CPS Data Analytics and Artificial Intelligence

Introduction and Context

A cyber-physical system (CPS) integrates physical processes, sensors, actuators, computation, communication networks, and human or automated decision-making. In such systems, data analytics and artificial intelligence (AI) transform measurements into operational knowledge. A temperature signal can reveal overheating; vibration can indicate bearing damage; camera images can identify defects; and a learned controller can adjust a process under changing conditions.

However, AI does not automatically create reliable engineering decisions. CPS data may be incomplete, noisy, asynchronous, biased, or generated under conditions not represented during model training. A model can achieve high laboratory accuracy yet fail when a sensor drifts, a product changes, a network introduces delay, or an actuator saturates. Therefore, CPS analytics must be treated as an engineering lifecycle involving sensing, data quality, modelling, validation, deployment, monitoring, safety, and governance.

This chapter develops the foundations required to apply signal analytics, anomaly detection, predictive maintenance, machine learning, computer vision, and reinforcement learning in mechatronic systems. It also examines explainability, deployment constraints, and the relationship between statistical predictions and physical risk.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Explain how CPS data move from physical sensing to operational decisions.
2. Assess data quality using accuracy, completeness, consistency, timeliness, and representativeness.
3. Apply basic time-domain and frequency-domain signal analytics.
4. Distinguish supervised, unsupervised, semi-supervised, and reinforcement learning.
5. Design anomaly-detection and predictive-maintenance workflows.
6. Explain the role of machine vision in inspection and process monitoring.
7. Evaluate AI models using metrics appropriate to engineering risk.
8. Identify deployment constraints involving latency, computation, communication, safety, cybersecurity, and explainability.

1. From Physical Signals to Engineering Decisions

A CPS data pipeline normally contains the following stages:

1. Physical phenomenon: vibration, force, temperature, pressure, current, position, or visual appearance.
2. Sensing and acquisition: a transducer converts a physical variable into an electrical signal; an analogue-to-digital converter samples it.
3. Pre-processing: calibration, filtering, synchronization, scaling, and removal of obvious errors.
4. Feature extraction: numerical descriptors are computed from raw or transformed signals.
5. Inference: a statistical or machine-learning model estimates a state, class, fault probability, or remaining useful life.
6. Decision and action: an operator or controller schedules maintenance, changes a set point, or safely stops equipment.
7. Feedback and monitoring: outcomes are recorded and used to detect model degradation and improve future decisions.

The distinction between a measurement, a feature, a prediction, and a decision is important. For example, accelerometer voltage is a measurement; root-mean-square vibration is a feature; probability of bearing failure is a prediction; and “inspect bearing within 24 hours” is a decision. These stages have different uncertainty and validation requirements.

Sampling and aliasing

If a continuous signal contains frequency components up to a maximum frequency f_max, the sampling frequency f_s should exceed twice that frequency according to the Nyquist condition:

Here f_s and f_max are measured in hertz. This condition assumes that the signal is adequately band-limited before sampling. In practice, an anti-aliasing low-pass filter is used because components above half the sampling frequency can appear incorrectly as lower-frequency components. This false representation is called aliasing and cannot generally be removed after sampling.

For a sampling interval Δt, the sampling frequency is:

For example, if vibration samples are collected every 0.5 ms, then f_s = 1/0.0005 s = 2000 Hz. The theoretical Nyquist frequency is therefore 1000 Hz, subject to appropriate analogue filtering.

2. Data Quality in Cyber-Physical Systems

Data quality is not a single property. A useful assessment considers:

Data validation should combine statistical checks with engineering knowledge. A pressure sensor may produce values within its electronic range while still being physically impossible for a particular process. Useful checks include range limits, rate-of-change limits, cross-sensor relationships, missingness patterns, timestamp ordering, and calibration status.

A missing value should not automatically be replaced by the mean. Interpolation may be reasonable for a short gap in a slowly varying temperature signal, but it can conceal a critical communication failure in a safety system. Imputation must preserve an explicit indicator that the value was reconstructed.

Data leakage and dataset splitting

Data leakage occurs when information unavailable at prediction time enters training. Examples include using a maintenance code written after a failure, normalizing a test set with statistics calculated from the entire dataset, or randomly splitting strongly correlated time-series windows from the same machine into training and test sets. Leakage produces optimistic results.

