Outputs of real-world processes can be considered as signals. Describing these signals in the form of signal models or times series is a well-known problem nowadays. In these cases by time series, we mean a sequence of data points, measured typically at successive time instants spaced at uniform time intervals. The importance of modeling time series has several reasons. Among them, these models are useful for simulating the source of signals and making a good theoretical basis for signal processing systems. The used model for these systems can be either deterministic model or statistical one. In this thesis, for modeling , training and classifying time series a statistical model named Hidden Markov Model (HMM) is used. Having a strong mathematical basis and its great and varied applications especially in the field of voice recognition were some reasons for selecting HMMs. On the other hand, combining models and classifiers in order to improve the performance and lower the error rate is used widely for a while in hard probelms. In these methods, several weak classifiers are combined with each other to construct a strong good classifier. If the basic classifiers have enough difference and performance of each one is higher than a random model, the obtained combined model expected to have higher performance in comparision to others. Though many researchers investigated classifier combinination in their papers, combining classifiers with the aim of obtaining the final classifier by means of internal structure of all basic classifiers which are particularly used for time series is considered less yet. Pervious methods wich use classifier combination techniques, do their combination in the final, decision level or are not specialized for time series structures. Clustering is also a classification procedure in which the used data don’t have any training labeled data and the purpose is to detect and classify similar data in the same group and the optimum case is when the internal distance among the data of the same clusters are minimum and the extra-distance between different clusters is as much as possible. Due to not having label for data in clustering, this problem is more difficult than classification. In this thesis the first attempt was to obtain a good model for time series by using Hidden Markov Models which can best describe the given data and achieve a good performance and superiority compared to other considered methods. Second, by using this training method and separating the models of different classes from each other, a new technique for time series classification by the means of Hidden Markov Models is proposed. The third problem was clustering. For this problem again by means of Hidden Markov Models and involving measures from pervious researches a new time series clustering method is proposed which take advantage of combining basic models and revealed better performance than other compared methods in the experiments. Keywords: Time series, Hidden Markov Model, classification, training, clustering