Data-driven Intelligent Fault Recognition Method For Rotating Machinery

Rotating machinery plays a pivotal and irreplaceable role in modern industrial production.Therefore,the condition monitoring and fault diagnosis technology to ensure the safe and reliable operation of rotating machinery has been rapidly developed in recent years.For the overall development trend of the technology,the basic consensus has been reached.The consensus is that the fault diagnosis must take the path of intelligent decision-making technology of industrial big data and guide by the principle of data science.The development goal of this technology should be to make rotating machinery operate with high quality and efficiency as soon as possible,which can meet the needs of smart manufacturing.One of the basic problems induced by this is to solve the problem of data resource protection,development,and utilization for rotating machinery.For the above reasons,rotating machinery is taken as the study object in this dissertation,and modern machine learning techniques is as the theoretical basis.Moreover,data-driven intelligent fault identification methods for rotating machinery based on signal processing,rough sets,information fusion,semi-supervised learning,domain adaptation,neural networks,deep learning and other intelligent technologies are investigated.The obtained research results and conclusions are as follows.(1)In order to solve the problem of sensitive fault feature extraction for rotating machinery,a method based on empirical wavelet transform(EWT)and weighted multineighborhood rough sets(WMNRS)is proposed.Firstly,the nonliear and strongly noisy vibration signal is decomposed using EWT,and a group of optimal mode components are selected for reconstruction by correlation.Time-domain features of the reconstructed signal are calculated to construct the high-dimensional original feature set.Then,the frequent itemsets are obtained by attribute reduction of the original feature set using WMNRS under different neighborhood radius.Lastly,the probability of occurrence for each feature in the original feature set when attribute reduction is performed using multi-neighborhood rough set is counted,which is weighted with feature value as an easy to classify and sensitive feature set.The experimental results show that the proposed method can effectively extract the vibration signal features of rotating machinery and correctly recognize the fault types according to the extracted feature vectors.(2)Aiming at the problem that traditional shallow fusion models are weak in nonlinear mapping and feature representation for complex data,a fault recognition method using convolutional neural networks(CNN)to fuse multi-sensor signal features is proposed.Firstly,feature extraction is performed on multi-sensor vibration signals separately,and the obtained feature vectors are used as one-dimensional feature surfaces to construct a set of multi-sensor feature surfaces,which is used as the input of CNN.Then,adaptive hierarchical fusion and extraction of multichannel features are realized with the deep network structure.Last ly,recognition results are output by a softmax classifier.The experiment al results show that the method has a better fault recognition ability,good robustness,and self-adaptability.(3)To address the problem that the imbalanced sample number of each category in the fault data set results in low recognition accuracy for minority samples,a fault recognition method based on Min-Max Objective CNN(MMOCNN)is proposed.Firstly,representative features learned from vibration signals by alternating convolution and pooling operations of the CNN are mapped to the class space through the fully connected layers.Then,the MMO function of the features is constructed in the class space.Finally,the MMO function is integrated into the loss function of the CNN.During the model training,the overall classification error is considered to be the smallest,and the learned features are required to keep the smallest intra-class distance and the largest inter-class distance to achieve effective fault recognition for imbalanced data.The imbalanced bearing datasets are used to test the recognition effect of the proposed method and the traditional CNN.The results show that the proposed method can improve the recognition accuracy of minority samples by more than 20%.Other supporting experiments further confirm the effectiveness of the proposed method in the case of imbalanced data.(4)For the condition that labelled samples are difficult and expensive to obtain,but unlabelled samples can be collected more easily,a fault recognition method based on semi-supervised convolutional neural networks(SSCNN)is proposed.First of all,a 1-d CNN is applied to learn class space feature s and generate class probabilities of unlabelled samples,based on which a class probability maximum margin criterion(CPMMC)method is used to construct the loss function of unlabelled samples.Then,the constructed loss function which aims at maximising the inter-class distance of class space features and minimising the intra-class distance of class space features is integrated into the cross-entropy loss function of CNN,and the SSCNN is esta blished.At last,the model is applied to analyse the vibration signals collected from rolling bearings,and a new intelligent fault recognition method using SSCNN is proposed.Two datasets are employed to validate the effectiveness of the proposed methodo logy.The results show that the established SSCNN can effective ly utilize unlabelled samples to train the model and enhance its fault recognition performances.Through comparing with commonly used semi-supervised deep learning methods,the superiority of t he proposed method is validated.(5)Aiming at the problem of insufficient model recognition due to the difference in the distribution of training(source domain)and test(target domain)data under dynamic working conditions,a fault recognition method based on adversarial domain adaptation convolutional neural netwo rks(ADACNN)is proposed.Firstly,a 1-d CNN is used to map the source and target domain training samples to the class feature space,and the predicted label space simultaneously.Then,the adversarial domain adaption is constructed in the class feature space and the maximum mean discrepancy domain adaption is constructed in the prediction label space,and an ADACNN model is built.Finally,a fault recognition method based on the ADACNN model is proposed.The effectiveness of the proposed method is verified on two rolling bearing data sets of public and actual measurements.The experimental results show that the method has more than 4% advantage in fault recognition accuracy compared with CNN under variable working conditions(load,speed).In this dissertation,an in-depth study on intelligent fault identification methods for rotating machinery is performed from a data-driven perspective.The proposed methods help to solve the problems in intelligent fault identification such as lowdimensional sensitive fault feature extraction,multi-sensor information fusion,data imbalance,insufficient labelled samples,and variable operating conditions.Issues worthy of further study in the direction of data-driven intelligent fault recognition include the integration of multiple improved algorithms,the design of model structure parameters,and technical means of algorithm application,etc.Solving these problems will be able to provide the theoretical basis of data science for the practical application of industrial big data technology.