Perpetual Learning Framework based on Type-2 Fuzzy Logic System for a Complex Manufacturing Process

This paper introduces a perpetual type-2 Neuro-Fuzzy modelling structure for continuous learning and its application to the complex thermo-mechanical metal process of steel Friction Stir Welding (FSW). The 'perpetual' property refers to the capability of the proposed system to continuously learn from new process data, in an incremental learning fashion. This is particularly important in industrial/manufacturing processes, as it eliminates the need to retrain the model in the presence of new data, or in the case of any process drift. The proposed structure evolves through incremental, hybrid (supervised/unsupervised) learning, and accommodates new sample data in a continuous fashion. The human-like information capture paradigm of granular computing is used along with an interval type-2 neural-fuzzy system to develop a modelling structure that is tolerant to the uncertainty in the manufacturing data (common challenge in industrial/manufacturing data). The proposed method relies on the creation of new fuzzy rules which are updated and optimised during the incremental learning process. An iterative pruning strategy in the model is then employed to remove any redundant rules, as a result of the incremental learning process. The rule growing/pruning strategy is used to guarantee that the proposed structure can be used in a perpetual learning mode. It is demonstrated that the proposed structure can effectively learn complex dynamics of input-output data in an adaptive way and maintain good predictive performance in the metal processing case study of steel FSW using real manufacturing data.