Decentralized and Intelligent Estimation: Theory and Applications

Abstract

Contemporary technological development has had a profound impact on the architecture and operation of modern systems. In particular, smart systems, defined by their capacity for adaptation, have emerged as a dominant paradigm across various sectors. This dissertation presents two complementary surveys that establish the conceptual foundation for the technical contributions that follow. The first is a comprehensive examination of smart system architectures, framed through the lens of cognitive dynamic systems, which decomposes smart systems into five core components: control, perception, knowledge, communication, and security. The second is a focused survey on Intelligent Estimation, which explores the intersection of estimation theory and learning-based systems.

Motivated by the increasing reliance on secure and distributed inference, the first technical contribution introduces Decentralized Estimation (DeEst), a novel data-driven decentralized estimation framework that integrates data-driven local inference with blockchain-based federated consensus. In DeEst, each node maintains a local estimator informed by historical observations and contributes parameter updates to a shared global model via a blockchain-federated learning protocol. This architecture eliminates the need for a central aggregator while ensuring robustness to communication failures, malicious nodes, and node data heterogeneity. The second contribution focuses on estimator robustness at the node level through the development of the Sliding Sigmoid Filter (SSF), an extension of the Sliding Innovation Filter (SIF). By incorporating a nonlinear sigmoid-based saturation function, the SSF enables smoother transitions across innovation magnitudes, improving estimation stability in the presence of abrupt deviations or measurement outliers. The SSF’s capacity to modulate updates adaptively makes it particularly well-suited for decentralized implementations, where maintaining local estimator reliability in the face of system faults is essential to ensuring system-wide accuracy. The final contribution presents a novel learning paradigm, referred to as Intelligent Estimation, which reinterprets neural network training as a probabilistic filtering problem. In contrast to conventional gradient-based optimizers such as SGD or Adam, which often suffer from poor convergence in noisy settings, Intelligent Estimation employs estimation methods, such as the SSF, to adaptively scale weight updates based on innovation magnitudes, enabling context-aware and noise-resilient learning. The approach is empirically validated across diverse benchmark datasets, demonstrating improvements in both convergence behavior and generalization performance.