• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Fault diagnosis of rolling bearings with noise signal based on modified kernel principal component analysis and DC-ResNet

    2023-12-01 10:46:42YunjiZhaoMenglinZhouXiaozhuoXuNannanZhang

    Yunji Zhao | Menglin Zhou| Xiaozhuo Xu | Nannan Zhang

    1School of Electrical Engineering and Automation,Henan Polytechnic University,Jiaozuo,China

    2Henan Key Laboratory of Intelligent Detection and Control of Coal Mine Equipment,Henan Polytechnic University,Jiaozuo, China

    Abstract In view of the influence of aliasing noise on the effectiveness and accuracy of bearing fault diagnosis,a bearing fault diagnosis algorithm based on the spatial decoupling method of modified kernel principal component analysis (MKPCA) and the residual network with deformable convolution (DC-ResNet) is innovatively proposed.Firstly, the Gaussian noise with different signal-to-noise ratios (SNRs) is added to the data to simulate the different degrees of noise in the actual data acquisition process.The MKPCA is used to project the fault signal with different SNRs in the kernel space to reduce the data dimension and eliminate some noise effects.Finally, the DC-ResNet model is used to further filter the noise effects and fully extract the fault features through the training of the preprocessed data.The proposed algorithm is tested on the Case Western Reserve University(CWRU)and Xi'an Jiaotong University and Changxing Sumyoung Technology Co., Ltd.(XJTU-SY) bearing data sets with different SNR noise.The fault diagnosis accuracy can reach 100%within 30 min,which has better performance than most of the existing methods.The experimental results show that the algorithm has an excellent effect on accuracy and computation complexity under different noise levels.

    K E Y W O R D S artificial intelligence, classification, deep learning, fault feature extraction

    1 | INTRODUCTION

    ‘Made in China 2025’ proposed the comprehensive implementation of an innovation-driven development strategy, and took intelligent manufacturing as the main direction,focussing on key areas, such as intelligent terminals and high-end equipment.Rotating machinery as an important device in the industrial production process is of great significance to the development of intelligent manufacturing.

    As the core component of rotating machinery, rolling bearing plays an important role in the safe operation of rotating equipment.Rolling bearing is prone to damage and failure due to its structure and characteristics.According to statistics,30%of mechanical equipment failures caused by vibration are caused by bearing failure [1].According to the fault data to grasp the working condition of rolling bearings,timely fault analysis and diagnosis will avoid the industrial production process caused by the failure of rolling bearings, improve the safety of the production environment and reduce economic losses.Therefore,it is necessary to correctly analyse the collected vibration data that may contain noise and other irrelevant factors.The existing bearing fault diagnosis technology is almost based on datadriven methods.According to the difference of diagnosis technology,the bearing fault diagnosis algorithm can be divided into bearing fault diagnosis technology based on pattern recognition[2–5]and deep learning methods[6,7].

    Bearing fault diagnosis technology based on pattern recognition includes support vector machine(SVM)[8,9],K-nearest neighbour[10]algorithm,and backpropagation(BP)[11]algorithm.Research on bearing fault diagnosis based on pattern recognition commonly used signal processing technology to deal with the fault data collected by the sensor and feature extraction,and then use pattern recognition method for fault diagnosis and classification.Among them, the commonly used signal processing methods contain Wavelet Transform (WT) [12],Fourier Transform (FT) [13], Empirical Mode Decomposition[14],and Variational Modal Decomposition[15].The above fault diagnosis method based on pattern recognition has strong robustness and realises the fault diagnosis and classification of bearings to a certain extent.However, when the number of samples is large or unbalanced, the training effect of the model is poor, and there is a large space for progress in multi-class bearing fault classification tasks.

    With the rapid development of computer science and artificial intelligence, a deep neural network has been widely studied and applied in the field of industrial process anomaly monitoring due to its strong self-learning and adaptive characteristics.The data-driven deep learning neural network is widely used in the field of bearing fault diagnosis.The deep learning algorithm in the field of bearing fault diagnosis is mainly composed of four basic network frameworks: stack self-coding network (SAE) [16], deep belief network [17],Convolutional neural network (CNN) [18], and recurrent neural network[19].It has developed to the recent Deformable Convolution Network(DCN)[20],Residual Network(ResNet)[21], Long Short-Term Memory Network [22], Capsule Network [23], Transfer Network [24] and their variants.

    The application of deep learning in the field of bearing fault diagnosis is mainly divided into three methods: fault feature extraction based on deep learning,fault classification based on deep learning,and end-to-end fault feature extraction and fault diagnosis based on deep learning[25].For example,for the first method,[26]designed a bearing fault diagnosis method based on deep convolution neural network feature extraction and SVM fault classification for the multi-fault vibration signal of rotating machinery, which was assisted by two additional time-domain features extracted by human experts, and the diagnostic accuracy was up to 98.71%.For the second method, Qi et al.[27]proposed a fault diagnosis model based on a stacked sparse autoencoder.Withoutrequiring too much prior knowledge,deep learning is used to automatically extract more abstract and representative features.Combined with statistical analysis technology,industrial process monitoring is realised.The researchon the third method is the most extensive among the three research ideas.Zhou et al.[28] used continuous WT to convert onedimensional vibration signal into two-dimensional timefrequency image, and then used the network combined with CNN and a non-linear autoregressive neural network to extract and diagnose the fault feature.Bai et al.[29] proposed a data enhancement technology based on multi-channel CNN and multi-scale shear fusion from the requirement of training data.Firstly,the data were transformed into time-frequency images by short-time FT.Secondly,the end-to-end bearing fault diagnosis with low complexity, high robustness, and high accuracy was realised by using the designed network.

    The existing methods based on deep learning can effectively extract fault features and complete fault classification tasks.However, in view of the strong non-linear coupling of the fault adjacent features of the bearing with noise in space and the low spatial discrimination,the existing methods mostly rely on the depth network for spatial decoupling.The stronger the coupling of data features is,the deeper the depth and width of the depth network are required.A deeper and wider model means a better non-linear expression ability, which can learn more complex transformations and thus fit more complex feature inputs.However, deeper and wider models require more training samples.Under actual conditions, the richness and diversity of fault samples are limited,which can easily lead to overfitting of depth models and reduce the generalisation ability of networks.

