Sensors & Data Acquisition interview questions

10 questions, easiest first

1. 1You're logging a signal that you believe only has meaningful content below 500 Hz, so you set your ADC to sample at 1 kHz, satisfying Nyquist on paper. After collecting data, you see a spurious low-frequency wobble that shouldn't exist physically. What's a likely explanation, and how would you confirm it?intro

   Answer this one, graded

   Or just show what a strong answer covers

   I'd suspect aliasing: any energy above 500 Hz — say noise from a nearby switching supply or motor — gets folded back into the sampled band because Nyquist was never actually enforced by a real anti-alias filter, it was just satisfied on paper by the sample rate choice. To confirm, I'd probe the raw analog signal directly with a scope or spectrum analyzer before the ADC to look for high-frequency content, or temporarily sample at a much higher rate to see the true spectrum, or add an appropriate low-pass filter ahead of the ADC and check whether the wobble disappears. If it does, that confirms aliasing was the culprit rather than a real physical effect.

   • Identifies aliasing as the cause — States that the wobble is likely aliased energy from content above the Nyquist frequency (500 Hz) folding into the passband, not real physical behavior.
   • Explains the missing piece — Notes that Nyquist rate alone is a theoretical requirement; without an actual anti-aliasing filter ahead of the ADC, out-of-band noise still gets sampled and folded down.
   • Proposes a concrete confirmation method — Suggests something verifiable: probing/scoping the raw analog signal for high-frequency content, oversampling temporarily, or adding a filter and checking if the wobble disappears.

   Where answers go wrong: Blaming ADC quantization noise or software bugs instead of aliasing. · Assuming that sampling at 2x the expected bandwidth automatically guarantees clean data. · Not proposing any way to actually verify the hypothesis.

2. 2You have a photodiode that outputs a few microamps of current proportional to incident light, and you need to feed that signal into an ADC input that expects a voltage between 0 and 3.3V. What circuit would you put between the photodiode and the ADC, and what determines how you'd size the key component in it?intro

   This is a basic front-end design question for turning a current-output sensor into something an ADC can read.

   Answer this one, graded

   Or just show what a strong answer covers

   I'd use a transimpedance amplifier: op-amp with the photodiode connected between the inverting input and ground, and a feedback resistor Rf from output to the inverting input, with the non-inverting input tied to a reference. That converts the diode's current directly into a voltage, Vout = -I*Rf, giving a low-impedance output the ADC can sample cleanly. Sizing Rf is a tradeoff: bigger Rf gives more voltage per unit current, but it interacts with the diode's junction capacitance and any stray capacitance to create a pole that limits bandwidth, and can make the loop unstable without a small feedback capacitor for compensation. I'd also want an op-amp with low input bias current so it doesn't add error current comparable to the diode's own signal, and enough gain-bandwidth product for the frequencies I actually care about.

   • Identifies current-to-voltage conversion need — States that the photodiode is a current source and needs a current-to-voltage stage, not a simple resistor-to-ground pickoff.
   • Proposes transimpedance amplifier — Describes an op-amp with a feedback resistor (photodiode across the inverting input to virtual ground, output = -I*Rf) as the conditioning circuit.
   • Discusses gain/bandwidth tradeoff — Explains that increasing the feedback resistor increases gain but reduces bandwidth (interacting with photodiode/stray capacitance) and can raise noise gain.
   • Mentions op-amp selection considerations — Notes the op-amp needs low input bias current and appropriate GBW so it doesn't distort the tiny signal or become unstable at the chosen gain.

   Where answers go wrong: Suggesting a plain resistor to ground and calling it done without addressing bandwidth or the current-source nature of the photodiode. · Ignoring that a very large feedback resistor for high gain will roll off bandwidth or introduce instability from stray capacitance. · Forgetting the amplifier needs to handle a signal that can be extremely small (nA to uA range), requiring low bias current.

3. 3A strain-gauge full bridge is excited with a 5V supply and has a rated sensitivity of 2 mV/V, so its differential output swings ±10 mV around zero as you apply the full rated load in both directions. Your ADC has a 0–3.3V single-ended input range and you want to use nearly the full range for best resolution. What gain should your instrumentation amplifier provide?easy

   Assume you'll handle centering the bipolar signal into the unipolar ADC range separately — just focus on the gain number needed to use the ADC's full span.

