City Research Online

Abstract

Predictive maintenance represents a data-driven methodology that applies machine learning and analytics to foresee equipment failures before their occurrence, thus reducing downtime and associated maintenance costs. Historically, the practice of predictive maintenance, especially in the context of vehicles, has relied on anomaly detection techniques applied to sensor data. In recent developments, vehicular predictive maintenance has transitioned from leveraging raw sensor inputs directly to employing fault events recorded within On-Board Diagnostic systems (OBDs). Rather than delivering raw sensor data, OBDs provide drivers and technicians with diagnostic information derived from various Electronic Control Units (ECUs) in vehicles, typically represented as Diagnostic Trouble Codes (DTCs).

Despite their significance and widespread use in the automotive industry, the application of conventional machine learning techniques to predictive maintenance using DTCs presents considerable challenges: DTCs are non-numeric, the spectrum of DTC codes is exceedingly vast, and they can coexist with additional attributes such as fault-bytes and Electronic Control Units (ECUs). These challenges have prompted researchers to examine a limited subset of DTCs at one time and to employ basic algorithms. Moreover, most algorithms applied to DTCs are heavily reliant on repair and warranty data to formulate problems within a supervised learning paradigm, such as categorizing sequences of events as faulty or non-faulty. However, in the absence of access to such data or clear indicators of vehicle non-operability, implementing supervised learning techniques, which necessitate substantial quantities of labeled data, can be difficult, if not infeasible.

This study initially re-conceptualized the task of vehicle fault prediction as a self-supervised next-DTC prediction problem, introducing a novel architecture that harnesses the strengths of deep learning models to directly confront the intrinsic complexity of DTCs. This is achieved by learning dense representations of DTC events through the use of neural embeddings, applied separately to each DTC attribute. Such a method enables our models to utilize sequential algorithms such as LSTM layers, thereby facilitating precise predictions of the subsequent event with consideration for all three attributes per timestep.

The second approach proposes an architecture that combines Gated Recurrent Units (GRUs) with an attention mechanism to succinctly encapsulate the complete DTC sequence into low-dimensional embeddings, facilitating efficient representation of multivariate event sequences, thereby enhancing accuracy, interpretability, and the capability for semantic search of individual DTCs and their associated sequences.

The third approach consolidates the advantages of transformer and GRU models, achieving a superiority of approximately 2% in the top-5 accuracy benchmark for the next-DTC prediction task. This model illustrates that large models while achieving outstanding performance in state-of-the-art research, do not necessarily operate optimally in domains constrained by data size.

Finally, we developed an improved variant of the DTC-TranGru model, referred to as DTC-GOAT, which incorporates various optimization techniques to augment prediction accuracy. Furthermore, we demonstrated how the ensemble approach, which entails the combination of multiple models for next-DTC prediction, can enhance top-5 accuracy outcomes compared to individual models.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science Q Science > QA Mathematics > QA76 Computer software T Technology T Technology > TL Motor vehicles. Aeronautics. Astronautics
Departments:	School of Science & Technology > Computer Science School of Science & Technology > School of Science & Technology Doctoral Theses Doctoral Theses

Export

Downloads