Adaptive Multi-Scale Wavelet Neural Network for Time Series Classification

Wavelet transform is a well-known multi-resolution tool to analyze the time series in time-frequency domain. basis diverse but predefined by manual without taking data into consideration. Hence, it great challenge select an appropriate wavelet separate low and high frequency components for task on hand. Inspired lifting scheme second-generation wavelet, updater predictor are learned directly from of series. An adaptive multi-scale neural network (AMSW-NN) proposed classification this paper. First, candidate decompositions obtained convolutional conjunction with depthwise network. Then, selector used choose optimal decomposition candidates. At last, fed predict label. A comprehensive experiment performed UCR archive. The results demonstrate that, compared classical transform, AMSW-NN could improve performance based different networks.