سيدمحمد قوامي

System identification is one of the fundamental challenges in the analysis an‎d modeling of complex industrial processes, as achieving accurate identification plays a key role in controller design, process monitoring, performance optimization, an‎d reliable decision-making in industrial operation. In this thesis, the distillation column process is selec‎ted as the case study for system identification, as it represents one of the most complex an‎d widely used industrial processes due to its highly dynamic behavior an‎d strong interactions among process variables. Considering the limitations of classical identification methods, as well as the challenges associated with conventional recurrent neural networks such as high parameter complexity an‎d limited interpretability a data-driven identification framework based on recurrent neural networks is proposed. In this framework, different attention mechanisms are incorporated to improve the overall performance of the model. In this study, baseline recurrent neural network architectures with varying structures an‎d depths are systematically analyzed an‎d compared in order to selec‎t a reference structure that provides an appropriate balance between identification accuracy an‎d computational complexity. Subsequently, different attention mechanisms are implemented on the selec‎ted architecture, an‎d their performance is eva‎luated from multiple perspectives. Model eva‎luation using various accuracy an‎d error metrics enables a comprehensive analysis of identification precision, prediction stability, an‎d robustness against noise. Simulation results demonstrate that incorporating attention mechanisms significantly improves identification accuracy an‎d reduces modeling error compared to conventional recurrent architectures, while simultaneously enhancing model interpretability. Specifically, the contribution of different historical time steps in the model’s decision-making process can be analyzed an‎d interpreted. These characteristics make the proposed approach a suitable an‎d effective solution for data-driven identification of complex industrial systems.

شناسايي سيستم , شبكه عصبي بازگشتي , فرايند برج تقطير , مكانيزم توجه

system identification , recurrent neural network , distillation column process , attention mechanism

Seyed Mohammad Ghavami

Dr.Seyed Majid Esmailzadeh