Fault root cause tracing in complex mechanical systems via probabilistic logic and reinforcement learning

When a failure occurs in a large-scale mechanical system such as a marine engine, the maintenance team typically conducts a sequence of inspections on various system components to progressively gather fault information and ultimately localize the root cause, a process known as fault root cause tracing. This process requires the inspection sequence to be planned efficiently. We formulate fault root cause tracing as a Markov decision process (MDP). This MDP is challenging to solve as its system dynamics are unknown. To capture the logical relationships between symptoms and root causes, we construct a Markov logic network (MLN) by integrating historical fault data with expert knowledge from failure mode and effects analysis (FMEA) tables. Parameters in the MLN are estimated using a variational expectation-maximization algorithm, in which the variational distribution is flexibly parameterized by a neural network. The constructed MLN facilitates the inspection sequencing for fault root cause tracing problems. The MLN-based reasoning enables efficient root cause inference by updating the posterior probabilities of root causes as new information is obtained through sequential inspections. We integrate the constructed MLN into a reinforcement learning framework to train a policy for fault root cause tracing problems. The effectiveness of the proposed method is validated through both simulation experiments and a real-world case study involving starting air system failures of harbor tug engines. Our method outperforms benchmark approaches in both inference accuracy and inspection efficiency.