    Therefore, in this paper, aiming at the fault data itself, a data preprocessing method of modified kernel principal component analysis (MKPCA) is introduced to decouple the spatial distribution, improve the spatial discrimination of fault categories,and realise the spatial decoupling of features.At the same time, a residual network with deformable convolution(DC-ResNet)with the ability to deal with noise interference is proposed to achieve high-precision and high-efficiency bearing fault diagnosis.Based on the decoupling ability of data and network model, this hybrid model of MKPCA and DCResNet is proposed.Firstly, Gaussian noise with different signal-to-noise ratios (SNRs) is added to the data to simulate different degrees of noise in the actual acquisition process.The improved KPCA MKPCA is used to map the kernel space of the fault signal containing different SNRs to reduce the data dimension and eliminate some noise effects.Finally,the DC-ResNet network is used to further filter the noise effects and fully extract the fault characteristics through the training of the preprocessed data.The algorithm is tested on the Case Western Reserve University (CWRU) and Xi'an Jiaotong University and Changxing Sumyoung Technology Co., Ltd.(XJTU-SY) bearing data sets with different SNR noise,and the fault diagnosis accuracy can reach 100%in 40 iterations.The experimental results show that the algorithm has an excellent effect on bearing fault diagnosis with noise.

    The innovations and contributions of this paper are summarised as follows.

    ? A novel modified KPCA (MKPCA) data preprocessing method is developed to decouple the spatial distribution,improve the spatial discrimination of fault categories, and realise the spatial decoupling of features.

    ? A new fault diagnosis method based on modified KPCA(MKPCA) and DC-ResNet deep learning algorithm is proposed to achieve 100% diagnostic accuracy for rolling bearings with noise signal.

    ? Comprehensive experiments are designed and carried out to fully verify the effectiveness of MKPCA and DC-ResNet model by comparison with CNN model and ResNet model under the noise with different SNRs.

    2 | RELATED WORK

    2.1 | Signal-to-noise ratios

    In the industrial processes, the collected bearing fault data by several sensors is normally in the form of vibration signals.Due to the mechanical characteristics of the rolling bearing,irrelevant information is inevitably contained in signals.This irrelevant information is called noise.In the process of fault diagnosis,the amount of noise contained in collected data will affect the accuracy of diagnosis to different degrees.The SNRs are proposed to quantify the level of noise contained in the signals.It is the ratio of signal power to the noise power,which is expressed by Equation (1).

    The higher the SNR,the less noise is mixed into the signal.When SNR is 0,the proportion of fault signal and noise signal is the same.Figure 1 shows the graphs of original signal,noise and their superimposed signal on CWRU data.

    wherePsdenotes the power of signals andPnis the power of noise.

    To realise a better real simulation of noise, Gaussian noise is used as the noise source to add to the public bearing data sets.In real environments, noise is often not caused by a single source, but by a complex of noise from many different sources.Assuming that real noise is regarded as the sum of many random variables with different probability distributions and each random variable is independent.In that way, the normalised sum tends to a Gaussian distribution with the increase of the number of noise sources according to the Central Limit Theorem.Based on this assumption, the use of synthetic Gaussian noise is a simple and accurate approximation to deal with this complexity without knowing what the noise distribution is.The method of adding different SNR noises is elaborated as Algorithm 1.

    Algorithm 1 Add noise with the given SNRs to the signal Input: Original signal: xi Output:Final signal x with given SNR:Fsignal 1 Generating random noise with the same dimension as the input signal:2 Pn= 1iΣi0x2i 3 Pn=Ps/10SNR 10 4 noise= ——--Pn√ / ————————————————Rnoise-R2noise■5 Fsignal=xi +noise

    2.2 | Modified kernel principal component analysis

    In the research and application of fault diagnosis based on deep learning, it is necessary to collect and analyse a large number of data that reflect the fault characteristics.These data often contain multiple variables,and there may be a correlation between these variables.A large amount of data will undoubtedly provide a lot of information to facilitate the training of more accurate neural network models, but at the same time, it will increase the complexity and calculation of the network, thereby affecting the diagnostic efficiency.However,blind reduction of data or variables may lead to the loss of important information and make the network produce wrong conclusions.

    Principal component analysis is based on the above problems.Through spatial mapping, while reducing the number of indicators to be analysed, it minimises the loss of information contained in the original data, and then makes acomprehensive analysis of the collected data.Principal component analysis is applied to linear separable cases.When the original data are non-linear,KPCA is needed.The data set is mapped from n-dimensional to high-dimensionalN>nwith linear separable, and then it is reduced from Ndimensional to a low-dimensional.

    F I G U R E 1 The signal diagram of original, noise and its superposition on CWRU data.(a) fault type 1; (b) fault type 2; (c) fault type 3; (d) fault type 4;(e) fault type 5.

    In this paper, the KPCA is improved, and a spatial mapping data preprocessing method based on modified KPCA(MKPCA) is proposed.Gaussian radial basis function is used to convert the inner product operation of the feature space after non-linear transformation into the kernel function calculation of the original space, to realise the non-linear transformation space, feature space, and category space is realised.The standard MKPCA mapping matrix of five fault categories is calculated first.Then kernel space mapping of fault signals with different SNRs is performed to reduce the data dimension and eliminate some noise effects.

    Firstly, the square of Euclidean distance between data is calculated by Equation (2):

    Secondly,the kernel matrix K is calculated by Equations(3)and (4).

    F I G U R E 2 Residual block

    Finally, the eigenvalues of the kernel matrixKare calculated, and the eigenvectors corresponding to the eigenvalues are calculated to form the mapping matrixP.

    2.3 | ResNet

    The core of the deep residual network is to introduce a deep residual learning framework into the network.Compared with the characteristic that each stack layer of a CNN such as VGG-16 is directly suitable for a desired underlying mapping,ResNet explicitly makes these layers suitable for a residual mapping[30].The residual network is composed of a series of residual blocks, as shown in Figure 2.

    A residual block can be represented by Equation (5):

    wherexlandxl+1are the input and output,F(xl,Wl) is the residual part, and refers to the gap between the predicted and observed values.The layer of the network can usually be seen asy=H(x), and a residual block of the residual network can be expressed asH(x)=F(x) +x,that is,in the unit mapping,y=H(x)is the observed value,but theH(x)predicted value,soF(x)corresponds to the residual, so it is called the residual network.

