Data Acquisition and Engineering

Introduction

Artificial intelligence in mechatronics depends on data that accurately represent physical processes. A machine-learning model cannot compensate reliably for a sensor installed in the wrong location, an incorrectly calibrated pressure transducer, unsynchronized measurements, excessive noise, or labels that do not correspond to the actual machine state. Data acquisition and engineering therefore form the bridge between a physical mechatronic system and an intelligent computational system.

A modern mechatronic installation may combine encoders, accelerometers, thermocouples, current sensors, force cells, cameras, programmable logic controllers (PLCs), robots, and industrial networks. These devices produce heterogeneous data at different rates, with different units, delays, accuracies, and failure modes. The engineering task is to transform these raw observations into trustworthy, time-aligned, documented, and analysis-ready datasets.

The data lifecycle considered in this chapter is:

physical process → sensing → signal conditioning → conversion → synchronization → storage → quality assessment → labeling → feature engineering → exploratory analysis → model development and deployment

The lifecycle is iterative. Discovering that a model performs poorly may reveal a sampling problem, a calibration error, data leakage, or an unrepresented operating condition rather than an inadequate algorithm.

Learning Objectives

By the end of this chapter, a student should be able to:

1. Design a data pipeline for sensors, actuators, and industrial systems.
2. Select sampling rates and synchronization methods for dynamic measurements.
3. Explain calibration, uncertainty, filtering, and signal conditioning.
4. Assess missing, duplicated, inconsistent, noisy, and implausible data.
5. Construct labels and features without introducing leakage.
6. Use exploratory analytics to identify relationships, regimes, and anomalies.
7. Evaluate an acquisition design in terms of technical performance, cost, safety, and governance.

1. Data Acquisition Architecture

1.1 From physical variables to digital records

A sensor converts a physical quantity into an electrical or digital signal. For example, an accelerometer measures acceleration, a resolver measures angular position, and a thermocouple produces a voltage related to temperature. The complete acquisition chain commonly contains:

• Transducer: converts a physical variable into an electrical quantity.
• Signal conditioning: provides amplification, excitation, isolation, linearization, and anti-alias filtering.
• Data converter: an analogue-to-digital converter (ADC) produces numerical samples.
• Time base: assigns timestamps or sample indices.
• Communication layer: transfers data through a fieldbus, Ethernet, wireless link, or storage interface.
• Data system: stores raw data, metadata, quality flags, and derived variables.

A digital sample is not the physical variable itself. It is an estimate conditioned by sensor sensitivity, quantization, noise, timing, and calibration. Each record should therefore contain more than a value: it should identify the sensor, unit, timestamp, operating context, calibration version, and quality status.

1.2 A reference pipeline

Figure: A closed-loop data acquisition pipeline for intelligent mechatronic systems.

1.3 Raw, processed, and derived data

Raw data are values as received from the acquisition system, retained in their original form. Processed data have undergone operations such as unit conversion, resampling, filtering, or correction. Derived data include features, labels, health indicators, and model outputs. Raw data should normally be preserved because later investigations may require a different filter, a new calibration, or an audit of the original measurement.

A robust data record commonly includes:

2. Sampling and Reconstruction

2.1 Sampling rate

If a continuous signal contains no frequency components above f_max, ideal uniform sampling requires a sampling frequency satisfying the Nyquist condition:

Here f_s is the sampling frequency in hertz and f_max is the highest frequency component of interest in hertz. This condition assumes band-limited input, accurate timing, and an appropriate anti-alias filter. Equality is not sufficient in practical systems because filter transition bands and noise require margin.

When a signal is sampled below the required rate, higher-frequency content can appear as false lower-frequency content. This phenomenon, aliasing, cannot generally be removed after sampling. An analogue low-pass anti-alias filter must attenuate frequencies above the intended band before the ADC.

Sampling should be selected from system dynamics, not from a convenient default. A vibration-monitoring channel may require thousands of samples per second, while a slowly changing tank temperature may require only one sample every few seconds. The engineer should consider:

• bandwidth of the physical phenomenon;
• control-loop and actuator dynamics;
• transient duration;
• required fault-detection latency;
• sensor and ADC limitations;
• storage and communication capacity;
• synchronization with other channels.

