A state classifier uses learning obtained from a plurality of training algorithms each adapted to differentiate between states of physical orientation of an object in response to input data from an tri-axial accelerometer. At least two of the training algorithms are trained using data from an accelerometer mounted at a non-ideal angle. The classifier is trained to distinguish between the desired states from data collected from an tri-axial accelerometer device mounted at a plurality of respective angles with respect to a optimal axis on the object, wherein the angles are in the range of −180 degrees to +180 degrees. The classifier may include a plurality of classifiers and a decision fusion module used to combine the decisions from the respective classifiers to ascertain a state.