Your operations team knows the failure modes: a flowmeter drifts and goes unnoticed for days. A pump cavitates itself to destruction before anybody sees it in the logs. A well-completion anomaly stays buried in manual well-log interpretation until production collapses. And every incident report ends the same way — the fault was detectable days earlier, hidden in telemetry nobody had rigged the systems to see.
The reason is not a shortage of sensors. Most brownfield plants stream thousands of telemetry points per second across SCADA, historians and distributed equipment. The reason is baseline confusion. A baseline that flags anomalies in one flowmeter type drowns in false positives on another. SCADA dark zones — assets that log no streaming telemetry at all — stay invisible until they break. And nobody has ranked the failures by actual downtime cost, so resources go into predicting rare events while the high-value ones stay reactive.
That is where a telemetry audit lives.
The Baseline Trap: Why Generic Models Fail on Brownfield Telemetry
A predictive-maintenance platform designed on greenfield data arrives at a brownfield plant and hits reality. The vendor's anomaly detector was trained on clean, uniform telemetry from an identical facility. Your plant has several generations of flowmeter hardware. Legacy SCADA talks to historians through a middleware layer. Some wells stream vibration data. Others only log completion records by hand.
The result: the model sees every anomaly from the older flowmeter types as noise. It misses the patterns it was trained to catch because your sensors produce different statistical signatures. Or worse — it flags so many false positives on the equipment with low variance that crews stop trusting the alerts within a week.
The reason is that brownfield plants are not broken. They work, which means they have accumulated a lot of operational wisdom embedded in their variance. A compressor that normally hums with a specific bearing signature is fine; the same acoustic profile at a backup pump is a fault. A produced-water flowmeter sitting at its normal operating rate that drifts slightly higher may be within tolerance, while the same meter reading well below its baseline can signal a blockage. And a well that drops sharply from its recent rate would be alarming, while a slow decline over the same period is normal.
One baseline does not hold all of that.
Mapping the Heterogeneous Reality: The SCADA Audit
The telemetry audit starts with the reality check: what actually streams, what sits dark, what is clean and what is not.
A typical brownfield site has flowmeter coverage that is neither complete nor uniform. Some header lines stream continuously into the historian; others log once a shift when an operator reads the gauge. Duplicate readings exist — once in SCADA, once in a stand-alone datalogger the operations team never deleted. Gaps appear when a flow computer went offline or was swapped out. The audit maps all of that, and for each source it answers three questions: how clean is the signal (how much missing data, how many spikes and dropouts), how far back does the history go (can you establish a baseline), and who currently uses it (is the data operationally real)?
Rotating equipment telemetry — vibration, bearing temperature, discharge pressure — is often the most fragmented. A new compressor might stream full Fourier-transformed spectral data. An older pump sends back a single vibration RMS at a coarse interval. A turbine logs fault events in text to a local file instead of streaming. The audit inventories what is available, the time resolution, and what failures each telemetry type has historically predicted.
Well-log data is often the biggest surprise. Most brownfield plants have decades of well-log scans from drilling and workover campaigns. Those scans sit in PDF archives or old databases, never correlated with production data. An audit that pulls and profiles those logs — measuring porosity, saturation, formation properties — can often unlock per-well baselines that are hidden from pure production telemetry alone.
SCADA dark zones are the gaps that kill predictive models. A satellite platform with no streaming telemetry. A gathering header that has a single pressure transmitter instead of flow measurement. A subsurface manifold with no instrumentation beyond periodic wireline runs. These are not failures of planning; they are the edges of what was economical to instrument decades ago. The audit names each dark zone, assesses whether adding instrumentation is worth the capex for the failure mode it would detect, and flags which anomalies are currently invisible.
The output of the audit is a map of data readiness by facility and asset type. It names the sources that can anchor per-asset baselines today, estimates the signal quality and history depth for each, and identifies where new instrumentation or data integration moves a currently-opaque failure mode into visibility.
Establishing Per-Asset Baselines: The Foundation of Detectable Anomalies
Once the audit is done, baseline work can start. This is not the work of building statistical models; it is the work of defining what normal looks like for every distinct asset type and operating regime.
The key principle is specialization. A baseline for a positive-displacement pump must handle the fact that discharge pressure scales with volumetric flow in a nonlinear way. A centrifugal pump's performance curve is different. A compressor running in a temperature-controlled building has different operating bounds than one in the field. The baseline that works for all three is worthless.
The audit's data inventory tells you which assets have enough history — several weeks of representative operation — to establish a defensible baseline. For each asset type, the work is to compute operating-regime-specific reference profiles. A flowmeter baseline must account for the temperature, pressure and fluid properties that affect its reading. A pump baseline needs separate profiles at different discharge pressures, because a pump running at low head has a different normal thermal signature than one running at high head.
This is tedious work, but it is where false-positive discipline lives. A detection model that drifts into noise on every equipment type that runs outside its training regime will never be trusted. A baseline that explicitly says this pump at this pressure and this temperature has this normal discharge temperature can flag a genuine departure from it, and crews will act on it.
