Model improvement using federated learning and canonical feature mapping

The method provides for receiving a plurality of trained models from a corresponding plurality of clients, wherein a respective trained model predicts a condition of an asset and is based on a data set associated with the asset of a respective client. The trained model is based on a seed model that includes a canonical set of features. The trained model includes a component that converts the data at a site to the canonical set of features used by the seed model. The plurality of trained models from the corresponding plurality of clients is assigned to two or more groupings, wherein a grouping includes trained models providing similar analysis. The one or more processors generate an improved model for a client with a limited amount of training data, obtaining the improvement by using multiple models that belong to the same grouping of the first client's model.