Federated Hyperdimensional Computing for hierarchical and distributed quality monitoring in smart manufacturing

Abstract

In emerging smart manufacturing, the integration of the Internet of Things (IoT) and edge devices is essential for in-situ sensing, communication, and adaptive learning. Federated Learning (FL) leverages edge-cloud collaboration to preserve data privacy and minimize communication overhead compared to centralized models. However, conventional FL approaches face significant challenges in manufacturing: (1) non-Independent and Identically Distributed (non-IID) data and diverse feature distributions complicate local model training within hierarchical, complex industrial data structures; (2) directly overwriting local models with a global model during updates causes clients to lose critical task-specific information unique to their environments; and (3) transmitting model updates causes communication overhead, limiting scalability. We propose Federated Distributed Hyperdimensional Computing ($\mathrm{FedDHD}$), an FL framework that employs Hyperdimensional Computing (HDC) to optimize communication for hierarchical manufacturing data. Unlike neural networks, HDC offers robust performance with lower computational demands and inherent resilience to noisy, non-IID data, enabling $\mathrm{FedDHD}$ to naturally handle data heterogeneity and reduce computational burdens on edge devices. $\mathrm{FedDHD}$ integrates a hierarchical graph-based learning model with a node pruning module to alleviate computational load and implements a novel client-cloud update strategy leveraging HDC’s high-dimensional representations to streamline synchronization, thereby minimizing communication costs and improving scalability. We validate $\mathrm{FedDHD}$ through a case study on machining using a Sinumerik edge device, focusing on the geometric quality assessment of two counterbore diameters. $\mathrm{FedDHD}$ achieved an F1-score of 95.3% and demonstrated performance improvements of up to 12.6% over state-of-the-art neural network-based FL methods, highlighting its superior efficiency and scalability in complex industrial settings.