2. Sensor data acquisition for industrial equipment

Anomaly detection begins long before any model is trained. It begins at the transducer. The physical quantity a sensor converts into an electrical signal, the frequency range it preserves, the rate at which that signal is digitized, and the integrity of the resulting samples together determine what anomalies are even detectable. A bearing defect whose signature lives at 4 kHz is invisible in a dataset sampled at 1 kHz, no matter how sophisticated the downstream algorithm. This chapter examines the six sensor families most common in industrial condition monitoring, the sampling decisions that govern each, and the data quality problems that quietly corrupt detection pipelines.

2.1 The Six Sensor Families and What They Measure

Vibration sensors are the workhorse of rotating machinery monitoring. Piezoelectric accelerometers convert acceleration into charge or voltage, typically covering 0.5 Hz to 10 kHz or beyond. They capture bearing defect frequencies, gear mesh harmonics, imbalance, misalignment, and looseness. Mounting matters enormously: a stud-mounted sensor preserves high-frequency content, while a magnetic mount attenuates it above roughly 2–3 kHz and introduces resonance of its own.

Acoustic emission (AE) sensors operate in a much higher band, commonly 100 kHz to 1 MHz. They detect stress waves released by crack propagation, friction, and incipient surface damage — phenomena that produce little low-frequency vibration until damage is advanced. AE is well suited to slow-rotating equipment and to crack growth monitoring, but it is highly sensitive to background noise from grinding, impact, and flow.

Temperature sensors — RTDs, thermocouples, and infrared pyrometers — are slow but indispensable. They reveal overheating bearings, lubrication breakdown, and electrical faults. Their bandwidth is low (often <1 Hz effective), so they are sampled at rates of 1 Hz or less and are usually treated as trend signals rather than waveforms.

Pressure sensors capture hydraulic and pneumatic system behavior, combustion dynamics, and pump discharge conditions. Piezoresistive and piezoelectric types differ sharply: piezoresistive devices measure static and dynamic pressure, while piezoelectric types respond only to changes. Choosing the wrong type for a static-pressure monitoring task yields a signal that drifts to zero.

Current sensors — current transformers, Hall-effect probes, and Rogowski coils — enable motor current signature analysis (MCSA). They detect rotor bar faults, stator winding asymmetry, and load anomalies without touching the machine. Because electrical fault signatures appear as sidebands around the supply frequency and its harmonics, current signals often need bandwidth into the low kilohertz range.

Oil debris sensors monitor wear particles in lubricant. Magnetic chip detectors, inductive particle counters, and optical sensors quantify particle count and size. They are effectively event-driven and low-bandwidth, but they provide direct evidence of metallic wear that vibration may only hint at.

2.2 Sampling Rate, Bandwidth, and Aliasing

The Nyquist–Shannon theorem states that a signal must be sampled at more than twice its highest frequency component to be reconstructed without aliasing. In practice, engineers sample well above this minimum. A common rule is to sample at 2.56× the desired analysis bandwidth, which conveniently yields a power-of-two number of samples for FFT processing.

Consider a motor bearing whose outer-race defect frequency is approximately 3.2 kHz. To capture that defect and its first few harmonics, an analysis bandwidth of 10 kHz is reasonable. Applying the 2.56 rule gives a sampling rate of 25.6 kHz, commonly rounded to 25.6 kS/s or 25.6 kHz. At that rate, a 1-second record contains 25,600 samples — a manageable size for both storage and FFT computation.

Example sampling configuration (illustrative):
  Sensor:        IEPE accelerometer, 100 mV/g, 0.5 Hz–10 kHz
  Analysis BW:   10 kHz
  Sampling rate: 25.6 kS/s  (2.56 × 10 kHz)
  Record length: 1.0 s  →  25,600 samples
  FFT lines:     12,800  (Nyquist-limited)
  Anti-alias:    analog low-pass at 10 kHz, ≥ 8th-order
  Window:        Hann for general spectra; flat-top for amplitude accuracy

Two details are frequently overlooked. First, the anti-aliasing filter must be analog and placed before the ADC; digital filtering after sampling cannot remove frequencies that have already folded into the baseband. Second, the sampling clock itself must be stable. Jitter in the sample timing smears high-frequency peaks and degrades envelope analysis, which is precisely the technique used for bearing diagnostics.