For time-dependent CPS data, chronological or machine-wise splitting is often more realistic than random row-wise splitting. A model should be tested on future time periods, machines, products, or operating regimes that represent its intended use.

3. Signal Analytics

Signal analytics converts raw measurements into interpretable descriptors. The appropriate method depends on whether the signal is stationary, periodic, transient, or nonstationary.

Time-domain features

For a sequence of N samples x₁, x₂, …, x_N, the arithmetic mean is:

The mean describes central level. The root-mean-square (RMS) value is:

RMS has the same units as x and is useful for vibration or electrical magnitude. For a zero-mean signal, RMS is closely related to signal energy, but the interpretation depends on the sensor and measurement interval.

Sample variance is commonly estimated as:

The denominator N − 1 provides an unbiased estimate of population variance under independent, identically distributed sampling. CPS signals are often autocorrelated, so this statistical interpretation must be used cautiously.

The crest factor is:

It is dimensionless and can increase when impulsive events occur, such as localized bearing damage. It should not be interpreted alone because speed, load, sensor placement, and noise influence it.

Frequency-domain analysis

The discrete Fourier transform (DFT) represents a finite sequence as frequency components:

Here x[n] is the sampled signal, X[k] is the complex coefficient, j is the imaginary unit, and N is the number of samples. The frequency spacing is Δf = f_s/N. The fast Fourier transform (FFT) is an efficient algorithm for computing the DFT; it does not change the underlying transform.

Finite windows cause spectral leakage when the observation does not contain an integer number of cycles. Window functions such as Hann or Hamming reduce leakage but alter amplitude and resolution. Analysts must therefore document the sampling frequency, window, record length, overlap, and scaling convention.

For changing-speed machinery, fixed-frequency spectra can be misleading. Order tracking expresses vibration relative to rotational speed, allowing components linked to shaft order to be distinguished from fixed-frequency disturbances.

Filtering

A low-pass filter reduces high-frequency content, a high-pass filter reduces low-frequency content, and a band-pass filter retains a selected range. Filters introduce trade-offs among noise reduction, phase distortion, transient response, and computational cost. A causal real-time filter cannot use future samples, whereas an offline zero-phase filter can process data in both directions but is unsuitable for immediate control decisions.

4. Anomaly Detection

An anomaly is an observation or pattern that differs sufficiently from expected behaviour. In CPS applications, anomalies may arise from equipment faults, sensor failures, cyberattacks, environmental changes, or previously unseen but legitimate operation.

Three common approaches are:

• Threshold methods: compare a measurement or feature with an engineering limit. They are simple and auditable but may produce false alarms when operating conditions vary.
• Statistical methods: model normal distributions, residuals, or temporal behaviour. They can quantify uncertainty but may be invalid when assumptions such as normality or independence fail.
• Machine-learning methods: learn complex boundaries or representations. They can detect subtle patterns but require representative data and careful validation.

A residual-based detector compares an observation y_t with a model prediction ŷ_t:

An alarm may be generated when |r_t| exceeds a threshold. The threshold should account for measurement uncertainty, normal variation, and the cost of false alarms and missed faults. For multivariable data, the Mahalanobis distance is useful when a covariance model is appropriate:

Here x is the feature vector, μ is the mean vector, and Σ is the covariance matrix. This method assumes that Σ is invertible and that the estimated distribution is meaningful. Highly collinear features may require regularization or dimensionality reduction.

Important performance measures include:

TP, FP, and FN denote true positives, false positives, and false negatives. Accuracy can be misleading for rare failures. In predictive maintenance, recall may be prioritized for safety-critical faults, while excessive false alarms can cause alarm fatigue and unnecessary maintenance.

5. Predictive Maintenance

Maintenance strategies include corrective maintenance after failure, preventive maintenance at scheduled intervals, condition-based maintenance using current measurements, and predictive maintenance using models to estimate future risk or remaining useful life (RUL).