    In the convolution network,the number of feature maps ofxlandxl+1is different, and then 1×1 convolution is needed for dimension elevation or dimension reduction, as shown in Figure 3.

    F I G U R E 3 Residual block with dimensions up or down

    The residual blocks in the residual network have two forms, as shown in Figure 4 respectively.Among them,Figure 4a is applied to ResNet at 18 and 32 layers, namely ResNet-18/32, Figure 4b is applied to ResNet at 50, 101, and 152 layers, namely ResNet-50/101/152.

    3 | FAULT DIAGNOSIS OF ROLLING BEARINGS WITH NOISE SIGNAL BASED ON DC-ResNET

    F I G U R E 4 Two forms of convolution in the residual block

    Convolutional neural networks have made breakthrough achievements in image classification, behaviour recognition,and other fields due to their strong feature extraction ability.The stacked layers in the network can integrate shallow,middle, and deep features, and achieve an end-to-end classification and recognition effect.Due to the above characteristics and intelligent classification characteristics, in recent years, many studies have introduced it into the field of intelligent fault diagnosis and it plays an important role in the field of bearing fault diagnosis.In the application process,some image classification and behaviour recognition processes need to use the deep features contained in the data, so it is necessary to deepen the number of network layers to extract deep features.In the actual training process, to a certain extent, with the increase in the number of network layers, the training loss will increase and the classification accuracy will decrease.This is because the increase in the number of network layers will lead to gradient disappearance and gradient explosion.Although the addition of BatchNorm,ReLu, Dropout regularisation, and other technical means can alleviate the problem of reduced training accuracy caused by overfitting to some extent, the effect remains to be improved,and the gradient disappearance and gradient explosion are not caused by overfitting.The purpose of ResNet can effectively solve the above problems, and its successful application in bearing fault diagnosis ensures the accuracy and efficiency of diagnosis [31, 32].Figure 5 gives the overall flow of the algorithm, and Table 1 gives the specific parameters of DCResNet used in this paper.

    The residual network used in this work is to add the residual unit shown in Figure 4b to the classical VGG deep learning model to form ResNet-101.Furthermore, the deformable convolution is introduced into ResNet, replacing the ordinary convolution at the first layer of the network with the deformable convolution [33] to form the proposed DCResNet.The deformable convolution can fully extract the fault information contained in the data at the first layer of the network with its adaptive convolution kernel [34].It provides convenience for the subsequent residual module to extract fault information and noise information.The deformable convolution operation is shown in Equation (6).

    F I G U R E 5 Overall flowchart of modified kernel principal component analysis (MKPCA) + DC-ResNet.

    T A B L E 1 Parameters of DC-ResNet model

    F I G U R E 6 Network framework of modified kernel principal component analysis (MKPCA) + DC-ResNet.

    where ?tnis the offset learnt by the extra convolution layer.

    The algorithm is mainly composed of spatial mapping preprocessing of noisy data and residual network fusion deformable convolution deep learning network model.Firstly,Gaussian noise with different SNRs is added to the original data collected by the sensor to simulate the different degrees of noise interference in the process of actual data acquisition.Modified kernel principal component analysis spatial mapping is used to reduce the data dimension and eliminate part of the noise data.The data after spatial mapping is converted into a two-dimensional graphical residual network fusion deformable convolution input deep learning network,and the bearing fault diagnosis model is trained.Figure 6 shows the network diagram of model training.The data preprocessing method of MKPCA can decouple the spatial distribution, improve the spatial discrimination of fault categories,and realise the spatial decoupling of features.The DC-ResNet has the ability to deal with noise interference to achieve high-precision and highefficiency bearing fault diagnosis.The hybrid model of MKPCA and DC-ResNet proposed in this paper has an excellent effect on bearing fault diagnosis with noise.

    4 | THE DESCRIPTION OF TWO CASE STUDIES

    In the actual industrial production process, due to external factors such as overload of mechanical equipment, excessive vibration frequency, or internal factors such as fatigue of bearing materials and long service life, various degrees and different types of bearing faults will be caused.According to the bearing components and structural characteristics of rolling bearings, rolling bearing faults are often divided into wear,fracture,corrosion, and deformation of different parts.In this study,deep learning technology is used to diagnose and classify the wear, fracture, and deformation of the outer ring, inner ring, and rolling body in different degrees.

    Due to the importance and wide applicability of rolling bearings, they are often used with different mechanical equipment in various industrial backgrounds and working environments.Therefore, rolling bearings under different working environments have different working mechanisms and failure modes.In this paper,the working environment and fault types of two common bearing data sets are described in detail.

    4.1 | The case study of CWRU

    CWRU is a bearing dataset provided by Case Western Reserve University.The data acquisition experiment platform is shown in Figure 7[35],which includes the data of fan end,drive end,and ball end.The three channel fault data are respectively collected by three sensors.The CWRU dataset used in this paper contains 5 types of faults that occur to the roller, inner ring,outer ring 3 points,outer ring 6 points and outer ring 12 points.The main fault types are shown in Table 2.In Table 2,12k represents the fault sampling frequency of 12 kHz, 0.007,0.014, 0.021 refer to diameter respectively.

    F I G U R E 7 CWRU testing rig

    The CWRU dataset used in this study includes 9900 training datasets and 375 test datasets.The sizes of the training and testing datasets are [9900, 2048, 2] and [375, 2048, 2]respectively.Each dataset has 2 dimensions, and each dimension contains 2048 data points.To a certain extent, the more data,the better the stability and robustness of the deep learning network.Therefore, in order to make the dataset contain a sufficient number of samples, the data is repeatedly selected with a certain step size to prepare the training dataset and test dataset.

