<p>Friction Stir Welding (FSW) is a solid-state welding process, which has strongly impacted welding technology, particularly for aluminum alloy applications. Reliable in-line process monitoring is not yet available for most common defects and downstream non-destructive and intermittent destructive testing are generally employed to validate weld seam quality. To reduce cost and production time significant efforts have been undertaken in the recent past to develop process-monitoring systems for FSW based on the evaluation of transient process-data. Neural Networks have been used widely to analyse FSW-process data and evaluate the process characteristics or weld seam quality. The data analysed includes welding parameters, thermal-/acoustic-measurement, image or video data and, most notably, the distinct and descriptive process feedback forces and torque. In this study, conducted within the scope of RWTH Aachen’s Cluster of Excellence (Internet of Production), a high granularity direct force measurement setup, which was adapted to the production environment, by integrating reliable, cost-efficient sensors into the machining spindle, was used. Weld data was recorded over a wide range of FSW applications with varying weld-parameters and Al-alloys. Convolutional Neural Networks (CNN) that were previously developed based on measurements of external force and torque sensors were adapted to evaluate the higher granularity data of the new sensor-system and detect volumetric defects within the welds. Good generalization was shown across the weld parameter sets, alloys and welding tool. An average classification accuracy of 98.04% was achieved over three network trainings. Due to the segmentation of data for the evaluation 100% of internal defects were successfully detected by each network iteration. The developed solution aims at offering a highly reliable, spindle integrated and cost-efficient quality monitoring solution for FSW to replace the required expensive and time-consuming testing.</p>