Research On Dynamic Neural Network Models And Algorithms For The Analysis Of Complex Time-varying Signals

Research on the pattern recognition of time-varying signals has always been an important and difficult issue in the field of intelligent computing.With the rapid development of the sensing and the Internet of Things technology,how to analysis the characteristics of time-varying signals accurately and fully utilize the time-frequency becomes increasingly important.Dynamic neural network plays an important role in signal processing because of its strong identification and learning ability,and the ability to approximate any non-linear input-output relationships.Aiming at the analyzing of complex time-varying signals,according to their distribution characteristics and structural properties,and the complexity,uncertainty,non-stationarity and with noise in various practice surroundings in various fields,this dissertation combines deep learning theories with the signal time-frequency analysis technologies to study several deep dynamic neural network models and algorithms for the analyzing of different types of signals.Therefore,the research of this dissertation has important theoretical significances and application values.The main contributions of this dissertation are as follows:(1)Aiming at the feature extraction and classification of the short-term time-varying signal which can reflect the information variation in a short period of time with variability,instability and with noise,this dissertation put forwarded a deep wavelet convolution recurrent neural network(DWRCNN)model and algorithms.The DWRCNN model can combine the wavelet transform mechanism,the acquisition ability of soft threshold method for signal characteristics,the learning and memory ability of the recurrent neural network for time series and the learning property and classification mechanism of convolutional neural network model for large-scale data sets.It has good pattern recognition applicability for the short-term,low-frequency,weak and unstable time-varying signals with noise.(2)Aiming at the analyzing of the continuous,uncertain,non-stationary,non-periodic and infinite signal-channel long-term time-varying signals,this dissertation put forwarded an integrated long-short term memory model(ILSTM).The ILSTM model adds fusion mechanisms of different scale features,which can consider both wide and local features of long-term signals.ILSTM has the ability of feature information association and long-term and short-term feature change memory,and can perform online real-time learning and prediction.(3)Aiming at the classification of the aperiodic multi-channel long-term time-varying signals with instability,noise,and with inaccurate,incomplete and other incomplete information,this dissertation put forwarded a Takagi-Sugeno process neural network model(TSPNN).The TSPNN model combines the processing ability of process neural network to time-varying signals with the T-S fuzzy classification mechanism,can realize the embedding of expert experience knowledge.It has an adaptive learning mechanism for sample set,is suitable for small sample set modeling.(4)Aiming at the periodic multi-channel long-term time-varying signals with fuzziness and randomness distribution characteristics,and multi-mode,uneven distribution of samples,this dissertation put forwarded a probability computing process neural network(PCPNN)model.PCPNN learning method integrates the time-varying information processing mechanism of process neural network and Bayesian decision rules.It combines dynamic time warping algorithm,C-means clustering and BP algorithm.It can integrate the feature knowledge of signal categories with fewer model parameters,and is suitable for the modeling and analyzing of small and unbalanced data sets.This dissertation presented a series of dynamic neural network models and methods to solve the above problems,which can extract the essential characteristics and conversion rules of the complex time-varying signals,and have good pertinence and applicability in mechanism.Experiments verified the feasibility and validity of the models in this dissertation.