On an approach to the integration of artificial intelligence into the fault tree analysis methodology

Abstract

In modern conditions of intensive operation of air transport, ensuring the reliability and failure-free operation of technical systems is a priority task, directly related to flight safety and economic efficiency [1]. Statistical data analysis indicates that technical malfunctions, ranking second among the causes of flight cancellations, create significant operational and financial risks. The problem is particularly acute in the context of aircraft engines, which are critically important components whose failure can have catastrophic consequences [1]. A crucial factor is also the time deficit between flights necessary for scheduled maintenance, which requires the implementation of reliable predictive diagnostic methods.

Traditional approaches to reliability assessment, based on classic analytical models (e.g., the exponential distribution law), although fundamental, often prove insufficiently sensitive to reflect the multifactorial and non-linear nature of wear and failure processes in complex modern systems [2]. In this regard, there is an urgent need to develop integrated methodologies capable of combining the structural logic of engineering analysis with the predictive power of artificial intelligence (AI).

The article proposes an integration that combines the classical Fault Tree Analysis (FTA) methodology with machine learning algorithms, namely Long Short-Term Memory (LSTM) recurrent neural networks [3]. The FTA methodology provides the necessary structural framework, establishing cause-and-effect relationships between basic events (component failures) and the top event (engine failure) through logical operators.

The key idea is to use LSTM networks for the dynamic prediction of the probability of failure P_i (t+Δt) of each individual critical system node. LSTM networks, known for their ability to effectively process large volumes of sequential data (time series), are trained on historical engine operational parameter data, which includes temperature, pressure, vibration, and other critical indicators Xi (t) [4]. This allows the model to capture complex, long-term dependencies that precede the actual failure.

The predicted failure probabilities P_i (t+Δt) of individual components are dynamically substituted into the FTA logical model (equation P = f (P₁, P₂, ..., P_n)), which allows for the real-time updating of the integrated reliability assessment of the entire system Psys(t). This mechanism not only increases the accuracy of the overall prediction but also allows for the operational reduction of the fault tree, narrowing the scope of the search for the most probable scenarios. As a result, engineering and technical personnel can quickly identify the source of a potential malfunction, which is critically important for transitioning to condition-based maintenance and minimizing the risks of unexpected failures.