Repository of DUTH

Development of novel cognitive network structures with applications in pattern classification, system fault detection and Time Series prediction (Doctoral thesis)

Καρατζίνης, Γεώργιος/ Karatzinis, Georgios

This thesis advances the field of Fuzzy Cognitive Maps (FCMs) by addressing their inherent limitations and extending their capabilities through novel methodologies, architectures, and applications. The research focuses on enhancing the adaptability, scalability, and practicality of Fuzzy Cognitive Networks (FCNs), a specialized class of FCMs, for complex, data-driven, and dynamic real-world scenarios. The foundational contributions include the introduction of functional weights, leading to the development of FCN-FW, which replaces static associations with dynamic, data-driven polynomials. This innovation eliminates the need for extensive fuzzy rule databases, minimizes human expert intervention, and improves the scalability and generalization capabilities of FCNs. Furthermore, specialized topologies tailored to classification tasks were formalized, enabling efficient and accurate convergence to equilibrium states representing class labels. This work also explores ensemble and hybrid architectures by developing Multiple Cognitive Classifier Systems (MCCS) and integrating FCNs with Convolutional Neural Networks (CNN-FCN), Echo State Networks (ESN-FCN), and Autoencoder-FCN (AE-FCN). These hybrid systems combine the interpretability of FCNs with the representational power of deep learning, demonstrating superior performance in feature extraction and classification tasks. A unified training mechanism was proposed, seamlessly blending FCNs with neural networks, further enhancing adaptability. The practical applicability of FCNs was validated through extensive real-world applications, including pattern recognition, time series prediction, and industrial diagnostics. Highlighted implementations include handwritten digit classification, stock market prediction, index tracking portfolio management, remaining useful life estimation for aircraft engines, incipient short-circuit fault detection in induction generators, motor rolling bearing fault detection and a wide set of classification benchmarks. These applications underscore FCNs’ versatility and robustness in addressing challenges in diverse application domains. This thesis also tackled key limitations of FCMs, such as dependency on expert-defined parameters, scalability challenges, and uncertainty handling, through rigorous theoretical and empirical validation. These contributions enhance the adaptability and robustness of FCNs, positioning them as a framework that balances interpretability with computational efficiency, while advancing their application in machine learning, predictive analytics, and decision-making systems.