Time series pattern recognition

Hi, I have a time series pattern recognition problem , and am seeking advices to further improve the recognition accuracy. The problem is as follows:

-I got training data for five water end-use categories, namely shower, faucet, clotheswasher, dishwasher and toilet (there are about 20000 samples for each category, and each sample is presented as a time-series vector). The task is how to classify a new unknown sample into those existing categories correctly.

I have tried the following techniques and the achieved accuracies are presented below:

1. ANN - When using this technique, the input is 4 features extracted from each sample, including volume, duration, maximum flow rate and most frequent flow rate, and the output is 5 categories. I have trained the model with different number of neurons varying from 5 to 50 using feed-forward network, and the achieved accuracy is about 60%.

2. Dynamic Time Warping algorithm - accuracy is about 65% and is really time consuming.

3. Hidden Markov model - (100 states) - accuracy is about 80%.

As I am trying to improve the accuracy of this problem, could anyone please advise me

what other techniques to be used or any suggestion on the modification of the above techniques.

Any help would be appreciated.

Khoi,

Figure 1: Examples of 10 clotheswasher and 10 dishwasher samples extracted from the training dataset.

Figure 2: Unknown sample to be classified