   Answer this one, graded

   Or just show what a strong answer covers

   The full-scale differential swing is ±10mV around zero, so the total peak-to-peak span is 20mV. To use nearly all of the ADC's 3.3V range, gain should be about 3300mV / 20mV ≈ 165. But that's not the whole story — since the bridge output is bipolar and the ADC input is unipolar 0 to 3.3V, I also need to add an offset or level-shift so the amplified signal is centered mid-rail instead of clipping on the negative side. I'd also sanity-check that a gain of 165 doesn't push the amplifier's noise floor or bandwidth limits for this application.

   • Computes full swing correctly — Recognizes the total peak-to-peak swing is 20 mV (±10 mV), not just 10 mV.
   • Computes correct gain — Arrives at gain ≈ 165 by dividing the ADC span (3.3V = 3300mV) by the total signal swing (20mV).
   • Flags the offset/centering issue — Notes that because the bridge output is bipolar and the ADC is unipolar, an offset/level-shift is also needed, not gain alone.

   Where answers go wrong: Using only the ±10mV half-swing instead of the full 20mV peak-to-peak swing, giving a gain off by 2x. · Forgetting that the bipolar sensor signal needs a bias/offset to fit into the ADC's unipolar range. · Ignoring amplifier bandwidth or noise limits when picking a gain this high.

4. 4You calibrate a pressure sensor against a reference standard at two points and it matches within spec. Six months later in the field, its output at zero pressure has shifted by 2% of full scale, but the slope between two known pressures still looks correct. What does this tell you about what happened to the sensor, and does your original two-point calibration still protect you?easy

   Answer this one, graded

   Or just show what a strong answer covers

   Since the slope between the two known points is unchanged but the zero point moved, this is offset drift specifically — likely from aging or temperature exposure — not a gain/span error. A two-point calibration only guarantees accuracy at the time and conditions it was performed under; it doesn't protect against drift that happens afterward. So no, the original calibration doesn't cover this. I'd want either a recalibration schedule based on the observed drift rate, or better, a way to periodically check the sensor against a known zero-pressure reference in the field to correct the offset without doing a full two-point recal every time.

   • Correctly identifies the drift type — States this is offset/zero drift, distinct from gain/span error, because the slope between the two known points is preserved.
   • Explains why the calibration doesn't still protect you — Explains that a calibration certificate/measurement is valid only for the conditions and time at which it was performed, and does not compensate for subsequent drift.
   • Proposes a mitigation — Suggests a concrete fix: periodic recalibration cadence, or an in-field zero-reference check to correct offset without a full recal.

   Where answers go wrong: Conflating this with a gain/span error instead of recognizing it's specifically an offset shift. · Assuming the original two-point calibration remains valid indefinitely. · Not proposing any way to detect or correct the drift going forward.

5. 5You calibrate a thermistor-based temperature sensor at two points: 0C using an ice bath and 100C using boiling water, fitting a straight line between the two raw ADC readings and the known temperatures. Both endpoints check out against your reference thermometer. But when you use the sensor to measure room temperature, around 22C, it reads 1.5C off from the reference. The sensor hasn't drifted and your ice bath and boiling water were good. What's actually going on, and what would you change about the calibration to fix it?easy

   This probes whether the candidate understands what a two-point calibration actually guarantees versus what it doesn't.

   Answer this one, graded

   Or just show what a strong answer covers

   The thermistor's resistance-to-temperature curve isn't a straight line, it's strongly nonlinear, closer to exponential. A two-point calibration fits a straight line that's forced to pass exactly through my two calibration points, so of course it checks out at 0C and 100C -- that's guaranteed by construction, not evidence the sensor is linear. Room temperature sits in the middle of the range where the actual curve bows away from that straight line, so I see an error there even though the endpoints are perfect. To fix it I'd either add more calibration points across the range I actually care about and use a piecewise or polynomial fit, or better, use the sensor's known physical model like the Steinhart-Hart equation to convert resistance to temperature, which tracks the real curve instead of approximating it with a line.

   • Identifies nonlinearity as the root cause — States that a thermistor's response is nonlinear and a straight line between two points will not track the curve in between, producing error at intermediate temperatures.
   • Explains why endpoints still check out — Notes that a linear two-point fit is forced to be exact at the calibration points by construction, so error there tells you nothing about mid-range accuracy.
   • Proposes a concrete fix — Suggests adding more calibration points across the operating range or using a nonlinear model (e.g. Steinhart-Hart, lookup table, polynomial fit) instead of a two-point line.
   • Distinguishes from drift — Explicitly rules out sensor drift or calibration bath error as the cause since the problem is systematic and reproducible at a specific temperature, not a shift over time.

