Improving Condition-Based Maintenance Transparency Through Causal explanations: An Aircraft Maintenance Scheduling Case Study

Deep Reinforcement Learning (DRL) has gained significant attention in aircraft condition-based maintenance scheduling (AMS), owing to its ability to quickly find optimal solutions. In our latest study, we introduced a strategy rooted in eXplainable Deep Reinforcement Learning (XDRL) to address this challenge. This approach shed light on the agent's prioritized objectives by utilizing a decomposed reward Deep Q-Network (drDQN) algorithm. While drDQN has shown promising interpretability and performance, human operators may desire a more profound comprehension of the agent's interaction with the environment and its predicted future outcomes when presented with a specific state and interventional action. To improve the transparency of our AMS-XDRL framework, we enhanced our existing drDQN agent by incorporating an action influence model based on structural causal modeling (SCM). This enhancement enables the framework to generate causal explanations for inquiries concerning action preferences. The responses encompass causal chains that link the action to the prioritized reward type, presented in natural language to articulate the relationships between causes and effects. At a broader level, the AMS-Causal xDRL (AMS-CxDRL) framework acquires a collection of structural equations describing the causal impact on environmental variables arising from the implementation of an interventional action. Following an action, AMS-CxDRL utilizes these equations to predict the values of environmental variables and component rewards, essential for establishing the causal chains in the explanations. The AMS-CxDRL not only provides explanations in natural language but also visually depicts causal chains in an action influence graph (AIG) through a user-friendly graphical interface, ensuring comprehension of causal explanations.