CBR-Based Decision Support System for Maintenance Text Using NLP for an Aviation Case Study

Recently, Prognostics and Health Management (PHM) has emerged to promote predictive maintenance as a methodological key to overcome the limitations of traditional reliability analysis. The Natural Language Processing (NLP) methods allow the maintenance log usage for maintenance diagnostics and decision making. The Maintenance Work Orders (MWOs) contain vital health indicators and decades of experience related to various maintenance actions. However, due to the unstructured nature of maintenance text, it is not common to develop a tool using these textual maintenance entries. This paper proposes a textual Case-Based Reasoning (CBR) approach combined with Technical Language Processing (TLP) to find solutions for new problems based on previous experiences. The Bidirectional Encoder Representations from Transformers (BERT) model is adopted for maintenance data using unsupervised finetuning technique Transformer-based Sequential Denoising AutoEncoder (TSDAE) for aviation case study. Results show that the pre-trained BERT model can adopt domain-specific data and produce semantic matches with only a small amount (1000 samples) of domain specific data.