Enhancing tool condition monitoring in friction stir welding with probabilistic neural network algorithm

Introduction Friction Stir Welding (FSW) is a critical industrial process in which a rotating tool generates heat through friction, enabling the solid-state joining of materials. This versatile method is widely applicable across numerous industries, including marine and auto-motive sectors. Method Real-time tool condition monitoring is essential for businesses to identify and address issues before they escalate into costly failures or product defects. While traditional methods such as visual inspection and endoscopy are used to observe tool conditions, they cannot be performed in real-time during welding operations. As a result, specific real-time tool condition monitoring methods are employed for continuous analysis. The real-time tool condition monitoring process involves acquiring vibrational data and extracting statistical features from the raw data. A feature importance study is conducted using a decision tree algorithm, which selects only the most significant features to reduce computational complexity. Result Feature classification is then performed using various machine learning and deep learning algorithms, including Support Vector Machines (SVM), Multi-Layer Perceptron (MLP), Cascade Correlation, GMDH Polynomial Neural Networks, and Linear Discriminant Analysis Among these classifiers, Probabilistic Neural Networks (PNN) consistently deliver the best results as 91.25% under 1,400 rpm. Discussion Based on these findings, the Probabilistic Neural Network algorithm is identified as a robust and reliable prediction model for monitoring FSW tool conditions.