The audit also establishes a per-asset monitoring cadence. Some assets are instrumented well enough to detect anomalies over hours; others have sparse data that only supports daily or weekly baselines. A model that tries to raise an alert on daily data from an asset that has only daily telemetry will either miss fast-moving faults or raise false positives when ordinary day-to-day noise gets interpreted as a trend.
Ranking by Recoverable Downtime: The Prioritization That Actually Matters
This is where most telemetry-audit projects go wrong. They treat all telemetry as equally valuable and all failure modes as equally addressable. The result is a platform that monitors everything and detects nothing actionable.
The ranking question is simple but hard: which failure modes, if detected early, return the most downtime value to the plant?
A flowmeter drift that causes production underestimation might cost hours per month in lost visibility. A pump cavitation that progresses to seal failure costs the plant days of downtime and an expensive replacement. A well-log anomaly that indicates a formation breach costs you not just downtime but an expensive remediation campaign.
The audit that talks to operations teams surfaces these trade-offs. Where do unplanned stops actually happen? Interview the maintenance crews. Dig into the last few years of downtime reports. A brownfield plant usually has a small set of recurring failure modes that account for the bulk of unplanned downtime. Those are your targets.
Then assess detectability. A pump bearing that fails can usually be detected by rising vibration RMS or bearing temperature over several days before catastrophic loss. That failure mode is detectable and high-value. A SCADA sensor going out of range is detectable with any basic monitoring, but it is a nuisance alert with low operational value. A well-log anomaly hidden in a PDF archive is high-value but requires extracting and correlating data that is not streaming, which means you need instrumentation work first.
The output is a ranked list: the top handful of failure modes, each with a recovery value (cost per prevented outage), a detectability confidence (what telemetry already exists to see it), and an effort estimate (how long to build and validate that detector). That list is what gets budgeted and sequenced in the transform phase.
Brownfield plants live on heterogeneous telemetry. Baselines that work for one flowmeter type fail on the next, and SCADA dark zones hide the failures your crew finds out about too late.
- 95%
- Detection at under 2% false positives
- 40%
- Less unplanned downtime
- 25%
- Lower maintenance cost
- 4-6 weeks
- To ranked roadmap
SCADA Integration: Where Brownfield Alerting Lives
Once baselines are established and high-value failure modes are prioritized, the audit also maps the integration reality. Predictive models do not live in isolation; they land on top of the control stack you already have.
A brownfield plant runs on SCADA and historians that were installed years or decades ago. The system works. It is stable. The cost of touching it is measured in unplanned downtime. This is why the models that actually shipped on these sites do not demand a rip-and-replace. They sit on top of the existing infrastructure and route their alerts into the existing maintenance workflow.
The audit identifies the data integration points: where SCADA historian APIs can be tapped, whether the system has user-defined variables that can be written into by an external model, what alerting machinery already exists. A well-scoped telemetry audit maps the middleware question: can the model read from the historian directly, or does it need to poll a SCADA workstation and extract data manually? Does the alerting need to hit a ticketing system, or can it write directly into the maintenance dashboard your crews already watch?
Where to start
A 4-6 week telemetry audit is the minimum viable assessment for any brownfield operations site. The work is not complex; it is just detailed and conversation-intensive.
Weeks 1-2: Inventory and access. Work with operations and engineering to list every data source: SCADA servers, historians, stand-alone datalogging equipment, third-party monitoring systems, archived well logs. Get read access to the historian or SCADA, or arrange for a trusted systems person to extract data for you. Map the SCADA architecture: which assets stream, which are dark, what latency and reliability each source has.
Weeks 2-3: Data profiling and baseline assembly. Pull representative periods of operational data — several weeks of history for each asset type — and compute basic statistics: mean, variance, drift patterns, missing-data rates. For each asset type (pump, compressor, flowmeter), establish operating-regime-specific baselines: what does normal look like at different discharge pressures, temperatures, or production rates?
Weeks 3-4: Failure-mode mapping and downtime audit. Talk to maintenance teams. Which assets have the highest unplanned downtime? Which failures cascade into the longest or most expensive outages? Which happen most frequently? Plot each recurring failure mode against the telemetry you have: is that failure detectable in the data, or is it happening in a SCADA dark zone?
Weeks 4-6: Roadmap and integration planning. Synthesize the ranked list of failure modes by recovery value and detectability. For the top handful, estimate the effort to build and validate a detector. Identify SCADA integration points and middleware questions. Produce a ranked anomaly-detection roadmap tied to your existing control infrastructure.
The output is not a deployable model. It is the prerequisite to one: clarity on which failures you can see, which ones are worth detecting, what integration work is needed, and what per-asset baselines will keep the detector honest as your operations evolve.
Every brownfield operations team knows that unplanned downtime is the margin killer. What they often do not know is that the telemetry to prevent most of it is already streaming through their SCADA, waiting for someone to ask: what does normal look like for this asset, in this regime, and how far does it have to drift before we call it a fault?
“Brownfield plants live on heterogeneous telemetry. Baselines that work for one flowmeter type fail on the next, and SCADA dark zones hide the failures your crew finds out about too late.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production — sovereign models, governed data, and agents you can audit. Start with a value-first assessment.