Design Considerations and Guidelines
improve the performance of the predictive model owned by a standalone client . Subsequently , this will bring potential organisations together to form an ecosystem in which , data can act as a strategic resource for the benefit of all . Figure 2-1 presents an example network ecosystem aligned with some example scenarios that benefit from FedL applicable in the context of smart manufacturing :
Large scale organizations can use on-premises computational nodes to train predictive models supporting product failure prediction using a variety of data gathered from manufacturing sites , thus forming localized data silos . On the contrary , small-scale organizations can often lack access to large training datasets and as a result develop less robust models . FedL can help in such scenarios where similar equipment operating on multiple sites of different organizations form a virtual network and share their failure patterns with each other to reach a consensus about the failure prediction model .
Prediction models deployed by an organisation require an initial supervised training procedure with the labelled data . Usually , labelling is performed manually by subject matter experts . To save time , auto-labelling using transfer learning is also feasible . Auto-labelling for un-seen data can benefit from FedL as a new participatory agent ( organisation ) can connect with the FedL ecosystem and gain access to a robust model trained with a variety of patterns observed by other agents over a longer period .
OEM provides equipment to multiple organisations and has the capacity to monitor its performance over time . This data can be utilised to optimise / fine-tune the performance of the product for their clients . Additionally , it can provide an opportunity to minimise downtime , preempt procurement delays by automating the procurement process of parts / equipment based on remaining useful lifetime of equipment or its components . However , OEM is often unable to gather the data from multiple organisations due to privacy and trust issues . FedL can be leveraged in such scenarios where an OEM distributes a set of global functions to the organisations along with the equipment . The functions take gathered equipment data as input and return the computed values to the OEM .
On-time product quality assessment is an important step towards cost savings and zero-defect manufacturing . An initial model can be deployed on factory floors to classify damaged products . However , it is observed ( Mohr , 2020 ) that different damages in the same product can be
22 March 2022