fault diagnosis method and device for rotating equipment
CN115841082A (en) * 2023-02-22 2023-03-24 天津佰焰科技股份有限公司 Gas station abnormity diagnosis system and method
CN115841082B (en) * 2023-02-22 2023-07-25 天津佰焰科技股份有限公司 Abnormality diagnosis system and method for gas station
CN117556344A (en) * 2024-01-08 2024-02-13 浙江大学 Fault diagnosis method and system for ball mill transmission system based on multi-source information fusion
CN117556344B (en) * 2024-01-08 2024-05-14 浙江大学 Fault diagnosis method and system for ball mill transmission system based on multi-source information fusion

Similar Documents

Publication Publication Date Title
CN114091539A (en) Multi-mode deep learning rolling bearing fault diagnosis method
CN110059601B (en) Intelligent fault diagnosis method for multi-feature extraction and fusion
CN111914883B (en) Spindle bearing state evaluation method and device based on deep fusion network
CN112308147B (en) Rotary machinery fault diagnosis method based on multi-source domain anchor adapter integrated migration
Li et al. WavCapsNet: An interpretable intelligent compound fault diagnosis method by backward tracking
CN112257530B (en) Rolling bearing fault diagnosis method based on blind signal separation and support vector machine
CN109858352A (en) A kind of method for diagnosing faults based on compressed sensing and the multiple dimensioned network of improvement
CN110110768A (en) Fault Diagnosis of Roller Bearings based on Concurrent Feature study and multi-categorizer
CN110132554A (en) A kind of rotary machinery fault diagnosis method that depth Laplce encodes certainly
CN112147432A (en) BiLSTM module based on attention mechanism, transformer state diagnosis method and system
CN112507915A (en) Method for identifying loosening state of bolt connection structure based on vibration response information
CN116793682A (en) Bearing fault diagnosis method based on iCORAL-MMD and anti-migration learning
CN115774851B (en) Method and system for detecting internal defects of crankshaft based on hierarchical knowledge distillation
CN113887136A (en) Improved GAN and ResNet based electric vehicle motor bearing fault diagnosis method
CN114091504A (en) Rotary machine small sample fault diagnosis method based on generation countermeasure network
Zhao et al. A novel deep fuzzy clustering neural network model and its application in rolling bearing fault recognition
CN111562109A (en) Deep learning state identification and diagnosis method for mechanical equipment
CN115824519A (en) Valve leakage fault comprehensive diagnosis method based on multi-sensor information fusion
Yang et al. Few-shot learning for rolling bearing fault diagnosis via siamese two-dimensional convolutional neural network
CN117524340A (en) Textile component quantitative characterization method based on multilayer one-dimensional CNN depth network
Wang et al. Auto-embedding transformer for interpretable few-shot fault diagnosis of rolling bearings
CN114676733A (en) Fault diagnosis method for complex supply and delivery mechanism based on sparse self-coding assisted classification generation type countermeasure network
CN114707534A (en) Rolling bearing fault diagnosis method under small sample data
Jianhui et al. New fault diagnosis method for rolling bearing based on PCA
CN112541524A (en) BP-Adaboost multi-source information motor fault diagnosis method based on attention mechanism improvement

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination