Search

Aiming at the problems of low accuracy and poor generalization ability of bearing fault diagnosis based on machine learning (ML), this paper proposes the method of deep learning (DL) to solve the above problems. Combined with the bearing signal one-dimensional characteristics, this paper proposes a one-dimension deep convolutional neural network(1D-DCNN). First of all, the original bearing vibration signal is directly input the 1D-DCNN frame structure, then 1D-DCNN frame structure is used to automatic feature extraction. Next, we use softmax regression to classified fault samples and normal samples, the confusion matrix shows that the accuracy of the 1D-DCNN model for fault diagnosis. Finally, T-SNE algorithm is used to reduce the dimension of extracted features and visualize data features, which proves that the 1D-DCNN model has a good feature extraction ability. We use bearing data sets from Case Western Reserve University (CWRU) to verify the model of 1D-DCNN. According to the result, we can see that the 1D-DCNN frame structure not only effectively extracts and diagnoses the original signal, but also has high fault identification accuracy. It shows the advantages of the 1D-DCNN model in extracting data features and fault diagnosis. The model of 1D-DCNN is better than the mainstream fault diagnosis method of support vector machine (SVM) and probabilistic neural networks (PNN).

In this paper we propose a new fault diagnosis model based on sparse semi-supervised GAN (SSGAN).The SSGAN-based fault diagnosis can use a large amount of unmarked data to improve the accuracy of the marked training part.Solved the problem that the general neural network requires a large amount of tag data.In particular, we improved the discriminator to get a more sparse network, which further improved the classification effect.At the same time we choose Leaky ReLU as the activation function which solve the problem that the ReLU activation function has a dead zone.Simulation studies on the Tennessee–Eastman (TE) benchmark process evaluate the performance of the developed method, which indicate that the SSGAN method performs better than BPNN.

A fault diagnosis method based on batch-normalization stacked sparse autoencoder (SSAE) is presented in this paper. This paper use the autoencoder to extract features for fault diagnosis on account of its good performance in feature extraction. In order to improve the accuracy of the extracted features, this paper use a sparse representation which is a constraint during the encoding process. The multi-layer structure of autoencoder has an internal covariate shift problem and the generalization ability of the network is critical, batch normalization is employed before the activation function in each layer of the autoencoder network. And a stacked method is utilized to optimize network structure and reduce training difficulty. So the features of the original signal are extracted by the network based on the above method and the extracted features are placed in the classifier to identify different health states. For the purpose of fault diagnosis, this paper uses the proposed method to experiment with the bearing data set provided by Case Western Reserve University (CWRU). The experiment proves that the proposed method has a better fault diagnosis performance compared with other traditional methods.

A process monitoring method based on robust slow feature analysis (RSFA) is presented in the paper. Slow feature analysis (SFA) is a new feature extraction technology which can find a group of slowly varying feature outputs from the high-dimensional inputs. In order to improve the robustness of SFA, RSFA is presented which applies locality preserving projection (LPP) to define the energy density function and set a reasonable threshold value to remove outliers mixed in the normal process data. This paper uses three spline interpolation method to interpolate the missing value caused by removing the outliers to maintain the integrity of the time series. For the purpose of fault detection, the D monitoring statistic index is adopted and its confidence limit is computed by kernel density estimation. Simulation on Tennessee Eastman (TE) benchmark process show that the proposed method has a better fault detection performance compared with the conventional SFA-based method.

Aiming at the problem on how to improve the diagnostic rate of rolling bearing fault diagnosis models, the original vibration signal is denoised and feature extracted by wavelet packet decomposition and reconstruction, and the fault pattern recognition is realized by Elman neural network. For the problem that Elman neural network has a slow convergence rate and it is easy to fall into the local optimal value, this paper uses the genetic algorithm (GA) with global search ability to optimize the weight and threshold of the Elman neural network through the steps of selection, crossover and mutation. It is proved by experiments that the Elman neural network optimized by genetic algorithm has high diagnostic precision, and this method can be better applied to fault diagnosis of rolling bearings.

Due to the complicated structure of most rotating machines, and the high dimension and strong correlation of process variables, this paper presents a multiple fault classification method of rotating machine based on Principal Component Analysis (PCA) and K Nearest Neighbor (KNN). Firstly, the dimensionality reduction model is established by using the normal historical data, which ensures a small memory footprint and computational consumption in fault classification. After combining different fault types data of rotating machine, PCA model is used to project them to low-dimension space. The KNN function is adopted to classify the training data after dimensionality reduction. The alternative data is utilized for online monitoring simulation by using SPE (Square Prediction Error) and T2 statistic, and then KNN training model is used to identify fault types from the abnormal data. The experiment result shows that the proposed method is accurate and effective in multiple faults identification.