A predictive-maintenance workflow is:

1. Define the failure mode and decision to be supported.
2. Establish operating context, including load, speed, environment, and duty cycle.
3. Collect synchronized sensor data and trustworthy maintenance histories.
4. Label failures and censor equipment that has not yet failed.
5. Engineer features or learn representations.
6. Train and validate using machine-wise or time-wise separation.
7. Select an alarm policy based on consequence and intervention time.
8. Deploy with monitoring, fallback rules, and periodic review.

RUL estimation is difficult because failure mechanisms, operating loads, and maintenance actions vary. A model that predicts a numerical lifetime should provide uncertainty or prediction intervals where possible. A point estimate without uncertainty can encourage false precision.

A practical decision should compare expected costs. For a simplified maintenance decision, let C_PM be the cost of preventive maintenance, C_F the cost of failure, and p the estimated probability of failure before the next service opportunity. Preventive maintenance is economically favoured when:

This is a simplified condition. It excludes downtime, false-alarm costs, production loss, safety consequences, and the effect of maintenance on future reliability. In real systems, a risk-based decision model is preferable.

6. Machine Learning for CPS

Learning paradigms

Supervised learning uses labelled examples to estimate a mapping from inputs to outputs. Classification predicts categories such as “healthy” or “misaligned”; regression predicts a continuous quantity such as temperature or RUL.

Unsupervised learning identifies structure without labels. Clustering and dimensionality reduction are useful for discovering operating regimes and visualizing high-dimensional sensor data.

Semi-supervised learning combines a small labelled set with a larger unlabelled set, which is common when fault labels are expensive.

Self-supervised learning creates learning tasks from the data itself, such as predicting masked sensor values. It can produce useful representations before task-specific fine-tuning.

Reinforcement learning (RL) learns actions through interaction with an environment. It is relevant to energy management, robot motion, scheduling, and adaptive control, but unsafe exploration must be addressed before deployment on physical machinery.

Model selection and validation

A model should be selected according to the decision, not fashion. Linear models are interpretable and computationally efficient. Decision trees and ensembles handle nonlinear relationships and mixed features. Neural networks can learn complex representations from large datasets, including images and raw waveforms, but often require more data, computation, and governance.

Overfitting occurs when a model learns training-specific variation rather than general relationships. Regularization, cross-validation, early stopping, simpler models, and independent test data can reduce it. Hyperparameter tuning must not repeatedly use the final test set.

For classification, a confusion matrix distinguishes actual and predicted classes. For regression, mean absolute error (MAE) is:

MAE has the same units as the target and is less sensitive to large errors than mean squared error. Root mean squared error is:

RMSE penalizes large errors more strongly. Metric choice must reflect engineering consequences.

7. Computer Vision in Mechatronic Systems

Computer vision applies image acquisition and analysis to inspection, localization, measurement, robot guidance, and safety monitoring. A typical system includes illumination, lens, camera, calibration target, image-processing pipeline, model, decision logic, and reject or control mechanism.

Classical methods include thresholding, edge detection, morphology, template matching, and geometric measurement. Deep convolutional models are widely used for classification, object detection, and segmentation. The method should match the task: classification assigns an image-level category; detection locates objects with bounding regions; segmentation assigns labels to pixels.

Performance depends strongly on lighting, viewpoint, focus, lens distortion, surface reflectivity, background, and product variation. A high-accuracy model trained under one lighting arrangement may fail after a lamp is replaced. Therefore, image datasets should cover expected variation and include difficult borderline examples.

Calibration is essential when pixels must be converted to physical dimensions. A camera model may correct lens distortion and establish a mapping from image coordinates to a calibrated plane. Measurements should be validated against reference artefacts, not merely against model predictions.

8. Reinforcement Learning and Intelligent Control

In RL, an agent observes a state, selects an action, receives a reward, and transitions to a new state. A Markov decision process is described by states, actions, transition probabilities, and rewards. The objective is to maximize expected discounted return:

where r is reward and γ, with 0 ≤ γ < 1, is the discount factor. The Markov assumption requires that the current state contain sufficient information for predicting future evolution given an action.

In a CPS, the state may include position, velocity, temperature, and production demand; the action may be a motor command or set-point adjustment. Reward design is a safety-critical engineering task. A reward that strongly prioritizes production rate may encourage excessive actuator wear or unsafe operation.

