Symbolic Time Series Analysis-Based Rolling Bearing Fault Diagnosis

As one of the key components in rotating machines, bearing’s reliability is critical to safe operation of the whole machine. Nowadays, with bearing’s working conditions becoming increasingly harsh, the probability of bearing failure is increasing. It is of great importance to put effort on bearing fault diagnosis. Symbolic time series analysis (STSA) is built upon the principles of symbolic dynamics and information theory. After decades of development, it has become an appealing signal processing technique for applications in various engineering fields. The purpose of this thesis is to investigate symbolic time series analysis-based bearing fault diagnosis approach, which focuses on the symbolic probabilistic finite-state automata (PFA)-based pattern classification. The main content of this thesis is as follows:1) It first presents a novel partitioning method called probability density space partitioning for symbol sequence generation. In this partitioning approach, a time series is divided into several equal-sized regions based on the probability density distribution and each region is represented by a symbol. To verify the effectiveness of the probability density space partitioning for symbolic time series generation, bearing test-to-failure experiments are conducted, and the results indicate that probability density-based method is more sensitive to detecting bearing anomaly than traditional partition based methods.2) A short time Fourier transform (STFT)-based probability finite state automation (PFSA) pattern extraction method is proposed. In this method, the STFT is applied to the original time series obtained from sensors. After converting the STFT coefficients to symbol images, the symbolic dynamics-based probabilistic finite state automation is used to capture the relevant information, embedded in the symbol images.3) K-nearest neighbor (KNN) is a memory-based classification method. With increase of the training set, the traditional KNN method requires intensive computational load and memory to store the training set. This thesis proposes a K-means cluster-based algorithm to accelerate KNN classification as well solving other shortcomings.4) Three groups of bearing signals are used to verify the effectiveness of the STSA-based bearing fault diagnosis approach. For comparison, experiments are also conducted using traditional KNN classifier. The results show that STSA-based approach is good for bearing fault diagnosis and the improved KNN classifier has shown better classification accuracy than the original one.