Abstract
We present a concept for increasing the reliability of distance-based pattern classification models, applied to time-variant processes. In fault detection, static pattern classification is often used to differentiate between normal sensor data indicating ordinary uncritical states for the monitored process and abnormal data, indicating states that need to be resolved by maintenance or even dangerous states, that require the process to be switched off. Depending on the process and its environment, it can be difficult to classify, properly, each state the process can be in by employing sensor data. Sensor hardware limitations may also cause uncertainty when gathering the data for different process states. Inspired by conventional linear control theory, we propose a Proportional-Integral-Derivative (PID) concept which allows the determination of a relevance factor for a current classification, by consulting previous data and expert knowledge. It is based on the assumption that features of sensor data have different alteration rates, when transitioning into different states. The PID concept can be applied to several classical distance-based pattern recognition approaches such as Kohonen's Learning Vector Quantization (LVQ) or the k-means algorithm and allows them to classify abnormal data more reliably. The performance of this PID concept is demonstrated with test data sets. Problems that might occur when applying the PID concept to real applications and their solution are pointed out.