For a periodic phenomenon with period T, the frequency is:

where f is in hertz and T is in seconds. For example, a shaft rotating at 1,800 revolutions per minute has a rotational frequency of 30 Hz. If harmonics up to the fifth order are relevant, the highest frequency of interest is 150 Hz. A theoretical minimum sampling rate is therefore greater than 300 samples per second; an engineering design might choose 1 kHz after considering filtering, transients, and implementation constraints.

2.2 Quantization and resolution

An ideal ADC with N bits and input range from V_min to V_max has an approximate quantization step:

The variable q is in volts when the input limits are in volts. This expression assumes a uniform ideal converter; real converters also exhibit offset, gain error, nonlinearity, noise, and effective-number-of-bits limitations. A signal variation much smaller than the effective resolution may not be measured reliably.

2.3 Synchronization and latency

Measurements from separate devices may have different clocks, timestamp conventions, communication delays, and sampling periods. Synchronization is essential when calculating relationships such as force versus position, current versus torque, or command versus response.

Common strategies include:

• a shared hardware trigger;
• a common clock distributed through an industrial network;
• timestamping at the point of measurement;
• software alignment using known events or interpolation;
• encoder-based angular synchronization for rotating machinery.

Clock drift can accumulate over time. If two clocks differ by a fractional rate error δ, the timing error after duration t is approximately:

where Δt and t have units of time and δ is dimensionless. A 20 parts-per-million clock mismatch produces approximately 72 ms of relative drift over one hour. This may be insignificant for temperature monitoring but serious for high-speed sensor fusion.

Timestamping should distinguish event time—when a physical event occurred—from arrival time—when a record reached a computer. Network delay must not be mistaken for process dynamics. Latency budgets should include sensor response, conversion, buffering, transmission, computation, and actuation.

3. Calibration and Measurement Uncertainty

3.1 Calibration concepts

Calibration compares an instrument with a reference of known accuracy and establishes the relationship between indicated and reference values. It does not mean that every measurement becomes exact. Calibration may estimate offset, sensitivity, nonlinearity, hysteresis, temperature dependence, and cross-axis effects.

For a simple linear sensor, a corrected engineering value can be represented as:

where x is the sensor output, y is the physical quantity, a is the scale factor, and b is the offset. Units must be consistent: if x is volts and y is newtons, then a has units N/V and b has units N.

A calibration procedure should define the reference, range, environmental conditions, number and distribution of test points, acceptance criteria, date, operator, and calibration version. In safety-critical applications, traceability to recognized standards and documented uncertainty are essential.

3.2 Static and dynamic calibration

Static calibration evaluates response to slowly applied or steady inputs. Dynamic calibration evaluates amplitude and phase response over frequency or during transients. A sensor can be accurate in static conditions but unsuitable dynamically because of bandwidth, resonance, damping, or delay.

Sensor placement is also part of calibration quality. An accelerometer mounted loosely may measure mounting resonance rather than machine vibration. A thermocouple installed near a hot surface but exposed to airflow may not represent the process temperature. Calibration therefore includes installation, orientation, cabling, grounding, and environmental conditions.

3.3 Uncertainty and error

Measurement error is the difference between a measured value and a reference or accepted value. Uncertainty expresses incomplete knowledge about that value. Sources include sensor repeatability, reference uncertainty, resolution, environmental effects, drift, and model assumptions.

For independent quantities with standard uncertainties u₁, u₂, …, a commonly used root-sum-square estimate is:

This formula assumes the uncertainty components are expressed in the same units and are sufficiently independent. Correlated components require covariance terms and cannot safely be combined by this simple expression.

4. Signal Conditioning and Filtering

4.1 Sources of unwanted variation

Noise may arise from thermal effects, electromagnetic interference, quantization, mechanical vibration, grounding problems, switching devices, or communication errors. Not every variation is noise: a rapid change may be a real process event. Filtering must therefore be based on process knowledge and measurement objectives.

Good acquisition practice includes shielded and appropriately routed cables, differential measurement where suitable, grounding designed to avoid loops, electrical isolation, sensor excitation within specification, and physical separation from high-current switching paths.

4.2 Digital filters

A moving-average filter of window length M is:

Here x[n] is the input sequence, y[n] is the filtered sequence, and n is the sample index. The filter reduces high-frequency variation but introduces delay and can blur short events. The effective time window is approximately M/f_s seconds for sampling frequency f_s.

