Faults diagnosis in a discrete system using a hybrid reinforcement learning and data-driven algorithms - دانشکده فنی و مهندسی
Faults diagnosis in a discrete system using a hybrid reinforcement learning and data-driven algorithms
نوع: Type: Thesis
مقطع: Segment: PHD
عنوان: Title: Faults diagnosis in a discrete system using a hybrid reinforcement learning and data-driven algorithms
ارائه دهنده: Provider: Mohsen Shiri
اساتید راهنما: Supervisors: Dr. Hadi Delavari
اساتید مشاور: Advisory Professors: Dr. Younes Solgi
اساتید ممتحن یا داور: Examining professors or referees: Dr. Mohammad Hassan Moradi, Dr. Seyed Jalil Sadati Rostami, Dr. Seyed Mohammad Mehdi Mousavi
زمان و تاریخ ارائه: Time and date of presentation: 2025
مکان ارائه: Place of presentation: 33
چکیده: Abstract: In this thesis, a novel artificial intelligence–based, data-driven fault diagnosis framework is proposed for identifying various sensor and actuator faults in model-free discrete-time nonlinear systems. To ensure satisfactory system performance under both healthy and faulty operating conditions, discrete-time sliding mode controllers—including a discrete adaptive sliding mode controller and a discrete nonsingular terminal sliding mode controller—are concurrently designed and developed to achieve accurate reference output tracking. The proposed fault diagnosis system consists of two main components: a fault detection module and a fault estimation module. In the fault detection stage, a novel discrete sliding mode observer, a novel discrete adaptive sliding mode observer, and a novel discrete nonsingular terminal sliding mode observer are introduced to enable fast and accurate fault detection. In the fault estimation stage, the type and magnitude of faults occurring in discrete-time systems are estimated using, respectively, a novel fractional-order neural network, reinforcement learning with a fractional-order observation function, and fractional-order actor–critic reinforcement learning. The proposed methods exhibit several key advantages, including a significant reduction in false alarms and mitigation of the chattering phenomenon in the fault detection module. Moreover, in the fault estimation module, online learning capability, convergence, and adaptability to new operating conditions and previously unseen faults are guaranteed through the integration of fractional-order neural networks and fractional-order reinforcement learning algorithms. The utilization of fractional calculus enables more accurate modeling of system dynamics, suppresses undesirable oscillations, and improves transient performance. Simultaneously, by reducing computational burden, the proposed framework ensures feasibility for real-time implementation. The learning and weight update processes in both neural networks and reinforcement learning algorithms are performed using a gradient descent scheme based on fractional-order derivatives. The stability of the closed-loop controlled system is analytically investigated and rigorously proven using Lyapunov-based methods. Furthermore, to optimally tune the parameters of the proposed controllers and observers and to ensure a fair and accurate performance comparison, optimization techniques based on ant colony optimization, neural networks, and reinforcement learning are employed. Finally, the effectiveness and superiority of the proposed approaches are validated through extensive simulations conducted on theoretical models, laboratory-scale systems, and industrial applications. The obtained results confirm a substantial improvement in fault detection accuracy, detection speed, and fault estimation quality compared with existing state-of-the-art methods.