A Deep-Coupled AE-KRLS Framework for Robust Measurement: Enhancing Precision and Calibration Transferability in Localization

Accurate and robust indoor localization remains a fundamental challenge in dynamic and noisy environments. Kernel Adaptive Filtering (KAF) methods, such as Kernel Recursive Least Squares (KRLS), offer powerful online signal processing capabilities, but their measurement accuracy is fundamentally limited by two factors: (1) noisy and low-quality RSSI signals, which degrade filter performance, and (2) environment-specific hyperparameters, such as kernel bandwidth, which require costly per-site re-calibration. To address these instrumentation challenges, we propose a novel framework that co-designs a deep autoencoder (AE) with the KRLS algorithm. First, we introduce a feature optimization algorithm guided by the theoretical minimum loss (J_min) of KRLS. This algorithm forces the AE to learn a mathematically optimized feature representation that enhances measurement precision under noisy conditions. Second, we design a transfer learning mechanism, based on MMD/JMMD, that trains the AE to normalize the feature space across different domains. This enables a robust calibration transfer mechanism, allowing a KRLS model trained in one environment to operate accurately in a new one using the same, fixed kernel hyperparameters. Experiments in diverse environments validate our framework, demonstrating significant improvements in measurement accuracy and, crucially, confirming its ability to significantly alleviates the recalibration burden: a model trained on the source domain can be directly deployed to a new domain using the same fixed source-domain hyperparameters, achieving high accuracy without re-tuning. This work provides a principled and practical approach for developing low-maintenance, intelligent measurement systems capable of reliable performance in dynamic, real-world settings.