A first-order exponential filter can be written as:

where 0 < α ≤ 1. A larger α responds faster and smooths less; a smaller value smooths more but increases lag. The initial condition y[0] must be specified. Filtering should be fitted or selected using training data only when its parameters are learned from data, and the same documented operation must then be applied to validation and deployment data.

Filtering can create artificial evidence. A non-causal, zero-phase filter may be appropriate for offline analysis but cannot be implemented in real time without future samples. It should not be used to claim real-time performance. Excessive smoothing can also remove the very fault signature that a diagnostic model is intended to detect.

4.3 Outliers and missing values

An outlier may represent a genuine shock, a sensor fault, a unit error, or a communication problem. Automatic deletion is unsafe. A useful procedure is to combine range checks, rate-of-change checks, physical constraints, redundant sensors, and maintenance records. Every replacement or removal should generate a quality flag and be recorded in provenance metadata.

For short gaps, interpolation may be acceptable for visualization or selected analytical tasks. Linear interpolation between values x₀ and x₁ at times t₀ and t₁ is:

This assumes a roughly linear change over the gap. It is inappropriate for abrupt events, long outages, or safety decisions where the missing condition is unknown.

5. Data Quality Assessment

Data quality is multidimensional. A dataset may be complete but inaccurate, precise but poorly synchronized, or statistically clean but physically impossible. Important dimensions include:

For a defined interval, a basic completeness rate is:

where both counts refer to the same channel, time interval, and expected sampling rule. The result is dimensionless and often reported as a percentage. Completeness alone does not establish correctness.

Quality checks should occur at acquisition, ingestion, storage, and analytical stages. Examples include schema validation, unit validation, timestamp ordering, duplicate detection, sensor saturation detection, impossible-state rules, cross-channel consistency, and monitoring of drift in distributions.

A data contract is a formal agreement describing field names, types, units, allowed ranges, timestamp semantics, quality flags, update frequency, and change-control procedures. It prevents silent changes from breaking downstream models.

6. Labeling and Feature Engineering

6.1 Labels and ground truth

A label is the target assigned to an observation or time interval, such as normal operation, bearing fault, collision risk, or completed assembly. Labels may come from expert inspection, maintenance records, controller states, test fixtures, or automatically generated events. Each source has limitations. A maintenance log may record the repair date rather than the true onset of degradation; a controller alarm may reflect a threshold rather than a confirmed fault.

Label definitions must specify the event, time boundaries, evidence standard, and treatment of uncertainty. When annotators disagree, disagreement should be retained or adjudicated systematically rather than hidden. In industrial data, the absence of a fault label does not necessarily mean that the machine was healthy.

Avoid data leakage, in which information unavailable at prediction time enters the features or labels. Examples include using a post-failure maintenance code to predict the failure that caused it, randomly splitting overlapping windows from the same run into training and test sets, or normalizing the entire dataset before the test period is isolated.

6.2 Feature engineering

A feature is a measurable representation used for analysis or modeling. Features should reflect system physics and the prediction task. Typical time-domain features include mean, variance, root-mean-square (RMS), peak, crest factor, and rate of change.

For samples x₁, x₂, …, x_N, the arithmetic mean is:

The population-style variance over the selected window is:

If the window is treated as a sample used to estimate a larger population, the sample variance uses N−1 in the denominator. The choice must be documented. RMS is:

RMS has the same units as x. For a sinusoid with zero mean, RMS equals peak amplitude divided by √2, but this special relationship does not apply to arbitrary waveforms.

Frequency-domain features may include dominant frequency, band power, spectral centroid, and harmonic amplitudes. They require an appropriate sampling rate, window length, detrending choice, and treatment of spectral leakage. Time-frequency methods are useful when machine behavior changes during a cycle.

Features can also be relational: motor current divided by speed when physically justified, tracking error between command and position, or temperature difference across a component. Derived ratios require careful handling near zero denominators and clear units. Feature engineering should be guided by mechanics, control theory, and failure modes rather than by blind generation of many variables.