    T A B L E 3 Working environment of XJTU-SY bearing data

    T A B L E 2 Five fault types of CWRU datasets

    T A B L E 4 Fault types of XJTU-SY bearing data set at 2100 rpm (3 5 Hz and 12 kN)

    F I G U R E 8 XJTU-SY testing rig

    4.2 | The case study of XJTU-SY

    XJTU-SY bearing dataset provided by the Institute of Design Science and Basic Parts, Xi'an Jiaotong University [36].The test bench is shown in Figure 8.There are 3 vibration frequencies, 2100 rpm (35 Hz) and 12 kN, 2250 rpm (37.5 Hz)and 11 kN, 2400 rpm (40 Hz) and 10 kN respectively.Each frequency contains five fault types.Therefore, the data set contains 15 kinds of complete operation rolling bearings fault data.These data are obtained by several accelerated degra-dation experiments.In this study, only five fault categories are included in the 35 Hz vibration frequency.The XJTU-SY bearing data set has two channels: vertical vibration and horizontal vibration.The bearing working data of two channels are collected by two sensors in a specific working environment, as shown in Table 3 and the fault type as shown in Table 4.

    T A B L E 5 Maximum test accuracy of the model under different signal-to-noise ratios (SNRs) within 30 iterations

    T A B L E 6 Test loss and accuracy on CWRU datasets with different signal-to-noise ratios (SNRs) in 30 iterations

    All kinds of fault type data in XJTU-SY rolling bearing dataset are composed of two-dimensional horizontal vibration signals and vertical vibration signals.Only horizontal vibration signals are used in this paper.

    5 | EXPERIMENTAL RESULTS AND DISCUSSION

    Intel (R) Core (TM) I7-9700 CPU @K3.60 GHz processor with memory of 32.0 GB was used in the experiment, and NVIDIA GeForce GTX 1080 GPU was used in the experiment under the 64-bit Window system.Python was used as the programming language for MKPCA spatial mapping data preprocessing and network structure construction under the PyCharm integrated development environment.In the training and testing process of DC-ResNet fault diagnosis model,Batch-size and learning rate are 128 and 0.001, respectively,and the training link is iterative 30 times.

    F I G U R E 9 Comparison for different signal-to-noise ratios (SNR) of Convolutional neural network (CNN), ResNet and DC-ResNet on CWRU.

    F I G U R E 1 0 Comparison for different signal-to-noise ratios (SNR) of Convolutional neural network (CNN), ResNet and DC-ResNet on XJTU-SY.

    In order to collect fault data and contain noise as widely as possible in engineering operations, noise with signal-to-noise ratio of -2, 1, 2, 10, 20, 30, 60 is added to the original fault data in this experiment.When the signal-to-noise ratio is negative, that is, the noise and other irrelevant factors in the signal are higher than the useful fault data.The lower the signal-to-noise ratio, the more the noise data contained.The maximum value of the signal-to-noise ratio is 60.When the signal-to-noise ratio is 0,the proportion of the fault signal and the noise signal is the same.This algorithm is verified on the public bearing datasets of CWRU and XJTU-SY.The maximum training accuracy and test accuracy of the model are 100%within 30 iterations,as shown in Table 5.The results of CWRU and XJTU-SY datasets under different SNRs are described in the following respectively.

    Tables 6 and 7 show the test loss and accuracy on CWRU and SJTU-SY datasets under different SNRs within 30 iterations.

    Since the lower the signal-to-noise ratio is,the more noise the fault data contain, Tables 6 and 7 only show the test loss and accuracy of the fault diagnosis model when the signal-tonoise ratio is -2, 1, 2, 10, and 20.When the signal-to-noise ratio is 30 and 60, the model reaches 100% diagnostic accuracy earlier than 15 epochs.

    In order to fully verify the advantages of the DC-ResNet deep learning fault diagnosis model proposed in this paper,the fault data under different SNR are tested on the ordinary CNN (CNN) and residual network (ResNet), and the diagnostic accuracy is compared.Figures 9 and 10 show the comparison curves of the diagnostic accuracy on CNN,ResNet and DC-ResNet between the fault data without noise signal and the SNR of-2,1,2,10,and 20 on the CWRU and XJTU-SY bearing data sets respectively.

    In Figures 9 and 10,red,blue,and green lines represent the diagnostic accuracy of DC-ResNet, CNN, and ResNet networks respectively.It can be seen from Figure 9 that the red line is close to the green line, and the blue line is relatively distributed in the lower half of the image, that is, the CWRU bearing data with noise have high diagnostic accuracy and consistent diagnostic effect on DC-ResNet and ResNet networks.The effect on CNN network is poor.It can be seen from Figure 10 that there are certain differences among the three lines, and the red lines are above the other two lines under different SNRs, that is, the XJTU-SY bearing data with noise have the best diagnostic effect on DC-ResNet.Overall,the DC-ResNet proposed in this paper has an excellent effect on bearing fault diagnosis with noise.

    F I G U R E 1 1 The diagnostic accuracy histograms of different noises on CWRU dataset on Convolutional neural network(CNN),ResNet and DC-ResNet.

    F I G U R E 1 2 The diagnostic accuracy histograms of different noises onXJTU-SY dataset on Convolutional neural network (CNN), ResNet and DCResNet.

    Figures 11 and 12 show the diagnostic accuracy of CWRU and XJTU-SY data under different noise levels in three deep learning networks in the form of a histogram, which more intuitively shows the superiority of DC-ResNet in the stabilityand diagnostic accuracy of CWRU and XJTU-SY bearing data with noise.

    T A B L E 8 Accuracy and computational complexity comparison of the proposed method and other methods

    The confusion matrix is the correspondence between the predicted label and the real label.The diagnostic accuracy of the noise-containing fault data on the DC-ResNet network model under different SNRs after MKPCA high-dimensional spatial mapping is 100%, that is, the predicted label is completely corresponding to the real label.In order to avoid overstatement,only the confusion matrix of the diagnosis results of XJTU-SY bearing data on the 3 deep learning network models is typical when the SNR is-2, 1, 10, and 30.The utility of the DC-ResNet model is expressed from the correspondence between the predicted label and the real label.Figures 13–15 give confusion matrix diagrams of three deep learning network models when SNR is-2, 1, 10, and 30 respectively.

    F I G U R E 1 3 Fault diagnosis confusion matrix diagram of XJTU-SY data on Convolutional neural network (CNN) model under different SNRs.

    F I G U R E 1 4 Fault diagnosis confusion matrix diagram of XJTU-SY data on ResNet model under different SNRs.

