AI-Assisted Diagnostics. Where AI actually earns its keep in the control room.
"AI" is the most overused word in industrial software marketing. Most "AI-powered" SCADA features are just statistical thresholding with better visualization. The interesting cases — the ones that genuinely change operator outcomes — are narrower than the marketing suggests but more valuable than the skeptics admit. Here's the honest version.
The honest taxonomy of "AI" in industrial software
Four classes of technique commonly labeled "AI" in this space. Knowing which is which changes how you evaluate vendors:
1. Statistical thresholding with smart visualization
Compute a moving average, flag deviation beyond N standard deviations, display it on a dashboard. Not really AI by any modern definition. Often labeled "AI-powered anomaly detection." Useful, but the math has been available in statistics textbooks since the 1950s. Cost should reflect that.
2. Classical machine learning for classification
Random forests, gradient-boosted trees, support vector machines trained on labeled historical data. Genuinely useful for tasks like "classify this vibration signature as one of seven known fault modes." Mature, well-understood, explainable.
3. Deep learning for sequence and image tasks
LSTMs, transformers, and CNNs for time-series prediction, vision-based defect inspection, audio-based fault detection. Higher capability, much higher data requirements, less explainable. Where the cutting edge actually lives.
4. Large language models for operator interfaces
LLMs to translate alarm dumps into natural-language summaries, query historians via conversational interface, draft maintenance reports. Newer category; promising but unproven for safety-critical workflows. Hallucination risk is real.
Where AI earns its keep — the actual use cases
Predictive failure on rotating equipment
Compressors, pumps, motors, gas turbines. Multi-mode signature analysis (vibration + temperature + load + electrical signature) trained on years of operating data can flag degradation 2-8 weeks before failure. This is the most-validated use case in industrial AI, and the one with the clearest ROI.
Fleet-wide correlation
Spotting patterns across hundreds of similar assets that aren't visible in any single asset's data. "Twelve sites on this repeater are degrading at the same rate" is a signal a human operator can't extract from 12 individual screens. See our radio degradation case study.
Alarm rationalization assistance
AI to cluster alarm storms, identify root-cause groupings, suggest bypass scopes. Not replacing operator judgment — augmenting it with pattern recognition the operator doesn't have time to do at 3 AM. See our alarm-flood case study.
Anomaly detection on network and protocol traffic
Used by Dragos, Claroty, Nozomi for cybersecurity. AI learns normal protocol patterns and flags deviations. Strong evidence base; primary value is reducing false-positive noise that would otherwise overwhelm a SOC.
Vision-based inspection
Cameras on critical equipment running CNNs trained to detect leaks, corrosion, foreign objects, weld defects. Mature for some applications (welding), still emerging for others (general visual inspection).
The hard problems no AI vendor talks about
Data sufficiency
ML models need labeled examples. Industrial failures are rare by design (good maintenance prevents them) and rarer in modeled data (historians compress away the interesting moments). Training a useful predictive model often requires 2-5 years of historian data and labeled incident data. Most operators don't have both.
Domain shift
A model trained on Operator A's data doesn't generalize cleanly to Operator B's equipment, environment, or operating regime. Vendors claiming "trained on the entire industry" are often glossing over how much per-customer adaptation is required.
Explainability and trust
An operator who doesn't understand why the AI is making a recommendation will reject it. Black-box models lose adoption even when they're right. Explainable-AI techniques (SHAP, LIME, attention visualization) help but add complexity. Aevus prioritizes explainability over absolute accuracy when there's a tradeoff.
Safety
An AI that's allowed to write to field equipment is an AI that can crash the plant. The industry-defining question is how to integrate AI capabilities while keeping the safety boundary architecturally intact. See IL-9000.
How to evaluate an "AI for SCADA" vendor
Six questions to ask:
- What specific class of model is this? Statistical, classical ML, deep learning, LLM? If the answer is "AI," they don't know.
- What training data was used? How much, from how many sites, over what time period? Labeled or unlabeled?
- What's the false-positive rate? Validated on what data, when, by whom? A precision number without context is meaningless.
- What's the explanation surface? When the model says "do X," what can the operator see about why?
- What can the model actually do? Read only? Recommend only? Authorized to write? Authorized to bypass operator? The answers should be in this order, in this priority.
- What's the architectural safety boundary? "We won't write to field equipment" is policy. "We architecturally cannot write to field equipment" is safety. The difference is the whole game.
How Aevus thinks about AI
Aevus uses classical ML for the bulk of its predictive workload (gradient-boosted trees on telemetry features, with explainability built in). We use deep learning for specific high-value cases (multi-mode vibration analysis, time-series forecasting). We do not use LLMs for safety-critical recommendations — only for operator-facing summary generation where hallucination is annoying rather than dangerous, and always with confidence-marked output.
Most importantly: none of our AI is authorized to write to your field equipment. The IL-9000 boundary makes that architecturally impossible, not just policy-restricted. Read the technical brief.
AI in the control room, done honestly.
If you're evaluating AI vendors and want to ask the six questions above without sales pressure, that's the conversation.
Related Articles
Watch AI catch degradation that classic SCADA misses.
SCADA, MES, CMMS, ERP into one decision surface.
The warning was there for 11 days before the outage.
Ready to see how this applies to your operation? Start a pilot conversation — no commitment, no field changes.