7. Exploratory Analytics

Exploratory data analysis (EDA) is the systematic examination of data before formal modeling. It should be performed separately on training and held-out data when it could influence modeling decisions. Useful activities include:

• plotting signals against time and operating mode;
• examining histograms and empirical quantiles;
• comparing normal and abnormal cycles;
• calculating correlations while remembering that correlation is not causation;
• inspecting scatter plots and lagged relationships;
• examining frequency spectra and spectrograms;
• measuring missingness by sensor, shift, product, and machine;
• checking feature distributions for drift;
• identifying clusters of operating regimes.

The Pearson correlation coefficient between paired variables x and y is:

It ranges from −1 to 1 when both variables have nonzero variation. The coefficient measures linear association under the selected observations; it does not prove a causal relationship and can be distorted by outliers, mixing of operating regimes, or time trends.

Visualization is an engineering diagnostic, not merely a presentation step. A plot may reveal a sensor stuck at a constant value, a periodic communication dropout, a sudden calibration shift, or a distinction between product recipes. EDA should lead to testable hypotheses and documented pipeline changes.

8. Applied Case: Predictive Monitoring of a Conveyor Drive

The following is an illustrative case, not a report of a specific company.

Background

A manufacturing cell uses a variable-speed conveyor driven by an induction motor and gearbox. Unplanned gearbox failures stop production. The engineering team proposes a model that identifies developing mechanical problems early.

Problem

The initial dataset contains motor current at 1 kHz, motor speed at 100 Hz, gearbox vibration at 5 kHz, temperature at 1 Hz, controller commands, and maintenance records. The channels use separate clocks. Several weeks contain missing vibration packets, and maintenance records identify gearbox replacement dates but not the exact onset of damage.

Analysis

The team first defines the prediction task as: “classify each operating cycle as normal or requiring inspection at least 30 minutes before a confirmed failure.” This definition prevents the ambiguous target of predicting a repair record at its timestamp. Cycles are grouped by production run, not randomly by individual windows, so adjacent windows from the same run cannot appear in both training and testing.

A shaft speed of 1,800 rpm corresponds to 30 Hz. Since gearbox harmonics and sidebands up to approximately 600 Hz are of interest, the 5 kHz vibration channel provides practical bandwidth margin after analogue anti-alias filtering. The speed channel is synchronized to vibration using hardware timestamps and resampled only after alignment. Temperature is retained as a slow contextual variable rather than artificially interpolated across long outages.

Calibration records reveal that the vibration sensor was replaced midway through the period. Data are separated by calibration version, and a controlled test is used to estimate whether the two sensors have comparable sensitivity. The pipeline retains raw acceleration, calibrated acceleration, filter settings, and quality flags.

Features are calculated over one-second windows: RMS acceleration, peak acceleration, selected band powers, motor-current RMS, speed, temperature slope, and tracking error. Windows with saturation, missing timestamps, or severe packet loss are excluded from model training but counted in a quality report. Maintenance notes and post-failure measurements are not used as predictive features.

Approach and solution

A baseline rule based on vibration RMS is established before machine learning. It provides interpretability and tests whether the proposed features contain useful signal. A supervised classifier is then trained using time-ordered or run-based validation. Class imbalance is reported using precision, recall, and the false-alarm burden rather than accuracy alone. Threshold selection reflects the operational cost of missed failures, unnecessary inspections, and delayed production.

The deployment system produces a health score, the contributing features, data-quality status, and an escalation recommendation. A score is suppressed when synchronization or sensor validity is inadequate. Engineers review alerts and record whether inspection confirms a fault, creating a feedback loop for label improvement.

Outcome

The analysis demonstrates that the largest initial improvement comes from correcting timestamp alignment and separating sensor calibration periods, not from replacing the classifier. The resulting monitoring system is accepted as an inspection-support tool rather than an autonomous safety controller. Its documentation specifies operating limits, retraining triggers, alarm ownership, and procedures for sensor replacement.

Learning points

1. The prediction target must be defined operationally and temporally.
2. Data splitting must respect machine runs and time order.
3. Calibration and synchronization are model-performance issues.
4. A simple physical baseline is valuable.
5. Quality flags and abstention can be safer than forcing a prediction.
6. Industrial success includes maintainability, interpretability, and response procedures.

9. Professional, Ethical, and Governance Considerations

Data acquisition can affect worker safety, equipment availability, and accountability. A model should not silently control a hazardous actuator merely because its validation score is high. Safety functions require appropriate deterministic safeguards, fail-safe behavior, verification, and compliance with applicable machinery, electrical, functional-safety, cybersecurity, and data-protection requirements.