    The number in the confusion matrix graph represents the correspondence between the predicted label and the real label,that is, the diagnostic accuracy of each fault category, 1 being the highest diagnostic accuracy.For 500 data with a real label of two in a fault category,if the predicted label is all 2,the number‘1’ appears in the rectangular block of the third row and the third column, and the number is all 0 in the other rectangular blocks of the third row.If the predicted labels are all 3,then the number ‘1’ appears in the rectangular block of the fourth column in the third row, and the number is 0 in other rectangular blocks in the third row, which is the same as the theorem.The colour depth of rectangular blocks in the graph also represents the diagnostic effect.As the colour column on the right side of each subgraph shows,the deeper the colour is,the greater the number is, that is, the more the corresponding number is.It can be seen from Figures 13–15 that for the noise-containing fault data with different SNR, the diagnostic effect of CNN network is worse than that of ResNet and DCResNet networks, and that of ResNet network is worse than that of DC-ResNet.

    In order to fully verify the usability of the proposed DCResNet model, this paper compares the computational complexity of the model with CNN and ResNet.Parameters(Params) and floating-point operations (FLOPs) are measurable indexes to show the computational complexity of a deep learning model.Table 8 presents the accuracy, Params and FLOPs of three deep learning models.It can be seen from Table 8 that the DC-ResNet model has the highest accuracy when the parameters are moderate.

    To sum up, through the experiments under different data sets and different SNR, and the comparison with CNN and ResNet network diagnostic accuracy,it is fully verified that the proposed algorithm in this chapter has high diagnostic accuracy and stable diagnostic effect for bearing fault data with noise.

    6 | CONCLUSION

    F I G U R E 1 5 Fault diagnosis confusion matrix diagram of XJTU-SY data on DC-ResNet model under different SNRs.

    The bearing fault diagnosis algorithm based on improved KPCA high-dimensional space mapping and fusion of deformable convolution residual network is designed in this chapter.Aiming at the bearing fault data containing noise factors collected by the sensor, MKPCA high-dimensional space mapping is applied to reduce the data dimension and eliminate some noise interference at the same time.The mapped data are converted into two-dimensional images to construct the training set and test set.The DC-ResNet deep learning network is used to fully extract the fault characteristics, noise information and other factors contained in the image of the training set to realise the classification and diagnosis of bearing faults.By adding noise with different SNR to CWRU and XJTU-SY common bearing data sets,the diagnosis experiment is carried out, which fully reflects the applicability of the algorithm designed in this paper to the fault diagnosis of bearing with noise, and shows the superiority of the algorithm designed in this chapter in the diagnosis accuracy.At the same time, the comparative experiments with CNN and ResNet networks show that the residual network designed in this chapter with deformable convolution is superior to the ordinary convolution neural network and the ordinary residual network in the fault diagnosis of noise bearing.

    ACKNOWLEDGEMENT

    We thank the anonymous reviewers for their valuable suggestions and comments on this work.We also wish to thank the authors of DCN for providing the source code.

    This research was funded by the Foundation of the National Natural Science Foundation of China grant number 61973105, 61573130 and 52177039; the Fundamental Research Funds for the Universities of Henan Province (NO.NSFRF200504); The Key Technologies R&D Program of Henan Province of China(NO.212102210145,212102210197 and NO.222102220016).

    CONFLICT OF INTEREST

    No potential conflicts of interest are reported by the authors.

    DATA AVAILABILITY STATEMENT

    The data that support the findings of this study are available from the corresponding author upon reasonable request.

