Fluidized bed processes are widely used in industries such as food, pharmaceuticals, and chemicals to produce granular products, including instant powders, catalysts, detergents, and fertilizers. These processes enable intense momentum, heat, and mass transfer, facilitating the formation of diverse particle structures. A key mechanism is granulation, where particles are bonded through solid bridges. The properties of the resulting granules are heavily influenced by process parameters like liquid spray rate and fluidization air temperature. While extensive research has investigated these effects using different mediums, the complex interactions between solid, liquid, and gaseous phases are not yet fully understood. Deviations from optimal process conditions can lead to defects or operational failures, such as over-wetting, particle clumping, and equipment blockages (e.g., bearding). These issues may arise from human errors, faulty equipment, or a lack of process understanding. Effective monitoring and control of process parameters are crucial to ensure consistent product quality and prevent costly production interruptions.

The goal of this project is to establish a systematic, data-driven anomaly detection framework for identifying and responding to critical faults in fluidized bed granulation. To accomplish this objective, the project employs retrofitted sensor technologies to enable the continuous monitoring of critical transient parameters, including spray pressure, temperature, and fluidizing gas velocity.

Based on systematic experimental investigations, a hybrid machine learning framework is being developed for real-time anomaly detection. This approach combines deep learning techniques for high-frequency pattern recognition with contextual heuristics to reliably detect both sudden equipment failures and slow, creeping process drifts. Since the quality of training data is a crucial factor for the accuracy of these models, meticulous data preprocessing is implemented, utilizing advanced signal smoothing techniques to handle noise in industrial environments.

Furthermore, the framework explores the integration of Generative AI to translate complex, multivariate statistical deviations into natural language diagnostic narratives. By formulating immediate, expert-level recommendations for responding to undesirable phenomena, this module lays the groundwork for advanced decision-support systems and automated intervention via the plant's operating software, ultimately guaranteeing improved product quality and process stability.