   Where answers go wrong: Blaming sensor drift or aging when the behavior is a repeatable nonlinearity error, not a time-dependent shift. · Assuming a two-point calibration guarantees accuracy everywhere in the range just because both endpoints matched. · Suggesting to simply recalibrate at 0C and 100C again without addressing the underlying curve shape.

6. 6You're reading a sensor through an ADC and the digitized values look noisier than the sensor's datasheet noise spec would predict — but only when a particular motor elsewhere on the board is running. Walk me through how you'd isolate whether this is a signal-conditioning problem, a grounding/EMI problem, or an aliasing problem.medium

   Answer this one, graded

   Or just show what a strong answer covers

   First I'd confirm the correlation is real by toggling the motor on and off a few times and watching the noise appear/disappear, ruling out coincidence. Then I'd probe directly at the sensor output and at the ADC input with a scope while the motor runs, to see if the noise is already present in the analog signal — that tells me it's not a purely digital or software issue. I'd look at the noise's frequency content: if it correlates with the motor's PWM or commutation frequency and that frequency is above my anti-alias filter's cutoff but still shows up in my sampled data, that points to aliasing, and I'd check the actual filter cutoff against the sample rate to confirm. If instead the noise looks like a broadband hash tied to motor current transitions and shows up as a shift in ground potential between the sensor and ADC reference, that points to a ground loop or shared return path, which I'd check by measuring the voltage difference between the two ground points directly. Based on which hypothesis holds up, I'd fix it with either a steeper/lower anti-alias filter, better grounding/layout separation, or differential sensing to reject common-mode noise.

   • Confirms correlation first — Starts by toggling the motor on/off (or load conditions) to verify the noise is truly correlated before diagnosing further.
   • Probes the analog chain directly — Proposes scoping the sensor output and ADC input directly to determine whether the noise is already present in the analog signal versus introduced digitally.
   • Considers EMI/grounding as a candidate cause — Mentions checking for shared ground return paths, ground loops, or radiated/conducted coupling from the motor's wiring or drive electronics.
   • Considers aliasing as a candidate cause — Checks whether the motor's switching frequency is above the anti-alias filter cutoff but still folds into the passband after sampling, and proposes verifying against the filter's actual cutoff and sample rate.

   Where answers go wrong: Jumping straight to 'add more filtering' without diagnosing which stage the noise is entering at. · Not measuring the analog signal directly before or at the ADC input. · Ignoring grounding/layout as a candidate cause and assuming it must be software or ADC-internal noise. · Assuming it's aliasing without checking whether the offending frequency actually falls above the anti-alias filter's cutoff.

7. 7You have an accelerometer and a gyroscope you want to combine to estimate tilt angle. The accelerometer is accurate over long timescales but noisy instantaneously; the gyroscope is smooth and accurate instantaneously but drifts over time. How would you combine them, and what does the crossover frequency of your fusion filter actually represent?medium

   Answer this one, graded

   Or just show what a strong answer covers

   I'd use a complementary filter: take the gyro's integrated angle for fast dynamics, but continuously pull it back toward a low-pass filtered version of the accelerometer-derived angle to remove long-term drift — something like angle = alpha*(angle_prev + gyro*dt) + (1-alpha)*accel_angle. The crossover frequency is essentially where I switch trust from one sensor to the other: below it, I trust the accelerometer because it doesn't drift; above it, I trust the gyro because it isn't noisy for fast motion. Practically it's set by how fast the gyro is expected to drift versus how noisy or vibration-prone the accelerometer is — a lower crossover corrects drift more slowly but rejects more accelerometer noise, and vice versa. One important caveat: if the system undergoes sustained linear acceleration, like a car cornering, the accelerometer's assumed gravity vector is corrupted, so the fusion estimate becomes wrong unless you bring in another reference like a magnetometer or wheel-speed data.

