Using Sensor Data for Asset Health

A vibration sensor reporting “4.2 mm/s” at 2:14pm on a Tuesday is not, by itself, useful information. Is 4.2 normal for this specific machine, or a warning sign? Was it 3.8 last week and climbing, or has it been steady at 4.2 for months? The raw number means almost nothing without context — and building that context is the actual work between “we have sensors” and “we have a trustworthy asset health assessment.”

What Turns a Reading Into an Assessment

A baseline. Before any reading can be judged abnormal, there has to be a documented understanding of what normal looks like for that specific asset — not a generic spec-sheet value, but the actual operating range observed across enough time and enough operating conditions (different products, different loads, different shifts) to represent genuine normal variation, not just a narrow snapshot that happens to look artificially tight.

Deviation, not absolute value. A health assessment cares less about the raw reading than how far it has moved from that asset’s own baseline, and how quickly. A reading that’s unusually high for this specific machine but well within a generic industry-standard range can still represent real developing wear; a reading that looks high in absolute terms but has been stable at that level for years on this specific asset may simply be that asset’s normal operating point.

Trend, not snapshot. A single elevated reading can be noise — a temporary load spike, a momentary process variation. A reading that’s been climbing steadily over several weeks is a genuinely different signal, and distinguishing the two requires looking at the trend, not reacting to any individual data point in isolation.

Multiple signals, where available. An asset monitored by more than one technique — vibration and temperature, for example — gives a more reliable health picture than any single signal alone, because a genuine developing failure often shows up across multiple indicators, while noise in one signal is less likely to correlate with noise in another at the same time.

Building the Baseline Correctly

The most common mistake in early condition monitoring deployments is setting alert thresholds too soon, from too little baseline data — before the asset’s normal operating variation across different products, loads, or seasons has actually been observed. A threshold set from two weeks of data on one product mix will generate false alarms the first time the line runs a different product with a genuinely different, but still normal, vibration signature. Taking the time to establish a real baseline before enabling alerting is a small delay that prevents a much larger trust problem — a maintenance team that’s been sent chasing several false alarms stops trusting the system quickly, and regaining that trust afterward is considerably harder than building it correctly the first time.

From Health Score to Action

An asset health assessment that only lives on a dashboard hasn’t finished its job. The point of the assessment is to trigger a specific action — a work order created automatically in the CMMS when an asset’s health score crosses a defined threshold, closing the loop into the digital maintenance program covered in Building a Digital Maintenance Program. An accurate health score that nobody acts on because it’s buried in a dashboard nobody checks delivers essentially none of the value a connected, work-order-triggering version does.

Where This Extends Into Full Predictive Maintenance

Asset health scoring — baseline, deviation, trend — is the foundation. Extending it into an actual remaining-useful-life prediction, the subject of AI for Predictive Maintenance, requires historical failure data to train against, which is a further step once enough operating history — including, ideally, some real failure or near-failure events — has accumulated. Health scoring alone, without that further step, is still genuinely useful on its own; it doesn’t require waiting for enough data to build a full predictive model before delivering value.

Data Without Interpretation Isn’t Diagnosis

Sensors generate data. Asset health, as an actual decision-support tool, requires baseline, deviation, and trend analysis layered on top — and a connection into the maintenance workflow that turns a health assessment into a work order, not just a number on a screen. SG2’s Manufacturing & Industry 4.0 practice builds this interpretation layer as a first-class part of any condition monitoring deployment, not an afterthought left for a plant to figure out once the sensors are already installed.