Embeddings For The Identification Of Aircraft Faults (Merit)

Vector representation concept proves its success in solving many real-world problems from a variety of applications. In this paper, we built a novel vector representation model for avionics system for two types of fault messages called MERIT. This new model aims to identify the relationship between the flight deck effects (FDEs) and the maintenance messages (MMSGs) through calculating the embedding co-occurrence matrix between them within a predefined flight leg window. The a vector space embeddings representation of MERIT is able to differentiate between the strong and weak relationship between messages. Moreover, we benefit from the negative sampling method to incorporate the weak relationship between the FDEs and MMSGs from different subsystems (chapters) in assessing this relationship precisely. We called the developed MERIT with specialized negative sampling approach subsystem-wise MERIT. Both developed models can be used as descriptive and predictive tasks based on the flight leg window used (one and three, respectively). The main advantage of the proposed latent aircraft system model (MERIT) is that it needs to be trained only once and can be easily queried using any similarity measurements between the embedding vectors, which means it is more feasible and computationally efficient than traditional machine learning algorithm, where it necessitates building a different model each time for every target FDE. We tested both models on a real Boeing dataset and the experimental results demonstrate the effectiveness of the proposed model in exhibiting the embedded relationships between fault messages and extracting the most relevant predictors.