The modelling, prediction and performance monitoring of manufacturing processes are key research aspects for the optimal design and quality control, in particular for complex thermomechanical processes. Numerical-based modelling techniques such as Finite Element and Computational Fluid Dynamics are widely and used approaches to successfully model complex thermomechanical industrial processes. For real-time applications, however, such modelling techniques are not suitable due to the significant computational cost. In addition, the lack of in-depth understanding of some complex processes, such as Friction Stir Welding (FSW), prohibits the creation of accurate physics-based models. Data-driven modelling offers an alternative solution to model-based analysis of complex processes via the creation of computational structures that are capable of 'learning' from process data. In this thesis, a new data-driven modelling framework is proposed, focusing on real-time processing capability of a complex (and ill-understood for some materials) thermomechanical process: FSW. Specific challenges that this research work addresses includes availability of low number of process samples/data, modelling in multiple process scales (micro-, meso-, macro-), real-time processing capability (hence low computational cost), creation of new monitoring techniques capable of automatically identifying abnormal behaviour (novelty detection) and process optimisation which acts in real-time to ensure optimal Process Operating Windows (POW) in multiple scales. A special research focus of the presented research work is human-centric systems in manufacturing, hence the aspects of natural language feedback to the user and simple (transparent to the non-expert) yet accurate models are also investigated. The proposed hybrid model-based framework is based on Soft Computing, due to the need for system transparency and computational simplicity. This includes Fuzzy Logic-based approaches as well as Neural-Fuzzy (NF) modelling structures, and evolutionary optimisation with multi-objectives (real-time capable). The initial stage of this research investigation includes the creation of NF models, which accurately describe the behaviour of FSW in multiple scales, despite the availability of limited data. FSW, which is a solid-state joining process, is widely recognised in industry (aerospace, shipbuilding, automotive and railway) as an efficient, versatile and environmentally friendly welding technique that produces very high quality welds. Despite this success, many challenges are still ahead, due to the need for process certification and ISO standard compliance (reliable monitoring, and Non-Destructive Evaluation - NDE). A new model-based process monitoring and novelty-detection framework is proposed, it not only accurately monitors and predicts the process performance in real-time and in multiple scales, but it also provides a measure of assessing and predicting the normal or abnormal behaviours of the processes. In particular, this assessment is automatically communicated to the end-user via natural language feedback which is based on Human-Centric System (HCS). This is achieved by mathematically linking a number of process performance indices to a Fuzzy Logic rule base. The end-user reads (automatically generated text) the process performance in terms of forecasted product quality, reliability of model prediction, detection of abnormal behaviour, and overall multiscale process performance. The proposed model-based monitoring and novelty detection system is then coupled with a real-time capable multi-objective optimisation technique: a micro-Genetic Algorithm (micro-GA). For the first time in this field the multiple scales of FSW such as cooling rates, microstructure, mechanical performance, and overall quality of the manufactured parts are optimised in real-time using the proposed approach. The real-time processing capability is achieved by introducing short-length encoding for the micro-GA. The proposed model-based approach covers the whole manufacturing process lifecycle for FSW: process forecasting, monitoring-NDE and optimisation, while it is also generic enough to be employed in other manufacturing processes too, following further development.