Industrial datasets may reveal production rates, operator behavior, proprietary recipes, or maintenance practices. Access control, encryption where appropriate, retention limits, audit trails, and role-based permissions should be designed from the beginning. Data minimization is important: collecting every available signal increases attack surface and maintenance cost.

Models may perform differently across machines, shifts, products, sensor replacements, and environmental conditions. Governance should define who owns labels, who approves pipeline changes, how incidents are investigated, and when a model is retired. Version control should cover data schemas, calibration files, code, feature definitions, model parameters, and deployment configuration.

10. Contemporary Developments and Best Practices

Edge computing permits filtering, feature extraction, and rapid anomaly detection near the machine, reducing bandwidth and latency. Cloud or plant-level systems support longer-term storage, fleet comparisons, and model management. Digital twins can provide simulated data and physics-based constraints, but simulated data must be validated against real operating conditions.

Self-supervised learning can exploit large volumes of unlabeled sensor data by learning normal structure before fine-tuning for a specific task. Federated or privacy-preserving approaches may allow multiple sites to contribute to model development without centralizing all raw data. These techniques do not eliminate the need for calibration, representative data, reliable labels, and deployment monitoring.

A practical checklist is:

1. Define the engineering decision and prediction horizon.
2. Identify variables, units, ranges, bandwidths, and failure modes.
3. Select sensors, locations, sampling rates, and anti-alias filters.
4. Establish a trustworthy clock and timestamp semantics.
5. Preserve raw data and complete metadata.
6. Calibrate and document sensor changes.
7. Implement quality checks and visible quality flags.
8. Label events using an explicit, auditable protocol.
9. Split data according to time, machine, or production run.
10. Engineer physics-informed features and establish baselines.
11. Validate under realistic operating conditions and costs.
12. Monitor drift, data quality, alarms, and human responses after deployment.

Challenges and Limitations

No acquisition design is perfect. Higher sampling rates increase storage and energy use; stronger filtering may hide faults; synchronization hardware adds cost; calibration consumes maintenance time; and labels can remain uncertain even after expert review. Sensors may fail in correlated ways, creating false confidence from apparent agreement. Historical data often overrepresent normal operation and underrepresent rare failures. A model trained on one machine may not transfer to another because of mechanical tolerances, mounting, control tuning, or environment.

Consequently, data engineering should be treated as an engineering design activity with explicit requirements, verification, and change control. Statistical performance is only one part of system validity.

Chapter Summary

Data acquisition and engineering convert physical behavior into reliable information for AI-enabled mechatronics. A complete pipeline includes sensing, conditioning, conversion, timestamping, communication, storage, quality assessment, labeling, feature engineering, and exploratory analysis. Sampling must reflect bandwidth and must use anti-alias filtering; synchronization must distinguish event time from arrival time. Calibration establishes measurement relationships but does not remove uncertainty. Filtering and imputation require assumptions and can introduce delay or artificial patterns.

Data quality includes completeness, validity, accuracy, consistency, timeliness, uniqueness, integrity, and representativeness. Labels must be operationally defined and protected from leakage. Features should combine statistical, spectral, contextual, and physics-informed information. EDA reveals regimes, faults, drift, and acquisition defects before modeling. Finally, industrial AI requires governance, safety boundaries, cybersecurity, traceability, and monitoring after deployment.

Key Terms

Aliasing: False low-frequency content created when a signal is sampled inadequately.

Anti-alias filter: An analogue filter that attenuates frequencies before sampling.

Calibration: Comparison with a reference to establish measurement correction and performance.

Data contract: Documented agreement defining data fields, units, timing, quality, and change rules.

Data leakage: Use of information unavailable at prediction time during model development.

Event time: The time at which a physical event occurs.

Feature: A measured or derived variable used for analysis or modeling.

Ground truth: Evidence accepted as the reference for a label or measurement.

Latency: Delay between a physical event, its measurement, computation, and response.

Label: Target classification or value assigned to an observation or interval.

Metadata: Descriptive information about measurements, sensors, units, and provenance.

Quantization: Representation of a continuous amplitude using discrete digital levels.

Quality flag: An indicator describing the validity or condition of a data record.

