HALT and HASS case studies with measured reliability results.

Commentary by Kirk A. Gray

This paper provides a useful practical example of what Highly Accelerated Limit Testing (HALT) can reveal when environmental stress is combined with continuous functional and parametric monitoring. Shibutani, Tsuboi, and Pecht subjected a programmable robotic system containing 12 servomotors and associated control electronics to cold, heat, rapid temperature change, vibration, and combined stresses. Rather than simply determining whether the system continued to operate, the researchers monitored electrical behavior and used machine-learning techniques to identify changes that preceded readily observable functional anomalies.

Practicing reliability engineers should pay particular attention to the thermal results. During cold testing, observable servomotor anomalies occurred at −70°C, while during heat testing shaft position began changing above 40°C and continued through the 80°C test limit. These failures were recoverable when temperature returned toward normal conditions—an excellent example of the type of soft, or operational, failure that HALT is intended to expose. Rapid thermal transitions also produced anomalous behavior that was not necessarily apparent through visual inspection alone.

Perhaps the most important lesson is the value of monitoring parameters rather than relying solely on pass/fail functional testing. The researchers found that anomaly scores could begin changing before an operating failure became physically observable. Their Gaussian graphical model detected changes even earlier than the individual k-nearest-neighbor analysis. The results also showed that interactions between the servomotors and their controller mattered—the behavior could not always be understood by considering an individual component in isolation.

For practicing reliability engineers: Do not just ask whether the product is still working. Monitor how it is working as environmental stress is increased. Voltages, timing, current, signal quality, error counts, communications behavior, sensor outputs, and other measurable parameters may begin shifting well before a conventional functional test declares a failure. Those changes can provide valuable clues to shrinking design margin and interactions within the system that deserve investigation.

— Kirk A. Gray