<![CDATA[A dense network of smart instrumentation is essential to autonomous process industries, with the accuracy of these sensors constantly being compromised by sensor drift, fouling, temperature changes and degradation over time. Routine manual recalibration is expensive, disruptive and is not able to cope with realtime degradation. Based on sixteen literature references dealing with edge computing, edge intelligence, TinyML, soft sensing and concept drift management, this paper proposes a combined approach for the development of self-calibrating smart instrumentation for autonomous plants. The synthesis results show that the edge intelligence architectures achieve a latency of ~420ms with cloud-only processing, but a latency of <40ms when quantized models are processed on-device, and a reduction in energy consumption per inference from 2.85 joules to 0.31 joules. Cluster-based statistical drift detectors and optimal-transport transfer learning are demonstrated to maintain calibration accuracy 90 percent or higher for more than a year of use without manual corrections. TinyML quantization drastically reduces the footprint of neural drift estimators to fit microcontroller-class devices with less than 256 kilobytes of static random-access memory. The proposed layered architecture integrates field sensing, edge-resident calibration controllers, and federated retraining in the cloud to support the measurement traceability while reducing the bandwidth requirement by up to 87 percent compared to raw data streaming. Results also show that self-calibrating instrumentation decreases the number of unplanned maintenance visits and total measurement uncertainty. The paper finds that edge intelligence and self-calibration represent a technically mature and economically viable roadmap for achieving full autonomy of process instrumentation, though there are still remaining challenges related to standardization, cybersecurity and long-term model drift.]]>
Copyrights © 2024
