Time sequence data mining using time-frequency analysis and soft computing techniques

This paper presents a new approach for time series data mining and knowledge discovery. The relevant features of non-stationary time series data from power network disturbances are extracted using a multiresolution S-transform which can be treated either as a phase corrected wavelet transform or a variable window short-time Fourier transform. After extracting the relevant features from the time series data, an integrated LVQ neural network and various feed-forward neural network architectures are used for pattern recognition of disturbance waveform data. The fuzzy MLP outperforms all the other different connectionist models and is used in the final stage for encoding knowledge in the connection weights that are used to generate rules for fuzzy inferencing of the disturbance patterns. Overall pattern classification accuracy of 99% is achieved for power signal time series data. The knowledge discovery from the data has then been presented for selected patterns using the new quantification procedures. The approach presented in this paper is a general one and can be applied to any time series data sequence for mining for similarities in the data.