Abstract

The realization of fully autonomous Industry 4.0 environments depends on wireless networks that are not merely connected, but predictively aware—capable of anticipating channel degradation before it disrupts mission-critical operations. In dynamic industrial environments such as automated warehouses, radio frequency propagation is governed by dense metallic infrastructure, continuously moving obstacles, and frequent line-of-sight to non-line-of-sight transitions that render the channel non-stationary and spatiotemporally complex. Conventional approaches face a fundamental dichotomy: deterministic simulators such as ray-tracing and NS-3 achieve high physical fidelity but require up to 38 hours to generate a single high-resolution SINR heatmap, making real-time adaptation computationally infeasible; empirical and statistical models are fast but site-agnostic, unable to resolve the localised shadow boundaries and multipath-sustained NLOS regions that govern autonomous robot connectivity. This dissertation addresses this gap through two progressive contributions that together constitute a physics-aware generative AI framework for spatio-temporal wireless channel prediction. The first contribution, WISVA (Wireless Infrastructure for Smart Warehouses using VAE), is a physics-informed variational autoencoder that replaces the NS-3 simulator as a fast planning surrogate. By encoding three domain-specific tensors—Euclidean distance, e!ective material permittivity, and access point location—WISVA learns a disentangled latent representation of the electromagnetic propagation manifold. It achieves a Mean Absolute Error of 0.33 dB at 88 ms inference, a speedup exceeding 104 → over NS-3, while generalising to entirely unseen warehouse layouts via a blind-quadrant extrapolation test evaluated across 207,936 spatial points, achieving > 94% recall in NLOS regions. The second and primary contribution, Evo-WISVA (Evolutionary WISVA), extends this framework to the temporal domain. A novel attention-driven Latent Memory Module (LMM) is integrated within the VAE latent space, maintaining a memory bank of past channel states and augmenting the current representation via scaled dot-product attention—providing a non-Markovian gradient shortcut that ensures e!ective spectral radius ω(Jef f ) ↑ 1 and enables long-range temporal credit assignment beyond the reach of standard recurrent units. The memory-augmented VAE reconstruction is then fused with the original physics tensors and processed by a Convolutional Long Short-Term Memory (ConvLSTM) network, which enforces spatial continuity while learning sequential temporal dynamics. The full pipeline is optimised end-to-end via a joint loss function Ltotal = εLVAE + ϑLKLD + ϖLSTM , with the optimal weighting (ε, ϑ, ϖ) = (1.0, 0.001, 1.0) identified through systematic sensitivity analysis. Rigorous experimental validation on high-fidelity NS-3 simulations of 60 GHz warehouse environments demonstrates that Evo-WISVA achieves 0.53 ms inference latency—supporting update rates of exceeding 1,000 Hz (theoretical maximum ↑ 1,887 Hz) under communication and scheduling overhead—and reduces prediction error by 47.6% over state-of-the-art recurrent baselines. Trained exclusively on a single moving obstacle, the model generalises to environments with ten simultaneously moving shelves with only 0.3–0.5 dB degradation in MAE, confirming that the architecture captures the essential spatio-temporal structure of the wireless channel rather than overfitting to specific mobility patterns. Architectural stability is confirmed across five independent training runs, yielding a 95% confidence interval of [0.317, 0.341] dB on the baseline MAE. By transforming wireless channel modelling from a reactive, simulation-driven process into a proactive, predictive intelligence layer, this work establishes a foundational framework for connectivity- aware autonomous systems. The sub-millisecond radio foresight provided by Evo-WISVA enables autonomous robots to navigate communication-safe paths before dead zones form, supports dynamic beamforming and proactive handover management, and advances the realisation of resilient, self-optimising wireless infrastructures for the next generation of industrial automation. This work shifts the wireless network from a passive medium to an active, intelligent participant in autonomous industrial operations.

Publication Date

5-13-2026

Document Type

Dissertation

Student Type

Graduate

Degree Name

Electrical and Computer Engineering (Ph.D)

Department, Program, or Center

Electrical Engineering

College

Kate Gleason College of Engineering

Advisor

Amlan Ganguly

Advisor/Committee Member

Cristian A. Linte

Advisor/Committee Member

Andres Kwasinski

Comments

This thesis has been embargoed. The full-text will be available on or around 6/12/2027.

Campus

RIT – Main Campus

Download

Share

COinS