   • Proposes a fusion architecture — Describes combining a low-pass filtered accelerometer angle with a high-pass filtered (or drift-corrected) integrated gyro angle — i.e., a complementary filter or equivalent.
   • Explains crossover frequency's physical meaning — Explains it as the tradeoff point between trusting the accelerometer (near-DC/steady state) and trusting the gyro (fast transients), and connects it to the drift-correction time constant.
   • Identifies the acceleration corruption failure mode — Notes that sustained linear acceleration (e.g. cornering) corrupts the accelerometer's implied gravity vector, causing the fusion estimate to be wrong unless compensated.

   Where answers go wrong: Only integrating the gyro and ignoring drift entirely. · Treating the crossover frequency as an arbitrary tuning knob without connecting it to a physical tradeoff. · Not mentioning that linear acceleration corrupts the accelerometer's tilt estimate.

8. 8You're monitoring vibration on a rotating shaft with an accelerometer sampled at 2 kHz, so your Nyquist frequency is 1 kHz. You expect meaningful vibration content up to about 900 Hz from known mechanical modes. During a test, the FFT of your data shows a strong, narrow peak at exactly 300 Hz that doesn't correspond to any known shaft speed, gear mesh frequency, or bearing defect frequency in your system. What's your leading hypothesis, and how would you confirm it on the bench?medium

   This is a diagnostic question about recognizing and confirming aliasing in real spectral data, not designing the anti-alias filter itself.

   Answer this one, graded

   Or just show what a strong answer covers

   My first suspicion is aliasing, not a real mechanical mode. At 2 kHz sampling the Nyquist limit is 1 kHz, so any real vibration energy above that, say from a bearing defect harmonic or structural resonance up around 1700 Hz, would fold back and show up at 2000 minus 1700 equals 300 Hz, landing right where I see the spurious peak. To confirm it, I'd change the sample rate, for example to 2.5 kHz, and watch what happens: a real 300 Hz mechanical signal stays at 300 Hz regardless of sample rate, but an alias will move to a different apparent frequency because the fold-back point changes with Nyquist. If the peak moves, that confirms aliasing, and the real fix is to add or tighten an anti-alias low-pass filter ahead of the ADC with a cutoff and rolloff steep enough to suppress everything above 1 kHz before it gets sampled, rather than just chasing sample rates.

   • Proposes aliasing as the hypothesis — States that the 300 Hz peak is likely an alias of a real frequency above the 1 kHz Nyquist limit rather than a genuine low-frequency mechanical mode.
   • Computes or reasons about the fold-back relationship — Shows how a frequency above Nyquist maps to an observed lower frequency, e.g. identifying that a real ~1700 Hz component would alias to 300 Hz at 2 kHz sampling.
   • Describes a concrete confirmation test — Proposes changing the sample rate (or adding/adjusting an anti-alias filter) and checking whether the 300 Hz peak shifts, which would prove it's an alias rather than a real signal.
   • Identifies the missing or inadequate anti-alias filter as the fix — Notes that the real fix is a low-pass filter ahead of the ADC with sufficient attenuation above 1 kHz, not just picking a different sample rate after the fact.

   Where answers go wrong: Assuming the 300 Hz peak must be a real, previously-unknown mechanical resonance without questioning the sampling setup. · Proposing to just notch-filter out the 300 Hz peak in software without investigating whether it's an alias corrupting the whole spectrum. · Failing to propose a concrete experimental test, offering only a definition of aliasing instead of a way to confirm it on this system.

9. 9You need to digitize a sensor signal with real content up to 200 Hz, sampled by an ADC running at 1 kHz (Nyquist = 500 Hz). You start with a single-pole RC anti-aliasing filter set at a -3dB cutoff around 200-300Hz. Roughly how many poles (i.e., what order of filter) would you actually need to suppress content at the Nyquist frequency by 60dB relative to the passband, given the narrow transition band between 200Hz and 500Hz? Is a single RC pole sufficient, and if not, what would you change about the overall approach?hard

   A single real pole rolls off at roughly 20 dB/decade. You need to go from -3dB near 200-300Hz down to about -60dB by 500Hz.

