Early detection of gradual concept drifts by text categorization and Support Vector Machine techniques: The TRIO algorithm

Abstract

During the normal operation of complex and risky industrial plants such as the nuclear or the aerospace ones, the safety heavily rests upon the capability of the diagnostic systems of detecting concept drifts which might imply incipient failures. In this paper we propound the TRIO algorithm for the online detection of signal drifts: the underlying idea is that a real signal may be categorized as correct or drifting by comparison with added sets of artificial signals known to be correct or drifted. More specifically, the TRIO algorithm is based on three performers, namely (i) a training set of artificial signals, (ii) the Text Categorization (TC) technique and (iii) the Support Vector Machine (SVM) technique. Initially, we construct an artificial training set constituted by one “correct” set of signals, embraced by two “suspect” sets of signals, the suspect-up and the suspect-down drifting signals. These signals are transformed in points within the signal space by the TC technique; then the SVM technique is applied for isolating the regions occupied by the suspect-up and by the suspect-down points. At this point the “artificial context” has been established and the real measurements come in. By resorting to the sliding window technique, at each epoch the actually measured data segment is analogously transformed into a point within the signal space and then declared correct or suspect (drifted) according to the region where it falls. In the latter case suitable actions must be taken by the plant operators. Numerical case-studies and a comparison with literature results are presented.