Piezoelectric Accelerometer Charge Amplifier Saturation Case Study

This case study follows a vibration measurement channel that looked credible in the configuration file but failed during machine startup because the piezoelectric accelerometer and charge amplifier were configured for low-level steady vibration, not for the transient acceleration that actually reached the sensor.

The failure is useful because the accelerometer was not broken. The charge amplifier was not wired incorrectly. The problem was a measurement-chain mismatch: high sensor sensitivity, small feedback capacitance, a long overload recovery time, and an acceptance procedure that trusted RMS summaries after the waveform had already clipped.

Case Summary

Item	Engineering relevance
System	Gearbox endurance test stand with bearing-housing vibration monitoring.
Sensor	Charge-mode piezoelectric accelerometer mounted on the driven-end bearing cap.
Interface	Charge amplifier, anti-alias filter, 16 bit ADC, vibration acceptance software.
Original objective	Verify that steady-state casing vibration remains below the acceptance limit.
Failure observed	Startup waveform flat-topped at the amplifier limit, then recovered slowly.
Hidden weakness	Charge-amplifier gain was selected from steady vibration, not startup shock and low-frequency base motion.
Main consequence	RMS and spectral values after clipping understated the true transient and corrupted the acceptance window.
Corrective action	Re-range the charge amplifier, delay acceptance until overload recovery is complete, add overload flags, and validate against an independent reference channel.

The central engineering question was:

Did the displayed vibration describe machine motion, or did it describe the saturated analog front end?

The answer was the latter during startup and early steady operation.

Initial Measurement Chain

The test stand used a high-sensitivity charge-mode accelerometer because the team expected small steady vibration after alignment and balancing. The signal chain was:

1. piezoelectric accelerometer;
2. low-noise coaxial cable;
3. charge amplifier with selectable feedback capacitance;
4. analog low-pass anti-alias filter;
5. 16 bit data acquisition module with a nominal \pm 10\ \text{V} input range;
6. software that calculated RMS vibration and spectral peaks.

The original acceptance procedure checked steady vibration after the gearbox reached operating speed. It did not explicitly check startup shock, overload flags, charge-amplifier recovery, cable-motion noise, or whether the first accepted time window began after the analog channel had returned to linear operation.

Simplified Channel Data

Use these screening values from the investigation.

Quantity	Symbol	Value
accelerometer charge sensitivity	S_q	1000\ \text{pC}/g
original feedback capacitance	C_f	1.0\ \text{nF}
feedback resistance	R_f	10\ \text{G}\Omega
amplifier linear output limit	V_{lim}	\pm 9.8\ \text{V}
ADC nominal input range		\pm 10\ \text{V}
ADC resolution		16 bit
steady acceptance limit		0.75g_{\text{RMS}}
measured startup transient on reference channel	a_{pk}	18g
target anti-alias cutoff		5\ \text{kHz}

The values are simplified, but the mechanism is common: a sensitive piezoelectric accelerometer can be a good sensor while the configured measurement range is still wrong.

Step 1: Check Charge-Amplifier Sensitivity

For an ideal charge amplifier, output magnitude is:

The charge generated by the accelerometer is:

Therefore the voltage sensitivity is:

With:

and:

the output sensitivity is:

The linear acceleration range before amplifier clipping is approximately:

Engineering Comment

The channel can resolve small steady vibration, but it cannot linearly measure acceleration above about 9.8g. That may be acceptable for a low-level condition-monitoring window. It is not acceptable for startup, impact, rub, loose-fixture events, or acceptance tests where the maximum transient has not been bounded.

Step 2: Predict the Startup Overload

The independent reference channel measured a startup transient near:

With the original charge-amplifier setting:

The amplifier can produce only about:

So the expected overload factor is:

The waveform must clip before the ADC sees it.

Engineering Comment

Once the analog front end clips, later digital processing cannot recover the missing peak. RMS, crest factor, spectral peaks, kurtosis, order tracking, and alarm logic are all derived from a distorted waveform. A clipped signal can look less severe than the real event because the peak is artificially limited and the waveform shape is changed.

Step 3: Explain the Slow Recovery

A charge amplifier uses feedback resistance to define the low-frequency time constant. A simplified estimate is:

With:

and:

the time constant is:

The corresponding low-frequency corner is:

After overload, the residual output can decay approximately as:

If the channel is driven to:

then after 5\ \text{s}:

After 30\ \text{s}:

Engineering Comment

The low corner frequency is good for preserving low-frequency dynamic content, but the same long time constant makes overload recovery slow. If the software starts accepting data five seconds after startup, the apparent vibration can include a large recovery transient from the instrument rather than the machine. The validation boundary must include when the channel is considered recovered and usable.

