Using temporal information in input features of neural networks

In many applications neural networks use temporal information, i.e. any kind of information related to a time series. Temporal information can be either represented within a network using a dynamic network paradigm or embodied in the input features of a network. The paper presents two methods for an explicit use of temporal information in input features. A least-squares approximation of signals with orthogonal polynomials is used to infer information about trends in a signal (average, increase, curvature, etc.). Input information about the length of a time series up to a certain point in time may act as a decreasing threshold making the network more and more sensible to changes in other input features. The advantages of the two methods are demonstrated by means of a real-world application example, tool wear monitoring in turning.