According to the characteristics of the noise of discharging billet, method based on scale product using wavelet transform for denoising is introduced in this thesis. Then evaluation criteria based on edge detection which evaluates the quality of image denoising is proposed, according to the requirements of captured pictures in the thesis. With the position of slab changing in image, gray of edge is changing. So it is unsuitable to detect edge by the fixed threshold value. Therefore, in order to improve image contrast between slab edge and background, image enhancement based on wavelet transform is adopted. Then, the slab edge is detected by Canny operator. After that using Hough transform extract complete and smooth edges of slab. In image, the size of slab is known. Taking advantage of reconstructed image, the relationship of pixel and actual size is assured. At last we can calculate the real position of the slab in reheating furnace.

In the paper, we use wavelet technique to detect edges in small scale along the direction of gradient maximum. Edges that we extracted are accurate and single-pixel wide. But the photo also contains a lot of noise, so we set threshold to extract the ideal edge points. Currently, the threshold is set mostly by people's experience that needing a lot of trial or set the average gray value of the image directly, but the overall effect is not satisfactory. In response to the problem, we propose a method of using two-dimensional otsu model to obtain the threshold, the two-dimensional otsu method not only considers the gray value of pixels but also takes the pixels outside their fields of space-related information into account and it takes a good performance in the presence of noise of image. And we do not need to set any parameter to get the threshold. After that, we propose the corresponding solution to the problem that some edge points can not be detected: local enhancement method. In the method, we first operate the fuzzy edges of the original image, and then use the method we have proposed to detect the edges again. Finally, the simulation shows the correctness and effectiveness of the algorithm and it can also detect the fuzzy edges with an advantage of positioning accurate and having low noise.

As an essential step of 3D reconstruction, research on the camera calibration methods has great important significance of theoretical study and practical value. In this paper, a new simply, flexible and more accurate coplanar camera calibration method is proposed based on neural network. This method only requires a coplanar target and without camera motion. The neural network is used to learn the relationships between the image information and the 3D information to emend aberrance of camera, and it neither requires the inner and outer parameters of the camera and any prior knowledge of the parameters. The experimental results of image simulation show that the proposed method is correct and effective.

As a necessary step of 3D reconstruction, and a premise and base for computer vision obtaining the spatial information of 3D objects, research on the camera calibration methods has great important significance of theoretical study and practical value. This paper introduces a new simply, flexible and more accurate classic camera calibration method based on coplanar points. This method only requires a coplanar target and without camera motion. We use step-by-step calibration thinking, and first calibrate the principal point, and then neglect the lens distortion and linearly solve all the parameters, when high accuracy is required, we can include Weng's lens distortion model and solve the distortion coefficient by nonlinear algorithm. The experiment of images proves the proposed method is accurate and effective.

As a new method of multivariate statistic analysis, non-negative matrix factorization (NMF) has been proposed for dealing with non-negative data. The results of NMF are non-negative and can be interpreted and understood easily, so they have specific physical meaning. In the production process, most data are non-negative. Due to this situation, a method of fault diagnosis based on NMF and SVM is presented in this paper. An on-line process monitoring model is built through NMF, and multi-fault classifiers are trained by SVM. The faults which are monitored by on-line monitoring model will be confirmed and identified by the fault classifiers. The simulation results of three water tank system show the effectiveness of this method.

The paper researchs the BSS application in audio separation. It gives a detailed analysis of audio mixtures separation, including a brief introduction of genetic algorithm and proposes an optimized analysis model for the audio separation. The analysis model avoids shortcomings of the traditional analysis which do not have a good as-tringency and the astringency influenced by step-length.

The fault detection and diagnosis methods based on principal component analysis (PCA) have been developed widely because they need no detailed information about the process mechanism model and really can detect faults promptly. However the existing diagnosis algorithms such as expert systems or contribution plots, etc. still have some trouble when they are applied in real industrial processes, which leads to more extensive research on this topic. In this paper, the proposed diagnosis method utilizes the on-line loading plot and cluster analysis to give accurate cause for abnormal process conditions, which is grounded on the fact that faults normally change the correlation of process variables which may indicate more direct information about the failure cause. Thus, the principal components score plot and square predicted error (SPE) plot are used to detect the abnormal process operation condition, the loading plot and cluster analysis are used to diagnose the faults. The result shows that accurate conclusion could be obtained easily by this method.