Artificial Intelligence for Fault Prediction, Detection & Diagnosis in Smart Asset Management

Author:
Preprint
Publisher website
Google Scholar
Abstract
This article attempts to address the three main questions any asset manager would have raised since the concept of an asset as an entity to be managed was developed. Question One: Why did I not know about this problem before? followed by Question Two: Is this issue really occurring? And then, finally, Question Three: How much and how long will it take me to correct it? This article proposes a cascaded structure of Artificial Intelligence (AI) for complex management decisions in Smart Asset Management for Fault Prediction Detection and Diagnosis (FPDD). Specifically, a Deep Learning (DL) neural network first performs a regression algorithm to predict future values as Fault Prediction to address question one. This structure is followed by another DL neural network that performs a classification algorithm as the Fault Detection & Diagnosis to define faults, therefore answering question two. The key value added by this article is that the asset fault prediction, detection and classification model is based on future predictions, rather than actual values; hence, issues can be proactively managed, reducing time, energy and cost, thus supporting the third question. The practical application of this proposed model is based on AI and data-driven control and automation, specifically smart building assets and their Heating, Ventilation and Air Conditioning (HVAC) systems. The two proposed AI algorithms analysed are Transformers and Long Short-Term Memory (LSTM) networks, which are broadly deployed in positional and time series applications such as text translation and trend prediction. Supported Vector Machines (SVMs) and Convolutional Neural Networks (CNNs) were discarded due to their learning time and convergence in a multi-classification application on this complex model. Experimental results demonstrate the proposed model successfully predicts, detects and diagnoses faults. On average, the Fault Detection precision for the LSTM is 95.30% and the Transformer is 94.43% against a 50% random precision, whereas the Fault Diagnosis precision for the LSTM is 49.57% and the Transformer is 47.89% against a 4.34% random precision for the eight-variable model. With one month of learning and prediction windows, the Fault Prediction precision of 95.43% for the Transformer and 60.37% for the LSTM against a 4.34% random precision for the four-variable model. The total number of data points ingested and analysed by the AI algorithms is two times 7.51E+09, equivalent to the return distance between the Planet Earth and the Sun in kilometres.

This publication has 36 references indexed in Scilit: