Federated Learning (FL) is changing the way we do Machine Learning (ML) by moving from standard centralized to decentralized and privacy-preserving approaches. A common FL strategy, the horizontal approach, consists in aggregating the weights of neural networks sharing the same structure, thereby enhancing the predictive abilities of the models in a collaborative manner. However, this approach doesn’t work when the models from different parties analyze diverse input data. In these situations, evaluating the performance improvement of models with varying architectures among parties requires considering the distinctions among them. In this work, we focus on a collaborative regression prognostics problem: to predict the Remaining Useful Life (RUL) of turbofan engines measured by multiple stations using different sensor spaces. We use two ML architectures: Multilayer Perceptron followed by a Kalman Filter (MLP-KF) and Functional Multilayer Perceptron (FMLP). We propose a non-horizontal approach, i.e., an approach that does not assume that all parties share the input data. Each party catches a different subsample of sensor signals to solve the federated problem utilizing index synchronization and the Vertical Asynchronous Federated Learning (VAFL) algorithm. We create a simulated environment where each party assesses multiple models by repeating the experiment for each ML architecture. Then, we apply appropriate key metrics to compare prediction variability among parties and experiments. Results show that it is possible to use non-horizontal federated approaches to tackle scenarios where the input space is different among parties. The evaluation revealed that FMLP outperformed MLP-KF, providing more stable predictions.