Inspection and Maintenance Planning in Large Monitored Structures

Abstract

Structural health monitoring (SHM) systems have become a popular technology to collect information about the condition of structural assets. SHM is typically paired with traditional onsite inspection and maintenance (I&M). Incorporating SHM information in I&M planning is non-trivial, but is key to an improved prediction of the structural condition and to save costs on I&M. In this paper, we propose a framework to combine SHM information with visual inspections and repair actions. It utilizes a hierarchical dynamic Bayesian network (DBN) to probabilistically model the deterioration in the structural system, including all component interactions. The DBN calculates the evolution of the system probability of failure given the monitoring information and I&M actions. By attributing discounted costs to I&M actions and to the consequences of failure, we calculate the expected life cycle-cost for a given I&M strategy by simulating monitoring and inspection histories. The cost optimization of I&M strategies is then approached with a heuristic approach. Finally, the value of information of the SHM system is calculated by comparing the optimal strategies with and without monitoring system.

Value of structural health information in partially observable stochastic environments

“…The respective observation matrix is also described by Eq. (24). As discussed in Section 3.2., this is a default observation at no cost for the purposes of evaluating the VoSHM.…”

“…Dynamic Bayesian networks are utilized in [22], to determine the underlying structural deterioration process. Based on the established dependencies, the cumulative life-cycle cost is evaluated and the policy space can be subsequently searched through optimization heuristics [10,23,24], genetic algorithms [25], or other relevant optimization solvers. POMDPs are also built within dynamic Bayesian network premises, so the two approaches can be seen as equivalent in terms of how the environments are simulated, however, their adopted optimization approaches and capabilities are completely different.…”

Deep reinforcement learning driven inspection and maintenance planning under incomplete information and constraints

“…Also known as the curse of history, this problem is optimally tackled by dynamic programming and POMDPs as explained in detail in the next section. Another approach to attack this complexity, however, often at the expense of solution efficiency, is to exploit problem-specific characteristics and employ simplified assumptions, including approaches that impose action periodicity, policy uniformity among components, component prioritization, ranking, or clustering [12,56,57,58,59,60]. Particularly in inspection planning, periodic inspection visits or non-periodic inspections that exploit similarity and/or prioritization of components is typical for deteriorating structural systems [10,58].…”