    ORCID

    Menglin Zhouhttps://orcid.org/0000-0003-4395-5256

    精品少妇一区二区三区视频日本电影| 免费看美女性在线毛片视频| 亚洲精品在线观看二区| 又粗又爽又猛毛片免费看| 日本黄大片高清| 男人舔女人的私密视频| 搞女人的毛片| 国内精品一区二区在线观看| 女同久久另类99精品国产91| 伊人久久大香线蕉亚洲五| 国产精品久久久人人做人人爽| 美女午夜性视频免费| 18禁国产床啪视频网站| 欧美黑人欧美精品刺激| 国产精品自产拍在线观看55亚洲| 久久性视频一级片| 夜夜夜夜夜久久久久| 一卡2卡三卡四卡精品乱码亚洲| 欧美3d第一页| 国产v大片淫在线免费观看| www.熟女人妻精品国产| x7x7x7水蜜桃| 国产又色又爽无遮挡免费看| 在线a可以看的网站| 日本熟妇午夜| 欧美大码av| 精品久久久久久久人妻蜜臀av| 波多野结衣巨乳人妻| 国产熟女xx| 国产1区2区3区精品| 精品久久久久久久久久免费视频| 免费观看精品视频网站| 精品国产乱子伦一区二区三区| 成人一区二区视频在线观看| 啦啦啦观看免费观看视频高清| 午夜成年电影在线免费观看| xxx96com| 亚洲色图av天堂| 日本免费一区二区三区高清不卡| 精品人妻1区二区| 精品国产美女av久久久久小说| 香蕉丝袜av| 18美女黄网站色大片免费观看| 叶爱在线成人免费视频播放| 在线观看美女被高潮喷水网站 | 日韩中文字幕欧美一区二区| 婷婷亚洲欧美| 国产高清视频在线播放一区| 熟女电影av网| 神马国产精品三级电影在线观看 | 老鸭窝网址在线观看| www.www免费av| 日本 欧美在线| 亚洲五月天丁香| 日本一区二区免费在线视频| 国产v大片淫在线免费观看| 欧美中文综合在线视频| 国产成人av激情在线播放| 精品一区二区三区四区五区乱码| 一级黄色大片毛片| 曰老女人黄片| 久久久久久九九精品二区国产 | av有码第一页| 香蕉丝袜av| 久久九九热精品免费| 狂野欧美白嫩少妇大欣赏| 三级毛片av免费| 99久久精品国产亚洲精品| 日韩欧美在线二视频| 人妻丰满熟妇av一区二区三区| 国产高清videossex| 国产1区2区3区精品| 老司机在亚洲福利影院| 老汉色∧v一级毛片| av免费在线观看网站| 国产午夜精品久久久久久| av欧美777| 嫩草影院精品99| 欧美乱妇无乱码| 校园春色视频在线观看| 人妻丰满熟妇av一区二区三区| 白带黄色成豆腐渣| 波多野结衣高清无吗| avwww免费| 一进一出好大好爽视频| 一个人免费在线观看的高清视频| 午夜免费成人在线视频| 欧美日韩福利视频一区二区| 国模一区二区三区四区视频 | 12—13女人毛片做爰片一| 国产亚洲精品第一综合不卡| 日韩欧美三级三区| 琪琪午夜伦伦电影理论片6080| 久久精品国产亚洲av高清一级| 成年人黄色毛片网站| 校园春色视频在线观看| 久久国产精品人妻蜜桃| 黑人操中国人逼视频| 国产激情偷乱视频一区二区| 成人精品一区二区免费| 欧美激情久久久久久爽电影| 日韩欧美三级三区| 欧美在线黄色| 国产精品一及| 天堂√8在线中文| 男女视频在线观看网站免费 | 亚洲天堂国产精品一区在线| 香蕉久久夜色| 国产精品久久久久久久电影 | 日本一本二区三区精品| 国内久久婷婷六月综合欲色啪| 九色国产91popny在线| 99国产精品99久久久久| 日本免费一区二区三区高清不卡| 欧美午夜高清在线| 大型av网站在线播放| 精品日产1卡2卡| 亚洲av片天天在线观看| 国产爱豆传媒在线观看 | 久久久久久亚洲精品国产蜜桃av| 怎么达到女性高潮| 亚洲午夜精品一区,二区,三区| 18禁观看日本| 少妇人妻一区二区三区视频| 人妻夜夜爽99麻豆av| 午夜免费观看网址| 51午夜福利影视在线观看| 欧美成狂野欧美在线观看| 亚洲人成伊人成综合网2020| 国产久久久一区二区三区| 中文字幕av在线有码专区| 婷婷丁香在线五月| 精华霜和精华液先用哪个| 欧美黑人精品巨大| 亚洲熟妇熟女久久| 露出奶头的视频| 成在线人永久免费视频| 国产成人精品无人区| 亚洲,欧美精品.| 无限看片的www在线观看| 日日夜夜操网爽| 丰满人妻熟妇乱又伦精品不卡| 成人18禁在线播放| 日本在线视频免费播放| 中文字幕av在线有码专区| 校园春色视频在线观看| 夜夜爽天天搞| 久久伊人香网站| 免费在线观看视频国产中文字幕亚洲| 午夜福利高清视频| www.999成人在线观看| 国产精华一区二区三区| 两个人看的免费小视频| 一区二区三区激情视频| 亚洲av美国av| 人成视频在线观看免费观看| 国产成人一区二区三区免费视频网站| 日韩欧美国产在线观看| 香蕉久久夜色| 人人妻人人澡欧美一区二区| 亚洲无线在线观看| 丝袜人妻中文字幕| 成人亚洲精品av一区二区| 国产伦一二天堂av在线观看| 国产亚洲av高清不卡| www日本黄色视频网| 亚洲av日韩精品久久久久久密| 日本a在线网址| 亚洲欧美激情综合另类| 国产视频内射| 99国产精品一区二区三区| 18禁观看日本| 国产高清视频在线观看网站| 国产成人精品无人区| 人成视频在线观看免费观看| 亚洲va日本ⅴa欧美va伊人久久| 搞女人的毛片| 日韩欧美 国产精品| 美女免费视频网站| 色综合站精品国产| 99在线视频只有这里精品首页| 欧美中文综合在线视频| 天堂影院成人在线观看| 女生性感内裤真人,穿戴方法视频| 在线观看美女被高潮喷水网站 | 天天添夜夜摸| 熟女少妇亚洲综合色aaa.