Separating sensor drift from bearing anomaly in low-frequency vibration telemetry

Predictive maintenance alerts on rotating equipment often fire on sensor drift rather than mechanical change. Operators lose trust in the system within weeks, and the InfiIoT value story breaks before it starts. This note sets up the Center of Excellence test we plan to use to separate the two signals.

The failure mode

A MEMS accelerometer on a motor housing streams vibration features to a cloud dashboard. Over weeks and months, the sensor's zero output and sensitivity drift. On a surface-micromachined variable-capacitance MEMS part, both zero and sensitivity drift are documented behaviours. The anomaly detector trained on early readings sees the drifted baseline as a change in the machine, and fires an alert. A maintenance technician opens the panel, finds nothing, and closes the ticket as a false alarm. After the third such ticket, the shift supervisor stops reading the alerts.

The irony is that the sensor is doing its job. The alerting layer is not distinguishing between a change in the sensor and a change in the asset. The literature agrees on both halves of the problem: MEMS accelerometer long-term stability is a published concern directly tied to false-alarm risk in condition monitoring, and bearing-fault analysis via envelope spectra at the ball pass frequencies (BPFO, BPFI, FTF, BSF) is a mature and well-benchmarked technique when the signal is trusted. The gap is the step in between.

The measurable question

Using the CoE industrial demo panel as a known-good baseline, can a combined approach, reference-channel differencing plus residual-based change detection on envelope spectra, distinguish accelerometer drift from early bearing wear, and at what signal-to-noise ratio does the separation stop being reliable?

Conditions we plan to publish with the result

• Rig: the Center of Excellence industrial demo panel, with a small motor-driven shaft and a serviceable bearing station. Hardware procurement and build are in progress with the current CoE vendor arrangement.
• Reference sensors: two accelerometers co-located on the same bearing housing. One industrial-grade piezoelectric (PE) reference sensor; one representative MEMS sensor of the class used in InfiIoT edge devices. Exact SKUs to be pinned.
• Sampling: 12 kHz on the drive end, with option to go to 48 kHz for broadband capture, matching the CWRU benchmark conventions.
• Shaft speed: three steady-state operating points (approximate: 1500 rpm, 1750 rpm, 1800 rpm), with ramp-up transients captured.
• Fault conditions: controlled bearing swap-ins with known defects (inner race, outer race, ball), as in the public benchmark corpus. Spalls or seeded defects in increasing severities.
• Drift conditions: injected via (a) controlled sensor temperature excursion (ambient to 60 degrees Celsius), (b) progressive sensor-cable disturbance, and (c) software-injected bias on the raw signal stream.
• Baselines:
• Classical envelope spectrum on the MEMS channel, raw.
• Classical envelope spectrum on the MEMS channel after a simple high-pass filter.
• PE reference channel as ground truth.
• Candidate method: residual envelope spectrum calculated as the MEMS envelope minus the PE envelope, with change detection on the residual at the characteristic bearing frequencies.
• Metrics:
• True positive rate on seeded bearing defects.
• False positive rate under injected drift with no bearing change.
• F1 at defect-detection, across SNR levels from clean to noise-dominant.
• Minimum defect severity detectable at each SNR.

Method we plan to test

1. Known-good capture. With a healthy bearing on the rig and both sensors mounted, capture steady-state and ramp vibration at the three speeds. Confirm both sensors see the expected shaft-rate fundamental and no bearing fault lines.
2. Characteristic frequency computation. Compute BPFI, BPFO, FTF and BSF for the specific bearing geometry and shaft speed. These are the lines the detector watches.
3. Baseline envelope analysis. Apply envelope spectrum analysis on each channel. Band-pass filter to the high-frequency range where defect-induced impulsive content lives, demodulate, spectrum-transform. Report the characteristic peaks.
4. Seeded fault runs. Swap in bearings with controlled defects. Capture the envelope spectrum changes and record the SNR at which the defect first becomes detectable above baseline.
5. Drift injection. With a healthy bearing again, inject the three drift modes individually and together. Record the false-positive rate for each baseline method.
6. Residual method. Compute the residual envelope spectrum (MEMS minus PE reference). Repeat the seeded-fault and drift-injection runs. Compare true positive and false positive rates against baselines.
7. SNR degradation. Add controlled broadband noise to the MEMS channel to simulate noisy site conditions. Report the SNR at which the residual method's separation accuracy drops below 80%.

The design is standard practice stitched together: envelope spectrum analysis is the baseline frequency-domain technique for bearing fault identification (ISO 15243 fault types, CWRU conventions). Reference-channel differencing is used in acoustic emission and structural health monitoring to isolate asset change from sensor-chain change. The lab contribution is the measured separation boundary for the specific MEMS-plus-PE combination used in InfiIoT-class deployments.

Observations

None yet. This note is published at Hypothesis stage.

Where this fails

• Single-sensor sites. The residual method requires a second reference sensor at or near the measurement point. Most InfiIoT greenfield deployments are single-MEMS by cost. The report must say when the method does not apply, not pretend it does.
• Shared drift sources. If both sensors drift together (shared power rail noise, shared enclosure temperature), the residual cancels the common-mode but misses the drift. We will test this explicitly and report the failure mode.
• Non-stationary speed. Variable-speed drives smear the characteristic frequencies. Order-tracking is well understood but complicates the detector. The first report targets steady-state and simple ramps; variable-speed sites are a follow-up.
• Bearing types outside the benchmark. Deep-groove ball bearings behave differently from tapered rollers and from spherical rollers. The result is reported per bearing type, not generalised across them.
• Transition from drift to drift-plus-fault. The hardest case in the field is a slowly drifting sensor on a slowly wearing bearing. The report will state the limits of separation in this joint-degradation case rather than paper over them.

What delivery teams can take from this

An InfiIoT reference pipeline for OEM clients deploying vibration-based condition monitoring on rotating equipment:

• A feature-extraction stack: band-pass filter, envelope detection, spectrum computation, characteristic-frequency peak tracking.
• A drift detector: reference-channel residual plus a change-detection gate on the residual.
• An alert gate: suppresses alerts until both the characteristic-frequency peak and the residual confirm the change is in the asset, not the sensor chain.
• A documented failure boundary table, so OEM clients can map their equipment class, sensor configuration and site SNR to an expected separation accuracy before deploying.

The pipeline ships only if the Controlled test holds. If the residual method does not improve on the baseline at realistic SNRs, we will publish that negative result and recommend the periodic recalibration alternative instead.

Open questions

• Is the CoE industrial panel instrumented well enough to inject controlled sensor drift (temperature, cable, software bias), or will we need to stage the drift conditions on a separate bench first?
• Can we source a known-defect bearing set at the Pune workshop level, or do we need to seed defects (EDM, spall insertion) in a controlled way at a university partner's facility?
• Which PE reference sensor do we standardise on for the test? The budget and the long-term stability both matter.
• At what cadence do we need to re-run drift tests across sensor batches to keep the residual-method parameters valid in production?
• Does the InfiIoT product roadmap allow a dual-sensor option for sites that cannot tolerate the drift-driven false-alert rate?

Resolving these moves this note from Hypothesis to Controlled test.

Running vibration-based predictive maintenance on rotating equipment and seeing operator trust erode after false alerts? Propose a challenge →