Step 4: Re-Range the Channel

The team changed the feedback capacitance to reduce voltage sensitivity. A practical target was to measure at least:

within a \pm 9.8\ \text{V} linear output limit. The required maximum sensitivity is:

Select:

The needed feedback capacitance is:

So:

With this setting, the 18g startup transient predicts:

The event is now inside the linear region with margin.

Engineering Comment

The new setting sacrifices some voltage sensitivity, but it converts an invalid measurement into a valid one. Sensitivity is useful only after range, recovery, noise, bandwidth, and calibration are satisfied. A high-gain saturated channel is not more accurate than a lower-gain linear channel.

Step 5: Check ADC Resolution After Re-Ranging

For a 16 bit ADC over a \pm 10\ \text{V} range, the voltage step is:

With the new sensitivity:

the acceleration increment per ADC count is:

The steady acceptance limit is:

The quantization increment is therefore:

or about 0.2\% of the acceptance limit.

Engineering Comment

The lower gain still leaves enough digitizer resolution for the acceptance decision. The more important remaining questions are analog noise, mounting resonance, cable motion, calibration uncertainty, anti-alias filtering, and whether the acceptance metric is robust to startup transients.

Step 6: Separate Machine Vibration from Instrument Artifacts

The investigation used four evidence checks.

Evidence	What it showed
Raw waveform review	Flat tops at \pm 9.8\ \text{V} proved analog clipping.
Overload flag logging	The DAQ recorded overload events during startup and occasional coast-down impacts.
Reference accelerometer	A lower-sensitivity channel measured transient peaks above the original range.
Re-range repeat test	With C_f=5.0\ \text{nF}, startup peaks remained inside range and RMS values changed materially.

The team also checked sensor mounting torque, cable strain relief, connector cleanliness, shield continuity, and bearing-cap surface preparation. These checks did not remove the startup transient, but they reduced the chance that cable motion or a loose mount was being mistaken for machine vibration.

Root Cause

The root cause was not a failed piezoelectric element. It was a configuration and validation failure:

1. Charge-amplifier gain was selected from expected steady vibration rather than the maximum credible transient.
2. The acceptance procedure did not require raw waveform review or overload flag review.
3. The software accepted early time windows before charge-amplifier overload recovery was complete.
4. The test plan did not compare the high-sensitivity channel with an independently ranged reference channel.
5. The measurement uncertainty budget omitted clipping and recovery as failure modes.

This is a measurement-chain failure. The physical effect, mounting, cable, amplifier, filter, ADC, software timing, and acceptance rule all had to be reviewed together.

Corrective Actions

The released configuration changed both hardware settings and evidence requirements:

1. Use C_f=5.0\ \text{nF} for endurance-test acceptance runs.
2. Record raw waveform data for every startup and coast-down transient.
3. Reject any acceptance window containing amplifier overload, ADC overrange, missing samples, or recovery baseline drift.
4. Start steady-state acceptance only after a documented recovery delay or after the channel baseline passes a return-to-zero criterion.
5. Keep a lower-sensitivity reference accelerometer during commissioning and after fixture changes.
6. Add a cable tap test and mounting verification to the pre-run checklist.
7. Store amplifier range, sensor sensitivity, calibration certificate, cable serial number, filter cutoff, sampling rate, and firmware scaling with the test record.

The goal is not merely to prevent clipping. The goal is to make clipped data impossible to treat as valid acceptance evidence.

Validation Matrix

Validation item	Acceptance evidence
Range check	Predicted maximum transient below 80\% of analog full scale.
Overload behavior	Injected or mechanical transient produces logged overload and automatic data rejection.
Recovery	Baseline returns within the specified tolerance before acceptance starts.
Bandwidth	Shaker or reference excitation confirms usable band and anti-alias cutoff.
Mounting	Torque, surface condition, adhesive or stud condition, and cable strain relief recorded.
Noise floor	Quiet-run measurement shows adequate signal-to-noise ratio for the smallest required spectral feature.
Cross-check	Reference accelerometer agrees within the stated uncertainty band during commissioning.
Traceability	Sensor, amplifier, ADC, calibration, and software scaling are recorded as one measurement chain.

Engineering Lessons

A piezoelectric accelerometer does not measure acceleration by itself. The useful measurement is produced by the installed sensor, mechanical load path, cable, charge amplifier, filters, digitizer, software, and acceptance procedure.

High sensitivity is not automatically better. If the channel clips, recovers slowly, or hides overload flags, the output can be precise-looking but physically invalid. A defensible vibration measurement states the expected maximum transient, steady measurement range, frequency band, sensor mounting, amplifier time constant, filtering, sampling rate, uncertainty budget, and release criteria.

The practical rule is simple: before trusting an RMS value or spectrum, inspect whether the time-domain signal was acquired linearly. If the waveform is clipped, the engineering decision must be suspended until the measurement chain is re-ranged and validated.