In this paper, we propose a resilient and fully decentralized federated learning framework specifically adapted for Open Radio Access Networks (O-RAN) named FedRA that enables collaborative intelligence between dApps. At first, we present how FedRA can be deployed in the O-RAN architecture by using already existing interfaces between the well-defined RAN components and how real-time ML applications (dApps) can interact to share their distilled intelligence. Furthermore, the software components of FedRA framework are briefly described, along with the step-by-step deployment workflow, outlining practical guidelines for its implementation. Finally, the FedRA framework is validated in a simulated environment by using both real and simulated datasets that report the time series of the throughout provided by multiple radio units. Two different Machine Learning models that are hosted in FedRA nodes are used, aiming to either detect anomalies in the upcoming network traffic or forecast the data-rate values. In both cases, the resulting accuracy confirms the validity of the FedRA framework in the training process of the cell-specific ML models, as well as the decentralized collaborative intelligence sharing among them. FedRA is also compared to typical client/server approach in terms of achieved model accuracy and network communication cost.