[20pt,42pt] Implementation, training and model validation using the approach from Sidhant Gupta

[ThomasKittler]

SP Spent[42]:

Tasks:

• Implement "calculate_background" function #64
• Implement database with average power consumption per known appliance #68
• Implement "subtract_emi_background" function #70
• Implement "event detection" function #69
• implement feature calculation functions #71
• Implement gupta model (outer function / wrapper) #62
• Add labels to training data #63
• Fit kNN classifier to labeled training data #67
• Hyperparameter optimization #60
• pack trained model incl. feature calculation into ONNX container (to be used by inference tool) #72

References:

Sidhant Gupta, "ElectriSense: Single-Point Sensing using EMI for Electrical Event
Detection and Classification [..]",PhD thesis, U-Washington, 2014

[mariahirsch]

Labelling training data more difficult than anticipated: there is no threshold value for peaks in Apparent Power Spectra that works as a switch detection for all appliances. Even the sum of the individual spectrum of a single time step is not comparable, although switches are clearly to be seen.

This are the sums of a time series of switches of a blow dryer:

And the sums of a time series of switches of an LED bulb:

Note the difference on the y axis.

[mariahirsch]

New approach: norm spectra before switch detection via sum of spectrum. Examples:

LED bulb:

or

blow dryer:

For switch detection: Once any switch event is detected for a normed, background subtracted spectrum, the detection of the switch direction (on or off) is preferrably done with the <switching_offset>th next spectrum. This approach is more robust against peaks resulting from mechanical switches.

[mariahirsch]

Hyperparameter tuning: 70% of training data used for fitting the classifier, 30% for validation.

Parameters tested: n_neighbors, feature scaling methods, number of peaks in power spectrum taken into account

Best results so far:

• for all appliances (including all similar appliances, e.g. two blow dryers etc):

MaxAbsScaler with n_neighbors of 1 and the 3 most prominent peaks in the power spectrum leads to a validation accuracy score of 0.808

• for only different kinds of appliances (both blow dryers map to the same output class, etc), using same number of peaks and scaling method as above:

• for different kinds of appliances using the 9 most prominent peaks in the power spectrum leads to an accuracy score of 0.992: