In Die Crack Detection (IDCD)
During my term at Magna, I worked on developing a real-time structural integrity monitoring system for an industrial hydraulic press. The press used multiple strain gauges embedded near the die to capture deformation data throughout each stroke. I built an end-to-end data pipeline that streamed this sensor data over MQTT, re-factored and synchronized the signals, and stored them in a PostgreSQL database for further analysis. I then created a web dashboard application that allowed operators and engineers to visualize the strain profiles for every part produced, compare them across production batches, and interactively flag or review anomalies. A screenshot of the dashboard detecting a crack can be seen below:
The core of the project was an anomaly detection algorithm I designed to identify when a metal part cracked during forming. I analyzed strain signatures across four sensors and derived a method to detect characteristic discontinuities that occur when material failure propagates. This detection pipeline ran in real time during operation and achieved 100% detection accuracy over a 350-stroke production validation run—correctly identifying every cracked part while maintaining zero false positives. The algorithm effectively removed defective parts from the line automatically, eliminating manual inspection steps. The algorithm flow chart can be seen to below.
​
A visual representation of the "normal" stroke thresholds developed by the alrogithm by looking at several control strokes (red lines) and the current stroke (blue line) for both a normal part, and a cracked part is shown below:
.png)
.png)
The system was smart enough to recognize when a different part was being stamped if it got a completely novel profile, and used a moving window average to not allow slight drifts in the sensor readings to be detected as false positives.
​
A significant technical challenge involved the limited memory capacity of the Beckhoff PLC controlling the press. To support high-rate sensor data aggregation without overrunning memory, I engineered a patent-pending data buffering system using custom semaphore-driven multithreading and staged data queues. This allowed the PLC to stream much larger datasets than originally possible. With this solution, the data collection rate improved from ~140 Hz to over 500 Hz, enabling much higher resolution analysis of material deformation events. Together, these improvements allowed for reliable real-time crack detection and meaningful production-level monitoring while operating inside strict hardware constraints. A flow chart of this logic is shown below:
