Logical analysis of maintenance and performance data of physical assets, ID34

Abstract

We present a novel approach to fault detection and Physical Asset Health Management (PAHM) called Logical Analysis of Data (LAD). LAD is a supervised learning, artificial intelligence, data mining technique that possesses distinctive advantages which proved to be of use in PAHM. This approach has been introduced by a group of researchers at Rutgers University in the USA, in the medical field, and it was adapted and used in the field of PAHM for the first time by our group of researchers. Unlike the traditional approaches that are either based on mathematical and statistical models or neural network modelling, this approach is based only on the advancement in the field of computer science, namely the speed of computation and the volume of data that can be processed, and proceeds in a logical manner like the human brain. That is why its potential for future applications seems promising. LAD does not assume that the data belong to a specific statistical distribution and therefore does not require statistical analysis of data prior to or after its use. Unlike the statistically based techniques, correlations and dependence between features or variables do not have any effect on LAD's per formance. LAD can handle this phenomenon, and moreover, it gives physical explanation to it. Unlike other data analysis techniques, such as neural networks and support vector machines, LAD is a transparent method; the output of LAD can be traced back to the specific root causes that resulted in the categorization of a specific observation into a certain class. The interpretability of all the results and of all the steps of calculation is a clear advantage that maintenance experts can use in order to find answers to their questions. LAD detects and diagnoses complex phenomena such as equipment fault due to the interaction of multiple factors. Moreover, LAD is not based and does not need data on failure events. Thus, for physical assets that have a lifetime of 25 or 30 years, degradation can still be analyzed and maintenance decision can still be taken. During the learning process, LAD analyzes the collected data that represents readings of some physical asset's features such as vibration signals, oil analysis, temperature, etc…, which are taken while the asset is in different states, for example new, failing, deteriorating, etc… It then generates patterns that can be easily interpreted and translated into meaningful physical rules. LAD automatically extracts features and generates patterns from the readings and, accordingly, classifies the asset into a certain state based on the patterns generated. The learning function of cbmLAD results in an accumulation and preservation of expert knowledge that can be used at any time by the user, even if the human expertise is lost due to retirement or resignation. It thus becomes a powerful knowledge management tool that automates and conserves knowledge. In the testing process LAD can detect and diagnose the asset's state based on the generated knowledge during the learning process. This process of learning and pattern generation serve to reinforce the theoretical knowledge and uncover new knowledge about a certain diagnostic problem in PAHM. LAD has been applied successful in different situations that will be presented in the following sections.