SIDDHESH
JOSHI.

In welding, quality is normally verified only after the weld is complete, through post-process inspection. That is slow and costly, and defects in the weld seam — porosity, slag inclusion, burn-through, irregular penetration — are caught only once the component is already produced, where they can compromise the strength and service life of the whole structure. Industry needs to detect and correct issues during welding itself.

The weld result depends on a set of process features — current, voltage, wire feed rate, gas flow rate, arc distance, travel speed, gas-nozzle angle and more. I developed a Python framework that connects to multiple hardware systems to capture these in real time: an NI USB DAQ with a signal-conditioning unit for current/voltage, a Baumer monochrome industrial camera, and an Optris thermal camera — all synchronized by timestamp. The framework supports configurable sampling rates, multi-rate logging and auto-trigger on arc voltage, with a GUI for live waveform display and camera preview. The architecture is built to extend to additional sensors as the project progresses.

BEAS Technology GmbH (Chemnitz) had already collected weld process data as current, voltage and timestamp values — features that strongly drive the quality and appearance of the weld seam. They needed a system to perform offline monitoring of the weld seam from that dataset: to find anomalies, pinpoint their position, identify the defect type, and report a confidence level for each.

I built an independent Python GUI implementing a full machine-learning workflow. The user loads training data and trains a PCA model, then loads a test dataset. Detection combines multivariate SPC — Hotelling's T² and Q-residual control limits at a 99% UCL — intersected with an Isolation Forest for robust anomaly identification. Detected regions are then segmented and classified by feature using a Random Forest / XGBoost ensemble. Results are presented as time-series and scatter plots and a table of defect regions, exportable to PNG/PDF/SVG and a detailed Excel report covering time, region, defect type and confidence.