Different sensor types demand different rates. Temperature and oil debris channels are typically sampled at 1 Hz or on an event basis. Pressure and current channels may be sampled anywhere from 1 kS/s to 50 kS/s depending on whether the goal is trend monitoring or spectral fault detection. A well-designed DAQ system therefore supports heterogeneous sampling rates rather than forcing every channel to the highest rate.

2.3 Data Quality Issues That Break Detection

Even a correctly specified sensor and sampling rate can produce unusable data. The most common failure modes are worth naming explicitly.

• Aliasing: high-frequency content folding into the band of interest, creating phantom peaks that a detector may flag as a fault. Symptom: spectral peaks that shift position when the sampling rate changes.
• Sensor drift and DC offset: slow baseline movement that masks low-frequency trends. Symptom: a rising DC level in an AC-coupled channel that should sit near zero.
• Dropouts and gaps: missing samples from network congestion, buffer overruns, or power interruption. Symptom: irregular timestamps and discontinuities in the waveform.
• Clipping and saturation: signal exceeding the ADC range, producing flat-topped waveforms and spurious harmonics. Symptom: repeated maximum-value samples.
• Ground loops and electrical noise: 50/60 Hz hum and its harmonics injected into low-level signals. Symptom: sharp peaks at the mains frequency regardless of machine state.
• Incorrect scaling or units: a channel configured in mV/g but recorded as g, or vice versa. Symptom: amplitudes off by a constant factor across all records.
• Timestamp misalignment: channels sampled by different clocks, so vibration and current records cannot be correlated. Symptom: phase relationships that drift over time.

2.4 A Practical Acquisition and Validation Workflow

The following workflow is a reasonable starting point for a new monitoring installation. It is written as a sequence of concrete checks rather than abstract stages.

1. Define the fault frequencies of interest. For a bearing, compute BPFO, BPFI, BSF, and FTF from the geometry and shaft speed. For a gearbox, identify gear mesh frequencies and their sidebands. These numbers set the required analysis bandwidth.
2. Select sensors and mounting. Match the sensor's frequency range to the fault frequencies. Use stud mounting where high-frequency content matters; document the mounting method, because it changes the effective bandwidth.
3. Set the sampling rate and anti-alias filter. Apply the 2.56 rule, verify the analog filter cutoff, and confirm the ADC resolution (16-bit is typical; 24-bit for low-level AE or current signals).
4. Record a baseline under known-good conditions. Capture at least several minutes of data at steady load and speed. This baseline becomes the reference for later comparison.
5. Validate the raw data. Check for clipping, dropouts, mains hum, and correct scaling. Plot the time waveform and spectrum before any processing.
6. Document metadata. Record sensor model, sensitivity, mounting, sampling rate, filter settings, load, speed, and ambient conditions alongside every file.

Step 5 deserves emphasis. A quick visual inspection of the time waveform and spectrum catches most acquisition problems before they propagate into a model. If the spectrum shows an unexpected peak at exactly the mains frequency, the problem is electrical, not mechanical. If the waveform is flat-topped, the gain is too high. If timestamps are irregular, the acquisition chain — not the machine — is the issue.

2.5 Signal Chain Overview

The diagram below traces the path from physical phenomenon to validated digital record, showing where each sensor family enters the chain and where the most consequential data quality risks arise.

2.6 Key Takeaways

Sensor selection and sampling configuration are not preliminary formalities — they define the ceiling of what any anomaly detector can achieve. Match the sensor's frequency range to the fault physics, apply the 2.56× sampling rule with a genuine analog anti-alias filter, and validate every record for clipping, dropouts, noise, and scaling before it enters a pipeline. Document metadata alongside the data, because a spectrum without its sampling rate, sensitivity, and mounting method is not interpretable. When acquisition is done correctly, the downstream detection problem becomes tractable; when it is not, no algorithm can recover the information that was never captured.