Data-driven Fault Diagnosis For Planetary Gearboxes

Planetary gearboxes can achieve a large torque/speed ratio in a compact package,and thus have been widely used in engineering machinery, wind turbines, helicopters etc.A planetary gearbox is a gear system consisting of multiple planet gears, a sun gear, aring gear, and a carrier. The planet gears not only rotate themselves but also revolveabout the sun gear. Because of the complicated physical structure, fault diagnosis forplanetary gearboxes has the following characteristics and challenges:(1) the vibrationsignals collected by sensors are non-stationary and may include many unknowncomponents;(2) simultaneous multiple gear meshes can result in sympathetic vibrationsor supressed vibrations due to faults; and (3) transmission paths of the vibrations arecomplex. Therefore, fault diagnosis technologies for fixed-shaft gearboxes generally donot work well for planetary gearboxes. Considering that planetary gearboxes play animportant role in rotating machinery, it is worth to study fault diagnosis for planetarygearboxes so that the system reliability can be increased and maintenance cost can besignificantly reduced.It is difficult to develop an accurate model for a planetary gearbox because of itscomplex structure. All reported models are based on many stringent assumptions andcannot cover many important details of the system. These simplified models cannotreflect field responses of faulty planetary gearboxes because many parameters areignored in the models. Data-driven methods do not require any physical models. Theyidentify health conditions of a system by comparing field data with experimental orhistorical data. Data-driven methods usually include stages such as experiment design,data collection, data cleaning, feature extraction, dimension reduction, and modeltraining. This thesis aims to investigate how to improve the performance of data-drivenmethods for fault diagnosis of planetary gearboxes. The contributions of the thesis arelisted below.1. Reported dimension reduction technologies have been investigated extensively,and three methods of dimension reduction have been proposed from unique perspectives.It further demonstrates that dimension reduction is important and necessary in data-driven methods.(1) A feature ranking criterion has been proposed for kernelized classificationalgorithms. Cosine similarity is the used measure in the kernel space, and the featureeffectiveness is estimated by class separability. Experimental results show that theproposed criterion can recognize important fault-sensitive features, and feature selectioncan be implemented in shorter CPU time by aid of this criterion.(2) A multi-criterion fusion feature selection method has been proposed toeliminate the limitations of using single criterion. The proposed method evaluates afeature from three aspects: effectiveness, correlation, and classification performance.The effectiveness criterion is used to estimate the generalization ability of a feature, thecorrelation criterion is used to reduce redundancy of a feature subset, and the criterionof classification performance is used to guarantee that a feature subset can obtain highclassification accuracy. Experimental results demonstrate that the proposedmulti-criterion fusion method can provide more comprehensive evaluation for a featureand reach higher classification accuracy than a single criterion.(3) Feature selection and feature extraction are two unique approaches fordimension reduction. A hybrid feature selection and feature extraction method has beenproposed to combine the advantages of the two types of dimension reductiontechnologies. Feature selection can reduce the input dimension of the entire system, andfeature extraction can reduce the input dimension of classifiers. In addition, theproposed method is kernelized and thus can recognize non-linear features. Experimentalresults show that the hybrid method does inherit the advantages of both feature selectionand feature extraction. That is, the hybrid method can achieve bigger reduction ratio andgenerate more informative features than the independent use of feature selection andfeature extraction.2. The topic of parameter selection of the Gaussian radial basis faction in SVM thatis a widely-used classifier in data-driven fault diagnosis has been investigated. Theparameters of SVM are crucial to its performance. In this study, an objective functionhas been established with respect to the kernel parameter in Gaussian radial basisfunction, and then obtained an analytical solution for the parameter. The proposedmethod contains no optimization search algorithms and thus has a great advantage inreducing algorithm complexity and CPU time. Experimental results show that the proposed method is fast and effective. Data-driven fault diagnosis methods usuallyrequire online learning and a real-time training. Therefore, the proposed method isextremely important to the online learning process.3. Semi-supervised learning to data-driven fault diagnosis has been introduced. Theperformance of data-driven methods deteriorates if the number of features is larger thanthe number of training instances. In other words, the classification accuracy decreases,the model complexity and the CPU time both increase. In earlier part of this PhD work,dimension reduction has already been utilized to improve the performance ofdata-driven methods. It here aims to improve their performance from another viewpoint,i.e. increasing the number of training instances by semi-supervised learning. Anon-parametric semi-supervised method that determines the most confident predictioninstance by instance ranking has been proposed. The most confident predictioninstances are obtained by semi-supervised learning, i.e. the k-nearest neighbor classifier.The size of the training set increases accordingly, and the performance could thus beimproved. Experimental results demonstrate that the proposed method can achieve ahigher accuracy than the conventional approach when the initial training set is small.