Sampling frequency: Number of samples acquired per second.

Sensor drift: Gradual change in sensor output or sensitivity over time.

Signal conditioning: Electrical or computational preparation of a sensor signal.

Synchronization: Alignment of measurements to a common time reference.

Timestamp: Recorded time associated with a data record.

Uncertainty: Quantified lack of complete knowledge about a measurement.

Window: A selected time interval over which features or statistics are calculated.

Review Questions

Short Answer Questions

1. Why is an anti-alias filter placed before an ADC?
2. Distinguish event time from arrival time.
3. What information should accompany a raw sensor value?
4. Why can a high completeness rate still represent poor data quality?
5. Define data leakage and give one industrial example.

Long Answer Questions

1. Explain how you would design a synchronized acquisition system for position, motor current, vibration, and temperature in a robotic cell.
2. Discuss calibration, uncertainty, and sensor placement as parts of a trustworthy measurement system.
3. Compare moving-average and exponential filters, including their benefits, delays, and limitations.
4. Describe a complete labeling protocol for predictive maintenance data and explain how label uncertainty should be handled.
5. Explain how exploratory data analysis can reveal both machine behavior and defects in the acquisition pipeline.

Analytical / Application Questions

1. A rotating shaft operates at 2,400 rpm, and the diagnostic task requires harmonics up to the sixth order. Calculate the highest frequency of interest and state the theoretical minimum sampling frequency. Explain why a practical design would sample faster.
2. An eight-bit ADC has an input range from 0 V to 10 V. Calculate its ideal quantization step and discuss two real-world effects that make the effective resolution worse.
3. A predictive-maintenance dataset contains overlapping one-second windows from the same production run. Propose a train-validation-test split and justify why a random window-level split could produce misleading performance.

Multiple Choice Questions with Answers

Question 1

Which condition is most directly associated with aliasing?

A. Sampling a signal too slowly for its frequency content

B. Using a sensor with a small physical size

C. Storing data in a relational database

D. Increasing the number of labels

Correct Answer: A. Sampling a signal too slowly for its frequency content

Question 2

A shaft rotates at 1,800 rpm. What is its rotational frequency?

A. 18 Hz

B. 30 Hz

C. 60 Hz

D. 1,800 Hz

Correct Answer: B. 30 Hz

Question 3

What is the primary purpose of calibration?

A. To guarantee that all measurements are exact

B. To increase network bandwidth

C. To establish the relationship between an instrument indication and a reference

D. To remove the need for quality flags

Correct Answer: C. To establish the relationship between an instrument indication and a reference

Question 4

Which timestamp best represents when a physical machine event occurred?

A. Event time

B. Arrival time

C. File creation time

D. Model training time

Correct Answer: A. Event time

Question 5

For a moving-average filter with M samples, which change generally produces stronger smoothing but more delay?

A. Decreasing M to zero

B. Increasing M

C. Changing the sensor unit

D. Removing timestamps

Correct Answer: B. Increasing M

Question 6

Which is an example of data leakage in predictive maintenance?

A. Using motor speed measured before a prediction

B. Using a post-failure repair code as a feature to predict that failure

C. Using a documented calibration coefficient

D. Using a quality flag generated at acquisition

Correct Answer: B. Using a post-failure repair code as a feature to predict that failure

Question 7

An ideal eight-bit ADC spans 0 V to 10 V. What is its approximate quantization step?

A. 0.0039 V

B. 0.0391 V

C. 0.391 V

D. 3.91 V

Correct Answer: B. 0.0391 V

Question 8

Which data-quality dimension asks whether all expected records are present?

A. Completeness

B. Accuracy

C. Uniqueness

D. Representativeness

Correct Answer: A. Completeness

Question 9

Why should raw data normally be retained?

A. Raw data always require no storage space

B. Raw data allow later auditing, recalibration, or alternative processing

C. Raw data automatically contain correct labels

D. Raw data eliminate synchronization requirements

Correct Answer: B. Raw data allow later auditing, recalibration, or alternative processing

Question 10

Which split is most appropriate when overlapping windows come from the same production runs?

A. Randomly split individual windows regardless of run

B. Place every window in the training set

C. Split by production run or time period

D. Split according to alphabetical sensor names

Correct Answer: C. Split by production run or time period