Time series problem involves several fields, such as economy, meteorology, waterconservancy, forestry etc. At present, time series data mining has been the focus of datamining, which has strong theoretical and practical significance. Because of some complexcharacteristics of time series data, for example, time variation, high dimension, noise jammingand volatility, time series data mining is always one of the difficulties in data mining research.Process neural network (PNN) is a development of traditional neural network in the timedomain. Process input of PNN relax synchronization instantaneous limit on inputs intraditional neural network models, and is more general artificial neural network model. It hasits own unique advantages in dealing with time related problems. In this thesis, process neuralnetworks is introduced into time series data mining to deeply studying on clustering,classification and prediction problems, combining with wavelet multi-resolution analysis andrelated technologies.Firstly, a time series clustering method is proposed base on wavelet and improvedself-organization process neural network for time series clustering problem. Original timeseries data is decomposed by wavelet. Under the principle of reserving clusteringcharacteristics, the signal is reconstructed. And then reconstructed signal which has beenfitted into time-varying functions is used as the input of process neural network.Self-organization PNN is trained by improved competition algorithm. Making use oftime-varying input characteristic of PNN, the timing signal characteristics processed bywavelet is considered adequately in clustering analysis. Network extracts implicit processmode characteristics of input function to self-organize. The improved competition learningalgorithm is given. Finally, clustering result of UCI datasets shows that the proposed approachcan be applied to timing clustering effectively.Secondly, a time series classifier is proposed based on competitive radial basis processneural network for time series classification. First, the topology of competitive radial basisprocess neural network is given. The compound competition process neuron hidden layer isadded to network. Discrete data in continues time points are fitted to time-varied functions asnetwork input. The classifier breaks through time series data unequal length restriction.Pattern matching and temporal aggregation operation of time-varied input is achieved bycompetition process neuron units. The linear connection weights calculation in the outputlayer is omitted to simplify network structure and training process. Then generalization ability of classifier is improved by using different clustering coefficient for each clustering sizes andnetwork training results. Learning algorithm is given. Finally performance and effectivenessof classifier are proved by multivariable time series classification simulation data.Thirdly, two process neural network models are proposed for time series prediction. Oneis genetic process neural network model. After input data are represented as a set oforthogonal basis expansions, using improved genetic algorithm to optimize genetic processneural network training process by introducing immigration operator. Prediction accuracy andgeneralization performance of the model is verified by per capita GDP prediction problem andcontrast analysis. The second model is an improved feed forward process neural network,giving combined process neural network algorithm and applying to CPI prediction. The resultshows that this modelâ€™s prediction accuracy is obviously higher than traditional neuralnetwork model.Lastly, a time series control and system identification method based on PNN is proposedfor control problem in time series context.First, based on time-varying feature and complexnonlinear characteristic of time series control problem, double hidden layer process neuralnetwork is introduced into time series control and system identification. Nonlinear andtime-varying characteristics are taken into the consideration of time series control for theadvantage of processing time-varying problem using PNN, and analyzed the model advantageused in time series control.Then, two wood drying control models are built by PNN trained byimproved learning algorithm for the characteristics of wood drying process control. Twomodelsâ€™ analysis results compared to traditional neural network models show that processneural network model has better prediction control precision and generalization ability, theperformance is superior to the traditional neural network model, PNN applied in time seriescontrol and nonlinear system identification problem is feasible.This thesis is mainly carried on some problems of tine series data mining, which involvesin some methods and models of clustering, classification and prediction. The performance andeffectiveness of these methods and models are verified by experiments, example validationsand comparative analysis. The research on time series clustering and time series classifierprovides a good theoretical basis and ideas for further time series data mining research. |