Why single-sensor monitoring misses faults

Single-sensor monitoring is still the default in most plants: one accelerometer per bearing, one camera per station, one thermocouple per cabinet. It is cheap, well understood and often good enough. But it misses faults in predictable ways, and those misses are built into the approach rather than caused by poor sensors.

This article explains the five structural blind spots of a single channel, works through an illustrative example and ends with a checklist for deciding when a second modality is worth it.

What single-sensor monitoring is, and why it is so common

Single-sensor monitoring means watching a machine or process through one type of measurement, usually one sensor per measuring point, and raising alarms from that channel alone. The classic form is a vibration transmitter on a pump bearing, wired to a PLC with a threshold.

It is common for good reasons. Vibration in particular has a mature standards base. ISO 20816-1 sets out general procedures for measuring and evaluating machine vibration and, importantly, evaluates both the vibration magnitude and changes in it. For a simple machine with one dominant failure mode, a single, well-mounted accelerometer with sensible limits works.

The trouble starts when the machine has several important failure modes, when false alarms are expensive or when nobody can go and check the sensor every time it complains.

Five ways a single sensor misses faults

1. The fault does not change the measured quantity

Every sensor is blind to faults that do not change its physics. A loose electrical connection heats up long before it causes any vibration. A misplaced part on a fixture changes geometry but makes no sound. A coolant flow problem shows in temperature, not in acceleration. A single accelerometer cannot see any of these, no matter how good its signal processing is.

2. Different faults produce the same signature

Many faults share indicators. A rise in overall vibration velocity at running speed can come from imbalance, misalignment, looseness or a resonance. Raised high-frequency energy can mean bearing wear, poor lubrication or cavitation. A trained analyst separates them with spectra, phase and experience, but an automated system on one channel often cannot.

ISO 13379-1:2025 frames diagnostics as working from a machine's measured operational parameters (inputs) and responses (outputs). The more independent responses you observe, the fewer candidate faults remain. One channel leaves many candidates open.

3. The sensor itself fails

Sensors loosen, cables chafe, connectors corrode, amplifiers saturate and adhesive mounts creep. A loose accelerometer produces impacts and broadband noise that look remarkably like a bearing defect. A dirty camera lens degrades a vision score slowly, just as a drifting process would. A thermocouple that has lost contact reads low and steady.

With one channel, the system has no grounds to distinguish "the machine changed" from "the sensor changed". Either it raises a machine alarm, which costs a technician's trip and some credibility, or it learns to ignore that sensor. Our article on self-monitoring sensors covers this failure mode in more detail.

4. The time scale does not match the fault

Each channel has a sampling or update rate, and the fault has to fit inside it. The Nyquist criterion, as Walt Kester's ADI tutorial puts it, requires the sampling frequency to be at least twice the highest frequency in the signal; otherwise aliasing folds higher components into the band of interest where they cannot be told apart from real signals.

The same logic applies to imaging. The FLIR Lepton 3.5 has an effective frame rate of 8.7 Hz. A 50 ms thermal event can fall entirely between two frames. Conversely, a vibration channel sampled at tens of kilohertz is poorly suited to a fault that develops over weeks unless its data is trended.

5. The operating context changes the signal

Load, speed, product type, ambient temperature and neighbouring machines all change what a sensor reads. A pump running at a different duty point vibrates differently. A cold morning shifts every thermal reading. A single channel cannot easily separate a context change from a fault unless context is recorded alongside it. Often another modality, such as a camera showing which product is on the line, is what provides that context.

An illustrative example: one pump, one channel versus three

The following scenario is illustrative, not a report from a specific installation. It shows how the same events read through one and through three channels.

A centrifugal pump has an accelerometer on the drive-end bearing housing. Over two weeks, the overall vibration level rises by a noticeable margin and the high-frequency band becomes noisier. On one channel, the possible explanations include early bearing damage, a loosened sensor mount, cavitation from a process change and a loose foot.

Now add a thermal view of the bearing housing and a microphone near the pump.

• If the housing is warming relative to the non-drive end and the microphone picks up a new tonal component, early bearing damage becomes the leading hypothesis.
• If the housing temperature is unchanged and the microphone hears broadband crackling, cavitation moves up the list.
• If neither thermal nor acoustic channel changes, a sensor or mounting problem is likely and a technician can check the mount first.

None of these channels diagnoses the pump on its own. The value is in the combination: each additional, physically independent observation removes candidates. The NASA Reliability-Centered Maintenance Guide states the same principle for predictive testing technologies: results should be correlated with each other to increase the probability of detection.

Checklist: is single-sensor monitoring enough for this machine?

Use this checklist per machine or per measuring point. Each "no" is a reason to consider a second modality.

• Does every important failure mode change the quantity your sensor measures?
• Can your sensor and its analysis separate the failure modes that matter, not just detect "something changed"?
• Can you tell a sensor fault from a machine fault without sending someone to look?
• Do the sampling rate and bandwidth match the fastest fault you need to see?
• Is the operating context (load, speed, product, ambient) recorded so a change can be explained?
• Are false alarms cheap enough that you can tolerate several per year on this asset?

A machine that passes all six is a good candidate to stay single-sensor. One that fails two or more usually benefits from multimodal monitoring, as described in our guide to multimodal sensing.

Adding a second modality without overbuilding

The goal is not to put every sensor on every machine. The most efficient step is usually one physically independent channel for the most critical failure modes. The table pairs common primary sensors with a second modality that does not share their blind spots or their typical disturbances.

For a deeper view of each modality's strengths and weaknesses, see what vision, thermal, depth, sound and vibration each see and miss. Where sound and vibration compete for the same job, our article on acoustic versus vibration monitoring helps choose the lead channel.

Once a second channel is in place, the evaluation needs to change too. Thresholds on single channels should give way to models that consider the channels together, and ideally to a ladder from anomaly detection to fault classification to severity grading.

How NeuroControls approaches this

Our systems are built on the assumption that one channel is not enough for most real machines. The Machine Health Kit, built to order, combines the Edge AI Hub with a Vibration Node and an Acoustic Node, and the Edge AI Hub in its Multi-Camera configuration carries global-shutter cameras and a radiometric thermal core. That gives structure-borne, airborne, thermal and visual channels on one asset.

On our cognitive sensors, signal health and dropouts are flagged at the edge, so gaps are visible in the record. Custom models trained by our engineers on customer data go from anomaly detection to fault classification and severity grading. More on the application side is on our machine condition monitoring and machinery industry pages.