Safe RL commonly combines a learned policy with action limits, constraint monitors, simulation, human approval, fallback controllers, and gradual deployment. Simulation-to-reality transfer requires attention to model mismatch, sensor noise, delays, unmodelled friction, and actuator saturation. RL should not bypass established safety functions.

9. Explainability and Trustworthy Deployment

Explainability concerns whether stakeholders can understand why a model produced an output. Intrinsic interpretability is provided by models such as sparse linear models or small decision trees. Post-hoc explanations approximate or inspect a more complex model using feature importance, local perturbations, saliency maps, or counterfactual examples.

An explanation is not automatically a causal explanation. A feature may correlate with a fault because it reflects operating context rather than damage. Engineers should combine explanations with physical reasoning, controlled tests, and sensor validation.

Deployment constraints include:

• Latency: the time from measurement to decision must meet process requirements.
• Throughput: edge hardware must process the required sample or image rate.
• Memory and energy: embedded devices may have strict limits.
• Connectivity: the system should tolerate network loss when safety requires local operation.
• Determinism: control loops may require bounded execution time.
• Model updates: changes require version control, testing, rollback, and authorization.
• Cybersecurity: data and models require authentication, integrity protection, access control, and monitoring.
• Safety: AI must not silently override independent protective mechanisms.

Cloud processing can provide scalable computation, while edge processing reduces communication delay and preserves operation during disconnection. Hybrid architectures often perform immediate filtering and safety decisions at the edge while sending selected data for fleet-level analysis.

Model lifecycle monitoring

After deployment, performance can degrade through sensor drift, concept drift, changes in products, maintenance practices, or environmental conditions. Monitoring should include input distributions, missingness, prediction confidence, alarm rates, calibration, latency, and confirmed outcomes. A model should have defined operating limits and a fallback mode.

10. Illustrative Case: Predictive Maintenance of an Industrial Pump

Background

An illustrative manufacturing plant operates centrifugal pumps driven by variable-speed motors. Unexpected bearing failures cause production interruptions. Each pump has vibration, motor-current, temperature, pressure, and speed sensors.

Problem

The existing rule generates an alarm when overall vibration exceeds a fixed threshold. Operators report both missed early faults and nuisance alarms during high-speed operation. The goal is to identify developing bearing faults early enough to schedule inspection without stopping healthy pumps unnecessarily.

Analysis

The engineering team first confirms sensor calibration and synchronizes vibration with shaft speed. Data are divided by pump and by time so that windows from the same event do not appear in both training and test sets. Normal operation is stratified by speed, flow, and load. Vibration RMS, crest factor, spectral-band energy, temperature trend, and current imbalance are calculated. Speed-dependent features are expressed relative to operating regime because a fixed threshold is inappropriate across all speeds.

Maintenance records are reviewed with technicians. Labels are assigned only when evidence supports a bearing-related failure; uncertain records are retained separately rather than forced into a false class. The team compares a rule-based baseline, a logistic regression model, and a tree ensemble. Evaluation emphasizes recall at a controlled false-alarm rate and the number of days of useful warning.

Approach and solution

The deployed design uses a two-stage system. A deterministic validation layer checks sensor range, timestamp continuity, and impossible combinations. A statistical model then estimates fault risk from validated features. An alarm requires persistent elevated risk across several consecutive windows and confirmation that the pump is operating in a supported regime. The system reports the principal contributing features, current operating regime, confidence limitations, and recommended inspection actions.

The model runs at the edge to maintain operation during network interruptions. A separate safety system remains responsible for emergency shutdown. Predictions and operator responses are logged for later review.

Outcome

In this illustrative case, the combined approach reduces nuisance alarms compared with the fixed threshold and provides earlier warnings for several developing faults. The result is not attributed solely to the machine-learning model: improved synchronization, operating-regime separation, maintenance-label review, and persistence logic are equally important.

Learning points

1. Data quality and problem definition precede model selection.
2. Operating context must be represented explicitly.
3. A baseline rule is valuable for comparison and fallback.
4. Alarm persistence can improve operational usefulness, but it adds detection delay.
5. Explainability and independent safety functions support adoption.
6. Reported performance must include deployment conditions, not only a test score.

