Distributed Autonomous Edge Analytics

<p dir="ltr">This thesis addresses the complexities of achieving real-time, efficient, and accurate autonomous behaviour in volatile environments. Inspired by the dynamic positioning of deep-sea vessels under various forces such as fluid dynamics, wind, ocean currents, waves, and volcanic eruptions, we developed an autonomous system for unpredictable conditions. The proposed solution combined the Receptor Density Algorithm (RDA) from the artificial immune systems with a homeostatic plastic spiking neuronal model (PSN) for position monitoring. While effective in stable conditions, integrating Deep Reinforcement Learning (DRL) into this hybrid bio-inspired model enhanced adaptability to unpredictable environments. A comparative analysis against the initial RDA-PSN-only system demonstrated that the DRL combined homeostatic system has consistent, predictable accuracy. It displayed the ability to predict and apply forces for future/unseen random environmental conditions. This super-human behaviour inspired us to analyse time series data on the environment, position, and system response. We transformed the data into sub-sequences by applying Piece-wise Aggregate Approximation (PAA) and Symbolic Aggregate Approximation (SAX) transformations to compress and preserve the data behaviour. A transitions matrix captured the frequencies of transformations and subsequently preserved the behaviour of the subsequences. The matrix was then reduced into a single digit using a novel digitisation technique, maintaining the explainability through the magnitude of the digit. We were inclined to transform the real-time environment, position, and system response data as single transitions in the transition matrix to reduce them into explainable digits. This insight converted the entire data spectrum into digits in the transition matrix. Moreover, the digits provided insights into the stability or volatility of the situation. We leveraged this aspect into our autonomous system by deploying the system into edge devices under various environmental conditions and replacing the input and output data with digits in the transition matrix. We proposed a model for closely emulating the behaviour of a supervised learning DNN in unpredictable environmental conditions via a novel distributed self-organising plane. The distributed plane consists of a network of edge devices connected to a centralised system of self-organising map (SOM) layers. The centralised system simultaneously gathers and disperses the most stable DNN structures through collaborative learning. This swarm-like model replacement behaviour enables the network of devices to converge into a single-state machine in unison, resulting in the behaviour of a supervised learning DNN. Our simulations originated a distributed self-organising plane, which caused the DRL’s Deep Neural Networks (DNNs) to cease backpropagation at the edge. The novel Distributed Autonomous Edge Analytics (DAEA) system is 27% faster than stand-alone autonomous systems with backpropagated DNNs. DAEA emerges as a resilient and effective solution for real-time systems in unpredictable conditions. In conclusion, this research presents a comprehensive approach to addressing unpredictable environmental conditions, offering an autonomous distributed learning system.</p>