   Answer this one, graded

   Or just show what a strong answer covers

   The gap between the filter cutoff (~200-300Hz) and Nyquist (500Hz) is less than a decade — roughly log10(500/250) ≈ 0.3 decades. To get from -3dB down to about -60dB in that narrow a band, I need something like 60dB / 0.3 decades ≈ 200dB/decade of roll-off, versus the ~20dB/decade a single RC pole gives. That means I'd need roughly a 8th-to-10th order filter, which is impractical to build as a passive single-stage RC network — a single pole is clearly not sufficient here. In practice, I wouldn't try to force a huge passive filter order. Instead I'd oversample at a much higher rate — say tens of kHz — so the effective Nyquist frequency is far above my noise sources, letting a much gentler analog anti-alias filter do the coarse job, and then digitally decimate/filter down to the final 1kHz-equivalent output rate. This is exactly what sigma-delta ADCs do internally, and using one would likely be the cleanest real-world solution rather than hand-designing a steep multi-pole analog filter.

   • Computes the frequency ratio / available transition band — Recognizes the ratio between Nyquist (500Hz) and the filter cutoff (~200-300Hz) is small (roughly 1.7-2.5x), giving less than half a decade of transition band.
   • Estimates required order/roll-off correctly — Estimates that reaching ~60dB attenuation within that narrow transition band requires a roll-off far steeper than 20dB/decade, implying a filter order in roughly the 6-10 pole range (or explicitly computes ~150dB/decade needed).
   • Concludes single RC pole is insufficient — States clearly that a single-pole RC filter cannot meet this requirement given the narrow gap between passband and Nyquist.
   • Proposes a realistic alternative — Suggests a practical fix: a much higher sample rate (oversampling) to relax the filter requirement, a sigma-delta ADC with built-in decimation filtering, or a multi-pole active filter, rather than trying to brute-force a passive single-pole design.

   Where answers go wrong: Assuming a single RC pole is 'good enough' because it's at the right cutoff frequency, without checking the actual attenuation at Nyquist. · Forgetting that roll-off is per decade and the available frequency range here is much less than a decade, making steep attenuation very hard to achieve. · Not proposing oversampling/sigma-delta as the realistic practical solution, and instead trying to force an unrealistic passive filter order.

10. 10You have two temperature sensors mounted near the same point on a board, a thermocouple and an RTD, each independently calibrated against a traceable reference and each verified accurate to within their stated tolerance. At steady state, running the same test, they disagree by 3C. Neither is out of spec on its own calibration certificate. How do you figure out what's actually going on, and how would you combine or reconcile their readings for a single trustworthy temperature value?hard

    This is a systems-level question about reconciling two independently-correct measurements that still disagree, common in DAQ systems using redundant sensors.

    Answer this one, graded

    Or just show what a strong answer covers

    The fact that both sensors pass their own calibration doesn't mean they agree with each other, because calibration only tells you each sensor is accurate against a reference under the conditions it was calibrated in, not that the two are sensing the identical physical point under these operating conditions. I'd look for real physical causes before assuming either sensor is simply wrong: a thermal gradient across the board if they're not mounted at exactly the same spot, different thermal mass or response time causing one to lag the other, and specifically for the RTD, self-heating from its excitation current, which is a well-known source of a small but real positive offset. To investigate, I'd try to get both sensors physically as close to the same point as possible, watch the transient response of each to see if one lags, and try reducing the RTD excitation current to see if its reading shifts, which would confirm self-heating as a contributor. Only after understanding the physical source of the 3C gap would I decide how to reconcile them, whether that's correcting the RTD reading, trusting one sensor as the reference for this specific application, or budgeting the disagreement as combined uncertainty -- blind averaging without that investigation would just hide a real physical issue.

    • Rejects the assumption that calibration guarantees agreement — Explicitly states that both sensors passing calibration against a reference does not guarantee they read the same value in a different physical setup or location.
    • Proposes specific physical causes — Names at least two plausible physical causes such as thermal gradient between the two sensor locations, self-heating from RTD excitation current, or differing thermal response time/mounting.
    • Describes an investigation method — Proposes a concrete way to isolate the cause, such as moving both sensors to the exact same physical point, checking response over a transient versus steady state, or varying excitation current to see if the RTD reading shifts.
    • Avoids naive averaging as the final answer — Explains why simply averaging the two readings is not a sound reconciliation strategy without first understanding and budgeting the source of disagreement.

    Where answers go wrong: Assuming one sensor must simply be wrong since the calibration certificates say both are within spec. · Jumping straight to averaging the two readings as the fused value without investigating the root cause first. · Ignoring self-heating effects from the RTD's excitation current as a plausible source of a real, physical temperature offset. · Failing to consider that the two sensors may not actually be sensing the same physical point due to mounting or thermal gradient.