Safety-critical Cellular Internet of Things (CIoT) systems demand a high degree of availability to ensure safe and uninterrupted operations. Given the distributed and heterogeneous nature of these systems, spanning IoT devices, edge nodes, on-premises servers, and cloud platforms, traditional monolithic software architectures are inadequate in delivering the necessary flexibility, scalability, and reliability. In contrast, the Microservices Architecture (MA) offers a promising alternative by promoting decentralization, modularity, and dynamic scalability. However, despite its advantages and some improvements for availability, MA does not necessarily meet high-availability requirements, particularly in safety-critical settings. Advanced availability management mechanisms tailored to the unique demands of safety-critical CIoT systems are required to enhance their availability. In this work, we propose an availability management framework for MA based safety-critical CIoT systems, integrating machine learning based anomaly detection to enable proactive fault-tolerance as reactive mechanisms often fail meeting availability requirements. Our framework aims not only at detecting and recovering from failures, but also predicts potential failures before they manifest, thereby avoiding/minimizing downtime and enhancing the availability of the system.