Classifying acoustic data

I have acoustic data taken near a pipe's valve which is in 3 different orientations (allowing different amounts of water depending on the valve's orientation). My task is to classify the orientation of the valve given the acoustic data. Samples have been taken for up to a minute in each situation (sampling rate is 12 kilohertz). I am assuming that the obtained signal is stationary (since the data are collected after the valve was positioned in its orientation), and so looked at the spectrum of different parts of each signal (each part is 0.5 second). In general, I could see that the overall shape of the spectrum of each orientation is different than the other. However, this might change due to other factors (change of flow pressure, temperature, etc).

I used Matlab NN toolbox to make a feedforward NN with one hidden layer, and used just the single sided spectrum of different parts of the signal as input from different valve's orientation, with 3 outputs. Playing around with the sizes, I had an error rate between 9-12%.

Would you suggest a better way to tackle this problem? I am thinking that I could apply PCA to reduce the dimensionality. Are there any more important features that I should investigate? from my research on the internet, I saw wavelet transform is used such situations, but I am not sure about this. I am not an expert in acoustics, nor in dsp. Any help is appreciated.