Federated learning is a decentralized approach used to train global machine learning models without sharing data between participants, and it has become a key solution present in situations where interested parties data must be protected. This is very important in data-driven prognostics, health management and anomaly detection systems as key data ownership is divided between several original equipment manufacturers and operators. However, a proper implementation of this technology requires significant upfront investment in infrastructure, as computing, energy and network capabilities must support an increased load on the edge, representing a shift from the centralized paradigm. Despite these demands, the automotive industry has shown substantial interest in the potential of this technology as a collaboration enabler. The privacy benefits of this technology are well recognized, however it is often applied indiscriminately, without thorough consideration of its appropriateness for each context. To address this, we conducted a detailed systematic literature mapping on the topic and through our analysis, we provide insights into the usefulness of federated frameworks in terms of their effectiveness in addressing identified specific challenges for predictive maintenance and anomaly detection applications in the automotive industry. Moreover, we contribute by identifying real world applications for automotive industry where the implementation of this technology really makes sense. Building on this, we performed an experimental analysis using widely adopted models and aggregation strategies to evaluate federated learning’s performance under various data split configurations that simulate real-world conditions. Our study tested how each method responds to different data scenarios. Results show that FedAvg performs best with balanced data, while FedProx excels with imbalanced distributions, where its regularization techniques address disparities. These findings highlight the need for tailored approaches to meet the unique demands of each application. While federated learning holds promise, its implementation may not always justify the costs, especially if only a few key challenges are addressed by the framework. Tailoring federated configurations can optimize predictive maintenance and anomaly detection in the automotive industry, but careful consideration about usefulness and infrastructure cost is crucial for long-term success.