Real-Time Event Detection in Time-Series Classification Based on Amplitude Rejection

Classification methods are widely used in several types of applications and a lot of research works report highly accurate results on their ability to predict in unseen data. However, results are usually based on strong assumptions related to data preprocessing that might not hold in real world applications. The training set in practice can significantly differ to that of testing, especially when the classification process is carried out in real-time and the required preprocessing is not applicable without prior knowledge on the testing signals such as its length and amplitude. Sampling methods like sliding or additive window are usually employed, but not always resolve the problem that in many cases results in false positives. This work proposes an algorithm for real-time classification of signals with unknown length, based on a feature transformation that enables the classifier only when the signal's amplitude is within the expected event range. The proposed transformation can be used to generalize a classifier in similar data by only requiring knowledge of the expected event amplitude. The real-time performance of the proposed algorithm is evaluated in two industrial processes and its generalization ability in two novel (a synthetic and an industrial) data sets.