| 两性午夜刺激爽爽歪歪视频在线观看 | 国产片内射在线| 久久久久久久久中文| 夜夜躁狠狠躁天天躁| 国产精品久久久久久久电影 | 成人高潮视频无遮挡免费网站| 亚洲专区中文字幕在线| 亚洲精品粉嫩美女一区| 99久久精品国产亚洲精品| 国产精品乱码一区二三区的特点| 色噜噜av男人的天堂激情| 草草在线视频免费看| 大型av网站在线播放| 男人的好看免费观看在线视频 | 成人特级黄色片久久久久久久| 国产高清激情床上av| 亚洲性夜色夜夜综合| 欧美成人一区二区免费高清观看 | 丝袜人妻中文字幕| 亚洲国产欧洲综合997久久,| 国产高清视频在线观看网站| 久久久久久久精品吃奶| a在线观看视频网站| 最新在线观看一区二区三区| 久久 成人 亚洲| 亚洲中文字幕日韩| 18禁国产床啪视频网站| 欧美zozozo另类| 亚洲男人的天堂狠狠| 成人av在线播放网站| 正在播放国产对白刺激| 午夜精品一区二区三区免费看| 成年版毛片免费区| 亚洲真实伦在线观看| 久久精品国产清高在天天线| 久久久精品欧美日韩精品| 母亲3免费完整高清在线观看| 一本一本综合久久| 亚洲av电影不卡..在线观看| 国产在线精品亚洲第一网站| 激情在线观看视频在线高清| 全区人妻精品视频| 午夜福利高清视频| 国产黄a三级三级三级人| 亚洲九九香蕉| 一进一出抽搐动态| 久久人妻av系列| 亚洲国产日韩欧美精品在线观看 | 成人手机av| 欧美3d第一页| 国产久久久一区二区三区| 国产激情欧美一区二区| 999精品在线视频| 国产欧美日韩一区二区三| 欧美性猛交╳xxx乱大交人| 亚洲一卡2卡3卡4卡5卡精品中文| 最新美女视频免费是黄的| 白带黄色成豆腐渣| 欧美精品亚洲一区二区| 欧美 亚洲 国产 日韩一| 亚洲精品一卡2卡三卡4卡5卡| 色在线成人网| 免费搜索国产男女视频| 亚洲欧洲精品一区二区精品久久久| 在线十欧美十亚洲十日本专区| av中文乱码字幕在线| 淫秽高清视频在线观看| 禁无遮挡网站| 午夜亚洲福利在线播放| 伦理电影免费视频| 久久久久久九九精品二区国产 | 在线看三级毛片| netflix在线观看网站| 天天躁狠狠躁夜夜躁狠狠躁| 中文字幕人妻丝袜一区二区| 久久九九热精品免费| 国产69精品久久久久777片 | 亚洲成人国产一区在线观看| 一边摸一边抽搐一进一小说| 精品久久久久久成人av| 激情在线观看视频在线高清| 久久婷婷人人爽人人干人人爱| 母亲3免费完整高清在线观看| 精品日产1卡2卡| 久久久久久亚洲精品国产蜜桃av| 久久中文字幕一级| 99热6这里只有精品| 91九色精品人成在线观看| 日韩欧美国产一区二区入口| 十八禁网站免费在线| 嫁个100分男人电影在线观看| 久久伊人香网站| 午夜激情av网站| 国产91精品成人一区二区三区| 午夜福利在线观看吧| 男女那种视频在线观看| 啦啦啦韩国在线观看视频| 一本久久中文字幕| 国内久久婷婷六月综合欲色啪| 老司机在亚洲福利影院| 久久久久久久久久黄片| 亚洲欧美日韩高清专用| 99国产精品99久久久久| 日韩精品青青久久久久久| 国产一区二区三区在线臀色熟女| 又爽又黄无遮挡网站| 亚洲第一电影网av| 欧美色视频一区免费| 国产激情久久老熟女| 欧美绝顶高潮抽搐喷水| 97碰自拍视频| 欧美日韩亚洲综合一区二区三区_| 亚洲一码二码三码区别大吗| 国产单亲对白刺激| 人妻丰满熟妇av一区二区三区| 国产精品99久久99久久久不卡| 一进一出好大好爽视频| 99在线视频只有这里精品首页| 精品久久久久久久人妻蜜臀av| 无遮挡黄片免费观看| 一个人免费在线观看的高清视频| 久久香蕉精品热| 丁香欧美五月| 久久中文字幕人妻熟女| 亚洲男人的天堂狠狠| 免费无遮挡裸体视频| 亚洲国产看品久久| 麻豆av在线久日| 国产在线精品亚洲第一网站| 12—13女人毛片做爰片一| 免费观看人在逋| 国产片内射在线| 亚洲精品中文字幕在线视频| 最近视频中文字幕2019在线8| 久久久久国产一级毛片高清牌| 小说图片视频综合网站| 亚洲精品粉嫩美女一区| bbb黄色大片| 日本熟妇午夜| 波多野结衣高清作品| 精品日产1卡2卡| 999精品在线视频| 国产精品av久久久久免费| 嫁个100分男人电影在线观看| 亚洲美女视频黄频| 久久久久九九精品影院| 国产激情偷乱视频一区二区| 日韩 欧美 亚洲 中文字幕| 日韩大码丰满熟妇| 国产麻豆成人av免费视频| www国产在线视频色| 成人永久免费在线观看视频| 天天躁夜夜躁狠狠躁躁| 日韩中文字幕欧美一区二区| 99在线人妻在线中文字幕| 非洲黑人性xxxx精品又粗又长| 不卡av一区二区三区| 亚洲男人天堂网一区| 欧美日韩一级在线毛片| 91九色精品人成在线观看| 巨乳人妻的诱惑在线观看| 制服诱惑二区| 欧美日韩国产亚洲二区| 日韩欧美一区二区三区在线观看| 别揉我奶头~嗯~啊~动态视频| 亚洲国产欧美人成| 免费观看人在逋| 日本一本二区三区精品| 美女午夜性视频免费| 1024视频免费在线观看| 真人做人爱边吃奶动态| 精品久久久久久,| 欧美成人一区二区免费高清观看 | av片东京热男人的天堂| 色综合婷婷激情| 欧美色视频一区免费| 高潮久久久久久久久久久不卡| 日韩欧美国产一区二区入口| 51午夜福利影视在线观看| 99热这里只有精品一区 | 三级毛片av免费| 国语自产精品视频在线第100页| 国产一区在线观看成人免费| 亚洲aⅴ乱码一区二区在线播放 | 黄色视频不卡| 午夜成年电影在线免费观看| 欧美又色又爽又黄视频| 久久久久国内视频| 亚洲一区二区三区不卡视频| 十八禁网站免费在线| 99热这里只有精品一区 | 国产精品永久免费网站| x7x7x7水蜜桃| 99热这里只有是精品50| 久久国产乱子伦精品免费另类| 青草久久国产| 亚洲国产精品合色在线| 国产精品爽爽va在线观看网站| 又爽又黄无遮挡网站| 日本黄色视频三级网站网址| 国产成人精品久久二区二区91| 亚洲国产日韩欧美精品在线观看 | 国产精品野战在线观看| 国产三级在线视频| 久久久久九九精品影院| 亚洲专区字幕在线| 国产黄a三级三级三级人| 正在播放国产对白刺激| 日韩成人在线观看一区二区三区| 母亲3免费完整高清在线观看| 亚洲av中文字字幕乱码综合| 99久久精品热视频| 中文字幕精品亚洲无线码一区| 日韩中文字幕欧美一区二区| 亚洲欧美一区二区三区黑人| 日本一本二区三区精品| 一a级毛片在线观看| 欧美色欧美亚洲另类二区| 亚洲中文字幕一区二区三区有码在线看 | 国产亚洲精品av在线| 