11. Ethical, Regulatory, and Professional Considerations

CPS AI can affect worker safety, employment, product quality, and access to services. Engineers must identify who is accountable when a model is wrong. Data collection should respect privacy and employment rules, particularly when cameras or wearable sensors monitor people. Data retention should be limited to legitimate operational needs, and access should be controlled.

Safety-related AI requires hazard analysis, verification, validation, traceability, and change control. A model should not be described as safe merely because it has high accuracy. Risk depends on the consequence of failure, detectability, exposure, independence of safeguards, and human response time.

Professional practice requires honest reporting of uncertainty, dataset limitations, excluded operating conditions, false alarms, missed detections, and known biases. Reproducibility benefits from versioned data, code, model parameters, calibration records, and deployment configurations.

12. Best-Practice Workflow

A robust CPS analytics project should:

1. Define the physical decision and its cost of error.
2. Identify sensors, sampling requirements, timing, and operating context.
3. Validate calibration, units, synchronization, and data lineage.
4. Establish a transparent engineering baseline.
5. Split data according to the intended deployment scenario.
6. Select features and models appropriate to available data and constraints.
7. Evaluate relevant metrics, uncertainty, robustness, and subgroup or regime performance.
8. Test abnormal, missing, delayed, and adversarial inputs.
9. Deploy with limits, fallback logic, monitoring, cybersecurity, and rollback.
10. Review outcomes continuously with domain experts and maintenance personnel.

Contemporary Developments

Current CPS research increasingly combines physics-based and data-driven models. Hybrid models can use conservation laws or machine dynamics while learning unknown parameters or residual behaviour. Digital twins support simulation, diagnosis, and what-if analysis, but their validity depends on calibration and continued synchronization with the physical asset.

Self-supervised learning is valuable where fault labels are scarce. Federated learning may allow multiple sites to improve a model without centralizing raw data, although it introduces challenges involving communication, privacy, heterogeneous equipment, and malicious updates. TinyML places compact inference models directly on low-power sensors. These developments do not remove the need for engineering validation; they increase the importance of lifecycle governance.

Chapter Summary

CPS data analytics connects physical measurements to engineering decisions through acquisition, validation, feature extraction, inference, action, and feedback. Reliable analysis requires attention to sampling, aliasing, filtering, time synchronization, missing data, leakage, and operating context. Time-domain and frequency-domain features provide useful descriptions of vibration, electrical, thermal, and process signals.

Anomaly detection may use thresholds, statistical residuals, or machine learning. Predictive maintenance depends on trustworthy failure histories, representative operating data, appropriate evaluation, and decisions that balance false alarms, missed failures, downtime, and safety. Machine learning paradigms include supervised, unsupervised, semi-supervised, self-supervised, and reinforcement learning. Computer vision supports inspection and guidance but is sensitive to imaging conditions. Reinforcement learning requires safe exploration, constraint handling, simulation validation, and fallback control.

Deployment is an engineering problem as much as a modelling problem. Latency, computation, connectivity, cybersecurity, determinism, explainability, monitoring, and change control must be designed explicitly. AI should augment—not obscure—physical reasoning, independent safety protection, and professional accountability.

Key Terms

Data leakage: Use of information during training that would not be available at prediction time.

Digital twin: A computational representation linked to a physical asset or process.

Drift: Change in sensor behaviour, data distribution, or system relationship over time.

Edge computing: Processing data near the physical source rather than exclusively in a remote cloud.

FFT: An efficient algorithm for computing the discrete Fourier transform.

Feature: A measurable attribute derived from raw data for analysis or modelling.

F1 score: The harmonic mean of precision and recall.

Machine vision: Image-based sensing and analysis for inspection, measurement, or control.

Nyquist condition: Sampling requirement that, for a band-limited signal, sampling frequency exceed twice the highest frequency.

Precision: Fraction of predicted positive cases that are actually positive.

Predictive maintenance: Maintenance planning based on estimated future condition or failure risk.

Recall: Fraction of actual positive cases detected by a model.

Reinforcement learning: Learning actions through interaction, rewards, and state transitions.

Residual: Difference between an observed value and a model prediction.

RMS: Root-mean-square magnitude of a signal.

RUL: Remaining useful life before a defined failure or performance limit.

