Symbolic aggregate approximation is an important way to extract time series features. However, traditional symbol aggregation approximation methods have many defects, such as the averaged segment number, equal segmentation interval, and mutation information that cannot accurately reflect the non-stationary sequence. Aiming at the abovementioned problems, we establish a new time series SAX model by introducing the local mean decomposition and a segmentation algorithm for improving wavelet entropy. The basic principle of the model is to denoise the original sequences by the local mean decomposition technique, obtain the segmentation number by the sliding window threshold method, and use SAX notation to represent the symbols, and the KNN classifier is used to realize classification performance test. Based on this improved model, we carry out an empirical test. The results show that the algorithm can effectively extract the information features of the sequences, has a high fitness degree, and achieves the purpose of dimensionality reduction. And more importantly, it improves the classification accuracy of the KNN classification algorithm in SAX notation.