Weld quality AI: 86% fewer defective parts

Every weld was photographed under a microscope and inspected. But machine settings drifted slowly. Each new weld strayed further from the standard. By the time the problem was obvious, defective parts had already been made.

The data existed but was scattered across thousands of spreadsheets. Sofixit linked microscope images, lab results and line data. Each result was tied to its workstation, robot and specific weld. The result: one shared source of information.

We built a system that analyzes weld images: it measures their geometry and detects irregularities. We compared its results with historical lab reports. As a result, quality assessment became automatic and repeatable, for every single part.

A single image shows how a weld looks now. A series of measurements shows where the process is heading. The system caught small deviations and predicted when quality would fall below standard. The team got a warning before the first defective part was made.

This is what working with the system can look like. The dashboard shows several robots, but robot seven on line three stands out. Its weld quality index drops to ninety-two points. The latest measurement is still within tolerance, but the geometric deviation keeps growing. The forecast shows about four hours until tolerance is exceeded, if the trend continues. An alert appears with the exact location and a recommendation to check the process.

The lab manager opens the alert details. They see the trend, weld images and recommendations: check part positioning, fixturing and robot settings against the approved recipe. The information goes to those responsible for production. The team makes the correction and logs it in the system. That doesn't close the case yet. New measurements show whether the deviation shrinks.

The timing of action changed. Before: inspect, find a defect, react after the fact. After deployment: forecast, warn, and correct before defects appear. Machine servicing could fit into planned maintenance windows.

Deployment result: eighty-six percent fewer defective parts and an end to the unplanned downtime described in this project. The value of this change is time to react before a problem starts generating costs. Fewer parts to scrap, and corrections built into the production plan.