We have learned by experience about some of the key problems involved with the design cycle of a Pattern Recognition system. We acknowledged the importance of having a good data set to work with. The combined large dimensionality of the data, and small number of data points to work with, made it difficult for us to run machine supervised feature extractions like FLD and PCA, and pushed us in the direction of building hand-crafted features.
When we tried several feature extraction and classification methods, we obtained 85% accuracy by running K-nn clustering on our hand-crafted features. It is interesting that most of the misclassified data has waveforms that match other classes. Essentially, data labeled as class A would have a waveform that belonged to class B, thus making it impossible to correctly discriminate the data, since our features are based in these waveforms. This brings up the question of whether we should label the devices by looking at their commonsense functionality, or their underlying structure in terms of electrical engineering.
We think it is possible to improve the performance of our methods by trying them on a larger data set, and by taking advantage of the Plug's other sensors. Data retrieved by the light, vibration, and sound sensors is likely to be correlated with the functionality of a device connected to the Plug. Any of this new dimensions can be a good candidate for new, better features for classification.