Componential coding in the condition monitoring of electrical machines Part 1: Principles and illustrations using simulated typical faults

Abstract

This paper (Part 1) describes the principles of a novel unsupervised adaptive neural network anomaly detection technique, called componential coding, in the context of condition monitoring of electrical machines. N umerical examples are given to illustrate the technique' s capabilities. The companion paper (Part 2), which follows, assesses componential coding in its application to real data recorded from a known machine and an entirely unseen machine (a conventional induction motor and a novel transverse ux motor respectively). Componential coding is particularly suited to applications in which no machine-speci c tailored techniques have been developed or in which no previous monitoring experience is available. This is because componential coding is an unsupervised technique that derives the features of the data during training, and so requires neither labelling of known faults nor pre-processing to enhance known fault characteristics. Componential coding offers advantages over more familiar unsupervised data processing techniques such as principal component analysis. In addition, componential coding may be implemented in a computationally ef cient manner by exploiting the periodic convolution theorem. Periodic convolution also gives the algorithm the advantage of time invariance; i.e. it will work equally well even if the input data signal is offset by arbitrary displacements in time. This means that there is no need to synchronize the input data signal with respect to reference points or to determine the absolute angular position of a rotating part.

Componential coding in the condition monitoring of electrical machines Part 2: Application to a conventional machine and a novel machine

“…H owever, when componential coding is applied to data recorded under faulty operation, the AD I values for each con guration are clearly much larger. This indicates that both the ICAN and JCAN are capable of separating the anomalous case 1 from the baseline [23], an anomaly is any characteristic of an unseen data-set that is different from the charact eristics of the training data-set. A fault can be thought of as a special case of an anomaly in which the anomalous characteristic happens to arise as the result of a real fault in the engineering system being monitored.…”

“…Section 2.1 of Part 1 of the paper [23] describes how componential coding may be implemented through either of two network variants. In summary, the Joint Channel Architecture N etwork (JCAN ) implementation can detect anomalous * correlations (including phase relationships) between simultaneously applied data channels.…”

“…Both the JCAN and the ICAN were studied for the separation of the two degrees of severity of voltage imbalance from the baseline condition. Section 1.2 of Part 1 [23] discusses how componential coding modelling is achieved by deriving basis vectors through the adaptive training process. The dimension of the basis vectors (i.e.…”

“…As cited in section 1 of Part 1 of this paper [23], these include time-frequency analysis, timescale analysis, higher-order statistical analysis and independent component analysis.…”

“…As explained in Part 1 [23], componential coding models data in an adaptive way and derives signal features automatically using a computationally ef cient algorithm. It therefore has the potential to perform accurate fault detection and is applicable to monitoring novel machines.…”