2. Data acquisition in industrial settings

Anomaly detection models are only as good as the data they are trained and evaluated on. In industrial environments, data acquisition is not a matter of pointing a camera at a process and recording — it is an engineering discipline shaped by sensor physics, timing constraints, network topology, and the reality that labeled faults are rare, expensive, and sometimes dangerous to produce. This chapter examines the full acquisition pipeline: what sensors are used, how fast they must sample, how their signals are aligned in time, why labels are scarce, and how to assemble a dataset that supports unsupervised learning.

Sensor Types and What They Measure

Industrial condition monitoring typically draws on several sensor families, each with distinct physical principles and failure signatures:

• Vibration (accelerometers, IEPE/piezoelectric): Captures rotating machinery faults — bearing defects, misalignment, looseness, gear mesh problems. Mounted radially and axially on bearing housings; sensitivity commonly 100 mV/g, bandwidth up to 10–20 kHz.
• Acoustic emission (AE): High-frequency (100 kHz–1 MHz) elastic waves from crack propagation, friction, or leakage. Used for slow-rotating bearings and pressure vessels.
• Temperature (RTD, thermocouple, IR): Slow thermal trends indicating overload, lubrication failure, or insulation degradation. Pt100 RTDs give ±0.1 °C stability.
• Pressure and flow (piezoresistive, differential pressure, Coriolis): Detect blockage, cavitation, valve degradation, and pump efficiency loss.
• Current and voltage (CT, Rogowski coil, Hall): Motor current signature analysis (MCSA) reveals rotor bar faults and load anomalies without intrusive sensors.
• Proximity and displacement (eddy current, LVDT): Shaft orbit and axial position in turbomachinery.
• Process and quality sensors: Torque, force, vision (2D/3D), and inline metrology tied directly to product quality.

The choice of sensor determines the frequency content you can observe, and therefore the sampling rate you must configure.

Sampling Rate and Anti-Aliasing

The Nyquist–Shannon theorem requires a sampling frequency f_s greater than twice the highest frequency of interest. In practice, industrial acquisition uses 2.5× to 10× oversampling plus an analog anti-aliasing filter. For a bearing with a 3 kHz defect frequency, a minimum of 6.4 kS/s is required; 25.6 kS/s is common to preserve envelope and sideband structure.

Different signals demand different rates, which is why industrial systems are almost always multi-rate:

Two practical pitfalls deserve attention. First, aliasing: if a 15 kHz resonance is sampled at 10 kS/s, it folds to 5 kHz and appears as a false feature. Always verify the analog front-end filter cutoff. Second, resolution vs. rate trade-off: a 16-bit ADC at 100 kS/s may only deliver 14 effective bits; check ENOB (effective number of bits) rather than the marketing figure.

Synchronization Across Channels and Devices

Anomalies often manifest as a relationship between channels — a vibration spike coinciding with a current transient, or a pressure drop aligned with a valve command. If channels drift by even a few milliseconds, that relationship is destroyed.

Three synchronization tiers are common:

1. Shared-clock (best): All channels sampled by the same ADC or a synchronized ADC bank. Jitter < 1 sample. Used in portable analyzers and compact DAQ chassis.
2. Hardware-triggered distributed: Multiple devices share a trigger line and a common timebase (e.g., IEEE 1588 PTP over Ethernet, or IRIG-B). Achievable skew: tens of microseconds to a few milliseconds.
3. Software timestamping (weakest): Each device stamps its own samples using host clocks. Skew can reach tens of milliseconds and drift over hours. Acceptable only for slow process variables.

For rotating machinery, an additional reference is essential: a tacho or keyphasor pulse once per revolution. This enables order tracking — resampling the vibration signal in the angle domain so that speed fluctuations do not smear bearing and gear frequencies.

Labeling Scarcity: Why It Exists and What to Do

Supervised fault classification needs labeled examples of each failure mode. In industry, those labels are scarce for structural reasons:

• Faults are rare by design. Maintenance aims to prevent them, so a healthy machine may run for years without a single labeled failure.
• Producing faults is destructive. Running a bearing to failure on a production asset risks collateral damage and lost output.
• Labels are ambiguous. A "fault" may be a gradual degradation with no clear onset; different engineers may disagree on when it began.
• Operating context varies. The same vibration signature means different things at different loads, speeds, and temperatures.

This is precisely why unsupervised and semi-supervised methods dominate industrial anomaly detection. The model learns the distribution of normal behavior from abundant healthy data and flags deviations. Labels, when available, are used for validation and threshold tuning rather than for training a classifier from scratch.

Building a Dataset That Supports Unsupervised Learning

A well-constructed dataset for unsupervised anomaly detection has four properties: it captures normal variability, it is temporally consistent, it is documented, and it includes enough anomalous events for honest evaluation.

1. Capture normal variability deliberately. Record across the full operating envelope — startup, shutdown, idle, part-load, full-load, seasonal temperature swings, and product changeovers. A model trained only on steady-state full-load data will flag every startup as an anomaly.

2. Preserve temporal structure. Store raw time series with accurate timestamps, not just aggregated features. Downsampling or averaging destroys the transient signatures that matter most. Keep the raw data and derive features separately.

3. Document metadata. Every recording should carry: asset ID, sensor ID and mounting location, sampling rate, units, calibration date, operating condition, and any known events. Without this, a dataset becomes unusable six months later.

4. Reserve anomalies for testing. Even a handful of confirmed fault records — from a test rig, a planned shutdown inspection, or a historical incident — should be held out as a test set. They are too valuable to consume during training.

A practical directory and metadata convention helps:

dataset/
  asset_A/
    normal/
      2024-03-01T08-00_run1_load75/
        meta.json          # asset, sensors, fs, units, condition
        vib_de.parquet     # drive-end accelerometer, 25.6 kS/s
        vib_nde.parquet    # non-drive-end, 25.6 kS/s
        current.parquet    # 10 kS/s
        tacho.parquet      # 1 pulse/rev
    fault/
      2024-05-12_bearing_outer_race/
        meta.json
        ...

Storing each channel as a columnar file (Parquet) keeps I/O efficient and preserves dtypes and timestamps. A meta.json per recording makes the dataset self-describing.

From Acquisition to Model Input

The figure below summarizes how raw multi-rate, multi-sensor signals flow through synchronization, windowing, and feature extraction into a form suitable for unsupervised anomaly detection.

Practical Checklist for a New Acquisition Campaign

Before recording a single sample, confirm the following. This checklist is written to be copied into a commissioning document.

With a synchronized, well-documented, multi-rate dataset in hand, the next challenge is deciding what "normal" looks like statistically — which is the subject of the following chapters on feature representation and density estimation.