TL;DR: The diagnostic engine sits inside the OT network, consuming telemetry from flow meters, pressure transmitters, level sensors, and temperature probes -- devices generating years of historical data per installation. The models were trained on telemetry across the entire installed base: millions of device-hours covering normal operation, degradation, and failure modes. The inference pipeline ingests real-time sensor data, device metadata (firmware, calibration history, installation date), and process variable correlations. A flow meter reporting zero flow while its downstream pressure transmitter shows a spike is not two independent anomalies -- it is a correlated fault signature the model recognizes as a blocked impulse line. This correlation capability separates ML-based diagnostics from rule engines. Critically, the training data came from real field failures, not lab simulations. A pressure transmitter failure at a chemical plant looks different from one at a wastewater facility, and the training corpus captures that variance. The system classifies faults into four categories. Sensor drift -- gradual deviation undetected for weeks -- is flagged when the trend emerges, not when it crosses a threshold. Electrical noise from ground loops, VFD interference, or failing analog cards produces characteristic frequency patterns matched against known signatures. Mounting issues -- incorrect insertion depth, impulse line blockages -- are inferred through cross-instrument correlation. And process condition changes -- two-phase flow, cavitation, unexpected fluid properties -- are distinguished from device faults, eliminating the most common support call: the no-fault-found dispatch. The system reads from existing plant historians (OSIsoft PI, AspenTech IP.21) and communicates with DCS via OPC-UA -- the same protocol connecting PLCs, HMIs, and SCADA. OPC-UA exposes not just process values but diagnostic parameters (signal quality, electronics temperature, sensor impedance) through standardized address spaces. The AI builds a multidimensional health view beyond the 4-20 mA signal the operator sees. The historian provides long-term memory: when an anomaly appears today, the AI queries five years of history to establish baseline behavior and correlation patterns. The 80% figure represents actual support case deflection. For each remotely resolved fault, the plant avoids truck rolls, phone-support cycles consuming operator attention, and downtime costs cascading from measurement faults in custody transfer or quality-critical applications. Device lifetime extension is the less obvious lever: detecting drift at 2% deviation and recalibrating preserves five years of useful life that would otherwise be lost to undetected degradation. The deployment validates principles applicable beyond instrumentation. Field data already exists in historians -- it was simply never structured for ML consumption. OPC-UA adoption is not optional for AI-driven diagnostics; its semantic richness provides the feature vectors that make classification possible. And the vendor relationship shifts from reactive support to co-engineering: when 80% of faults never generate a call, the vendor's value is in maintaining the model, not answering the phone. The next threshold is autonomous resolution -- AI diagnosing the fault, identifying the corrective action, and executing it through the DCS without human approval. That future is closer than most plant managers think.
For the complete architectural breakdown -- including the inference pipeline data flow, the four-category fault taxonomy in full detail, and the OPC-UA integration pattern with historian retrospective analysis -- read the full analysis at susiloharjo.web.id:
[Link] https://susiloharjo.web.id/ai-plant-measurement-fault-diagnosis/
Related on Susiloharjo: