Domain-Aware Hyperdimensional Computing for Edge Smart Manufacturing

TL;DR

This paper questions the universality of how HDC parameter choices affect performance across domains and demonstrates that domain-aware encoding is essential for edge manufacturing. By evaluating two representative tasks—image-based LPBF defect detection and signal-based CNC quality monitoring—the authors show that nonlinear nonlinear Random Fourier Features (RFF) are beneficial for CNC while linear Random Projection (RP) suits LPBF, with optimal hypervector dimensionality $D$ and projection spread $\sigma_b$ varying by task. A formal complexity analysis plus Bayesian multi-objective optimization reveals that inference, training, and retraining scale linearly with feature size and hypervector dimensionality, but retraining costs depend non-monotonically on encoding choices, making closed-form optima elusive. Empirically, domain-aware tuning yields HDC models that match or surpass state-of-the-art deep learning and Transformer baselines while delivering at least 6× faster inference and more than 40× lower training energy, highlighting HDC as a practical path to real-time industrial AI on constrained hardware. The work provides actionable guidance for deploying HDC in edge manufacturing and points to future directions in adaptive encoding, broader modality evaluation, and hardware-focused optimizations.

Abstract

Smart manufacturing requires on-device intelligence that meets strict latency and energy budgets. HyperDimensional Computing (HDC) offers a lightweight alternative by encoding data as high-dimensional hypervectors and computing with simple operations. Prior studies often assume that the qualitative relation between HDC hyperparameters and performance is stable across applications. Our analysis of two representative tasks, signal-based quality monitoring in Computer Numerical Control (CNC) machining and image-based defect detection in Laser Powder Bed Fusion (LPBF), shows that this assumption does not hold. We map how encoder type, projection variance, hypervector dimensionality, and data regime shape accuracy, inference latency, training time, and training energy. A formal complexity model explains predictable trends in encoding and similarity computation and reveals nonmonotonic interactions with retraining that preclude a closed-form optimum. Empirically, signals favor nonlinear Random Fourier Features with more exclusive encodings and saturate in accuracy beyond moderate dimensionality. Images favor linear Random Projection, achieve high accuracy with small dimensionality, and depend more on sample count than on dimensionality. Guided by these insights, we tune HDC under multiobjective constraints that reflect edge deployment and obtain models that match or exceed the accuracy of state-of-the-art deep learning and Transformer models while delivering at least 6x faster inference and more than 40x lower training energy. These results demonstrate that domain-aware HDC encoding is necessary and that tuned HDC offers a practical, scalable path to real-time industrial AI on constrained hardware. Future work will enable adaptive encoder and hyperparameter selection, expand evaluation to additional manufacturing modalities, and validate on low-power accelerators.

Figures (14)

• Figure 1: Domain-dependent HDC tuning. The optimal hyperparameters, such as hypervector dimensionality, projection variance, and training set size for maximizing accuracy and minimizing latency, vary across different application domains.
• Figure 2: The HDC framework, illustrating the four core phases: encoding, training, inference, and retraining.
• Figure 3: The experimental setup, showing the EOS M270 LPBF machine lane2016thermographic, the top view schematic of the built part, and the side view schematic with dimensions, including a 40.5 degree overhang. Moreover, captured optical images from the process of the overhang and the bulk areas.
• Figure 4: Experimental setup for the CNC drilling task, showing the steel workpieces, drilling locations, and representative signal channels collected during machining. Z-score-based labels (Under Drilling, Nominal Drilling, Over Drilling) are assigned based on measured hole diameters.
• Figure 5: Classification accuracy for the CNC task using (a) Cosine and (b) RP encoding, and for the LPBF task using (c) Cosine and (d) RP encoding.