We are interested in multi-site production systems consisting of n plants co-operating following a k-out-of-n reliability structure. The production is interrupted when at least k among n plants fail due to their high deterioration level. Maintenance of plants is therefore crucial to retain a nominal total production rate. Each plant gradually deteriorates following a homogeneous gamma process, and fails when its deterioration level exceeds a fixed threshold. To optimize the system performance, preventive and corrective maintenance of each plant should be coordinated into a predictive maintenance policy at the system level. Maintenance actions require some specific devices which are not always available, for which maintenance resources are considered limited. The aim of this paper is to propose predictive maintenance policies that properly exploit available resources to keep the highest system availability at an acceptable cost. The performances of these policies are evaluated and optimized through maintenance cost models. Various numerical examples illustrate these policies.