久久精品夜夜夜夜夜久久蜜豆 | 国产精品野战在线观看| 久久精品aⅴ一区二区三区四区| 99久久国产精品久久久| 欧美中文综合在线视频| 色精品久久人妻99蜜桃| 亚洲人成网站高清观看| 亚洲国产精品sss在线观看| 亚洲精品久久国产高清桃花| 亚洲一区二区三区不卡视频| 男女床上黄色一级片免费看| 欧美极品一区二区三区四区| 日韩免费av在线播放| 99国产精品一区二区蜜桃av| 国产精品野战在线观看| 国产不卡一卡二| 日韩精品免费视频一区二区三区| 亚洲成人国产一区在线观看| 在线看三级毛片| 麻豆成人av在线观看| 欧美性猛交╳xxx乱大交人| 日本一本二区三区精品| 亚洲自拍偷在线| 真人一进一出gif抽搐免费| 国内精品一区二区在线观看| 免费高清视频大片| 精品无人区乱码1区二区| 久久香蕉精品热| 国产黄色小视频在线观看| 丁香六月欧美| 久久亚洲精品不卡| 成年女人毛片免费观看观看9| 伊人久久大香线蕉亚洲五| 午夜精品在线福利| 亚洲午夜精品一区,二区,三区| 亚洲无线在线观看| 日本精品一区二区三区蜜桃| 色av中文字幕| 亚洲乱码一区二区免费版| 欧美成狂野欧美在线观看| 亚洲色图av天堂| 舔av片在线| 亚洲 欧美一区二区三区| 一进一出好大好爽视频| 熟女少妇亚洲综合色aaa.| 午夜免费成人在线视频| 久久九九热精品免费| 亚洲va日本ⅴa欧美va伊人久久| 久久久久久国产a免费观看| 一本精品99久久精品77| 成人国产综合亚洲| 亚洲一区中文字幕在线| 久久久久久久午夜电影| 九色国产91popny在线| av欧美777| 国产亚洲欧美98| 18禁观看日本| 亚洲欧美日韩高清专用| 国产精品一区二区精品视频观看| 舔av片在线| 看片在线看免费视频| 久久婷婷人人爽人人干人人爱| 性欧美人与动物交配| 久久久久久九九精品二区国产 | 国产精品一区二区三区四区久久| 成年免费大片在线观看| 99国产极品粉嫩在线观看| 俺也久久电影网| 成年版毛片免费区| 日日干狠狠操夜夜爽| 一区二区三区国产精品乱码| 国产蜜桃级精品一区二区三区| 欧美色欧美亚洲另类二区| 18禁观看日本| 久久天躁狠狠躁夜夜2o2o| 亚洲自偷自拍图片 自拍| 国产黄a三级三级三级人| 亚洲欧美日韩高清在线视频| 国产不卡一卡二| 亚洲免费av在线视频| 午夜视频精品福利| 一本一本综合久久| 亚洲电影在线观看av| 高清毛片免费观看视频网站| 亚洲精品一卡2卡三卡4卡5卡| 男女之事视频高清在线观看| 两个人看的免费小视频| 久久精品亚洲精品国产色婷小说| 19禁男女啪啪无遮挡网站| 欧美一级a爱片免费观看看 | 非洲黑人性xxxx精品又粗又长| 18禁观看日本| 欧美日韩一级在线毛片| 日本免费一区二区三区高清不卡| 欧美日韩福利视频一区二区| 每晚都被弄得嗷嗷叫到高潮| 91国产中文字幕| 一级黄色大片毛片| 搡老熟女国产l中国老女人| 久久伊人香网站| 国产主播在线观看一区二区| 欧美极品一区二区三区四区| 国产精品久久久av美女十八| 国产亚洲欧美98| www.自偷自拍.com| 国产成人av教育| 999久久久国产精品视频| 香蕉国产在线看| 不卡一级毛片| 禁无遮挡网站| 精品久久久久久久久久久久久| 99久久久亚洲精品蜜臀av| 在线观看免费日韩欧美大片| 又黄又粗又硬又大视频| 香蕉丝袜av| 成人高潮视频无遮挡免费网站| 亚洲aⅴ乱码一区二区在线播放 | 久久香蕉国产精品| 国产99久久九九免费精品| 日韩欧美一区二区三区在线观看| 国产精品 国内视频| 久久九九热精品免费| 国内精品一区二区在线观看| 两个人视频免费观看高清| 国产黄片美女视频| 国产成人精品久久二区二区91| 亚洲午夜精品一区,二区,三区| 成人午夜高清在线视频| 超碰成人久久| 18禁美女被吸乳视频| 成人国产综合亚洲| 18禁美女被吸乳视频| 免费在线观看视频国产中文字幕亚洲| 免费看美女性在线毛片视频| 天堂av国产一区二区熟女人妻 | 又爽又黄无遮挡网站| 欧美日本亚洲视频在线播放| 女人被狂操c到高潮| www.自偷自拍.com| 很黄的视频免费| 女人高潮潮喷娇喘18禁视频| 国内精品久久久久久久电影| 97碰自拍视频| xxxwww97欧美| 国产在线观看jvid| 久久久久国产一级毛片高清牌| 日韩成人在线观看一区二区三区| 97超级碰碰碰精品色视频在线观看| 精品久久久久久,| 可以免费在线观看a视频的电影网站| 午夜免费成人在线视频| 日本撒尿小便嘘嘘汇集6| 亚洲真实伦在线观看| 国产亚洲欧美98| 中文字幕人成人乱码亚洲影| √禁漫天堂资源中文www| 他把我摸到了高潮在线观看| 亚洲第一电影网av| 99热只有精品国产| 最近最新中文字幕大全电影3| 18禁观看日本| 日本 av在线| 91麻豆av在线| 草草在线视频免费看| 久久人妻av系列| 国产精品日韩av在线免费观看| 搡老熟女国产l中国老女人| 国产成人精品无人区| 在线观看免费视频日本深夜| 午夜精品久久久久久毛片777| 天天躁夜夜躁狠狠躁躁| 亚洲成a人片在线一区二区| 亚洲成人精品中文字幕电影| 免费看日本二区| 亚洲美女视频黄频| 两个人看的免费小视频| 长腿黑丝高跟| 日韩高清综合在线| 成年人黄色毛片网站| 亚洲色图av天堂| 午夜福利视频1000在线观看| 亚洲 国产 在线| 色综合欧美亚洲国产小说| 国产私拍福利视频在线观看| 脱女人内裤的视频| or卡值多少钱| 欧美精品啪啪一区二区三区| 一区福利在线观看| 中文在线观看免费www的网站 | 免费电影在线观看免费观看| x7x7x7水蜜桃| 无遮挡黄片免费观看| 欧美极品一区二区三区四区| 色综合亚洲欧美另类图片| 看免费av毛片| av在线天堂中文字幕| 欧美日韩精品网址| 一本久久中文字幕| 99久久久亚洲精品蜜臀av| 亚洲欧美激情综合另类| 巨乳人妻的诱惑在线观看| 日本在线视频免费播放| 亚洲国产欧洲综合997久久,| 亚洲中文日韩欧美视频| 成人av一区二区三区在线看| 欧美午夜高清在线| 国产精品一区二区免费欧美| 色综合站精品国产|