Sampling frequency: Number of measurements acquired per unit time.

Concept drift: Change in the relationship between inputs and target outcomes.

Explainability: Ability to provide understandable reasons or evidence for model outputs.

Review Questions

Short Answer Questions

1. What is aliasing, and how can it be prevented?
2. Distinguish a measurement, feature, prediction, and decision in a CPS pipeline.
3. Why can accuracy be misleading for rare-fault detection?
4. What is the purpose of a fallback controller in an AI-enabled CPS?
5. State two factors that can cause computer-vision performance to degrade after deployment.

Long Answer Questions

1. Explain a complete predictive-maintenance workflow, including data quality, labelling, validation, deployment, and monitoring.
2. Compare threshold, statistical, and machine-learning approaches to anomaly detection.
3. Explain how sampling, windowing, filtering, and the DFT affect vibration analysis.
4. Discuss the benefits and risks of deploying AI models at the edge rather than in the cloud.
5. Explain why reinforcement learning requires special safety provisions in physical systems.

Analytical / Application Questions

1. A vibration sensor is sampled every 0.5 ms. Determine the sampling frequency and theoretical Nyquist frequency. Explain why an anti-aliasing filter is still needed.
2. A fault detector produces TP = 36, FP = 9, and FN = 4. Calculate precision, recall, and F1 score, and discuss whether the detector is suitable for a safety-critical application without further analysis.
3. Design an analytics architecture for a variable-speed robotic production cell. Specify sensors, data-quality checks, features, model type, evaluation strategy, edge/cloud allocation, fallback behaviour, and monitoring requirements.

Multiple Choice Questions with Answers

Question 1

A vibration signal contains relevant frequency components up to 800 Hz. Which sampling frequency satisfies the basic Nyquist condition?

A. 800 Hz

B. 1,000 Hz

C. 1,600 Hz

D. 400 Hz

Correct Answer: C. 1,600 Hz

Question 2

Which metric is calculated as TP/(TP + FP)?

A. Recall

B. Precision

C. Accuracy

D. F1 score

Correct Answer: B. Precision

Question 3

What is the main purpose of an anti-aliasing filter?

A. To increase sensor resolution

B. To remove frequency components that could be misrepresented after sampling

C. To label maintenance data

D. To reduce model overfitting

Correct Answer: B. To remove frequency components that could be misrepresented after sampling

Question 4

Which situation is an example of data leakage?

A. Splitting data chronologically

B. Using a calibration record available before prediction

C. Using a post-failure maintenance code as a training feature

D. Removing duplicate timestamps

Correct Answer: C. Using a post-failure maintenance code as a training feature

Question 5

Which feature is especially associated with impulsive events in a signal?

A. Crest factor

B. Sampling interval

C. Camera focal length

D. Network bandwidth

Correct Answer: A. Crest factor

Question 6

For a predictive-maintenance model, why is machine-wise splitting often preferable to random row-wise splitting?

A. It guarantees zero measurement noise

B. It better tests generalization to different machines

C. It eliminates the need for labels

D. It always increases accuracy

Correct Answer: B. It better tests generalization to different machines

Question 7

In reinforcement learning, what does the discount factor control?

A. The camera exposure time

B. The importance assigned to future rewards

C. The sensor sampling frequency

D. The number of classes in a classifier

Correct Answer: B. The importance assigned to future rewards

Question 8

Which deployment property is most directly concerned with bounded execution time?

A. Determinism

B. Precision

C. Image segmentation

D. Data uniqueness

Correct Answer: A. Determinism

Question 9

A model has precision 0.80 and recall 0.50. Which interpretation is correct?

A. Half of all predictions are correct

B. Eighty percent of predicted positive cases are true positives, while half of actual positives are detected

C. The model has zero false negatives

D. The model is necessarily safe for emergency control

Correct Answer: B. Eighty percent of predicted positive cases are true positives, while half of actual positives are detected

Question 10

Which statement best describes explainability?

A. It proves that a correlation is causal

B. It provides understandable evidence or reasons for a model output

C. It guarantees perfect prediction accuracy

D. It removes the need for independent safety systems

Correct Answer: B. It provides understandable evidence or reasons for a model output