Data-Driven and AI-Enabled Signal Analysis

Introduction and Context

Classical signal analysis is built around explicit models. A signal is represented through its time-domain waveform, spectrum, transform, state-space model, or statistical description; system properties are then inferred from known physical laws and carefully designed experiments. This model-based perspective remains indispensable. However, modern sensing systems generate data at unprecedented scale and complexity. Signals may be nonlinear, nonstationary, multimodal, high-dimensional, weakly labelled, and affected by changing operating conditions. In such settings, data-driven and artificial-intelligence (AI) methods complement analytical signal-processing theory by learning representations, predictors, classifiers, and decision rules from observations.

Data-driven signal analysis does not mean abandoning signal-processing principles. Sampling theory, filtering, Fourier analysis, probability, estimation, and system identification provide the structure within which learning algorithms operate. AI is most effective when physical knowledge, measurement constraints, and statistical reasoning are integrated rather than treated as competing alternatives.

This chapter examines the mathematical foundations, workflows, models, evaluation methods, applications, and limitations of AI-enabled signal analysis, with emphasis on postgraduate electrical engineering practice.

Learning Objectives

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

1. distinguish model-based, data-driven, and hybrid signal-analysis approaches;
2. formulate supervised, unsupervised, self-supervised, and reinforcement-learning problems for signals;
3. construct suitable signal representations in time, frequency, time-frequency, and learned domains;
4. explain the operation and assumptions of neural architectures used for signal analysis;
5. design valid training, validation, and testing procedures for temporally dependent data;
6. interpret performance metrics, uncertainty, robustness, and generalization;
7. evaluate an AI-enabled signal-analysis system critically, including ethical and governance risks.

1. From Model-Based to Data-Driven Analysis

A model-based method begins with a presumed relationship such as

where x(t) is the measured signal, s(t; θ) is a signal model governed by parameters θ, and n(t) is measurement noise. The analyst estimates θ or tests hypotheses about the model. For a linear time-invariant system, the output is described by convolution:

where h(t) is the impulse response. Such equations are interpretable and often computationally efficient, but their accuracy depends on assumptions such as linearity, stationarity, and known system structure.

A data-driven method instead learns a mapping from examples. For classification,

where x is an input segment, ŷ is the predicted class, and φ denotes learned parameters. For regression, the output may be a continuous estimate such as fault severity or remaining useful life. Learning is generally posed as minimizing an empirical loss:

Here N is the number of training examples, yᵢ is the target, and L is a loss function. This optimization is meaningful only when the training data represent the intended operating population and the loss reflects the engineering objective.

A hybrid method combines both perspectives. Examples include a neural network constrained by Kirchhoff’s laws, a residual-learning model that estimates the error of a physical predictor, and a classifier supplied with physically meaningful spectral features. Hybridization can improve sample efficiency, interpretability, and extrapolation.

Important Terminology

• Sample: A finite observation or signal segment used by an algorithm.
• Feature: A measurable or learned attribute used for prediction.
• Label: The target category or value associated with a sample.
• Representation: The mathematical form in which a signal is supplied to a model.
• Training: Adjustment of model parameters using data.
• Inference: Production of an output from a trained model.
• Concept drift: Change in the relationship between signals and targets over time.
• Distribution shift: Difference between training and deployment data distributions.
• Data leakage: Unintended use of information unavailable at deployment or from the test set during training.

2. Signal Representations for Learning

The representation determines which structures are easy for a model to detect. No representation is universally optimal.

2.1 Time-Domain Representation

A sampled signal is commonly represented as a vector

where N is the number of samples and the sampling interval is Tₛ seconds. The sampling frequency is fₛ = 1/Tₛ hertz. Time-domain features include mean, variance, root-mean-square value, crest factor, zero-crossing rate, and autocorrelation.

For a finite sequence, the sample mean and variance are

The factor N−1 gives the usual unbiased variance estimator under independent, identically distributed observations. For strongly correlated signals, its statistical interpretation must be treated cautiously.

The root-mean-square value is

and has the same physical units as x. For a zero-mean sinusoid of peak amplitude A, x_RMS = A/√2.

2.2 Frequency-Domain Representation

The discrete Fourier transform (DFT) of x[n] is

where j = √−1. The corresponding frequency bins are fₖ = kfₛ/N for the nonnegative-frequency portion. The DFT assumes that the analyzed finite record is treated as one period of a periodic sequence. Discontinuities at the record boundaries cause spectral leakage; windowing reduces leakage but changes amplitude and resolution.

A periodogram-like power estimate is proportional to |X[k]|², with exact scaling dependent on the selected normalization and window. Therefore, amplitude and power comparisons must state the DFT scaling, window, and one-sided or two-sided convention.

2.3 Time-Frequency Representation

A short-time Fourier transform (STFT) is

where w[n] is an analysis window, m indexes the frame, R is the hop size in samples, and ω is angular frequency in radians per sample. Short windows provide better time localization; long windows provide better frequency localization. This is a resolution trade-off, not a defect of a particular algorithm.

Wavelet transforms, Wigner–Ville distributions, and learned time-frequency maps are alternatives. Quadratic time-frequency distributions may offer sharper localization but can create cross-terms. Wavelet scales are especially useful for transients and multiscale phenomena.

2.4 Learned Representations

Deep models can learn filters directly from raw waveforms or from spectrograms. A learned representation may be more task-relevant than manually selected features, but it can also be sensitive to nuisance variables, sampling changes, and spurious correlations. Engineering interpretation requires examining what properties the representation preserves and what it discards.

3. Learning Paradigms

3.1 Supervised Learning

In supervised learning, each input xᵢ is paired with a target yᵢ. Classification predicts a discrete class; regression predicts a continuous quantity. Common classification loss is cross-entropy. If pφ(c|xᵢ) is the predicted probability of class c and yᵢ is the true class, then

where yᵢc is 1 for the correct class and 0 otherwise. The logarithm requires probabilities in (0,1], and numerical implementations use stable log-softmax operations.

For regression, mean squared error is

It penalizes large errors strongly and is appropriate when squared error reflects the cost. Mean absolute error is less sensitive to outliers.

3.2 Unsupervised and Self-Supervised Learning

Unsupervised learning discovers structure without target labels. Clustering, principal component analysis (PCA), and density estimation are examples. PCA finds orthogonal directions of maximum variance. If X is a centered data matrix, principal directions are eigenvectors of its covariance matrix

provided rows of X contain observations and columns contain features. High variance does not necessarily mean high diagnostic value; PCA is descriptive rather than inherently task-optimal.

Self-supervised learning creates targets from the signal itself. A model may predict masked samples, future segments, transformed versions, or the ordering of windows. The resulting encoder can be fine-tuned using relatively few labelled examples. For signals, temporal order and augmentation design must respect physics: an augmentation that changes phase, polarity, or frequency may destroy the phenomenon of interest.

3.3 Reinforcement Learning

In reinforcement learning, an agent observes a state, takes an action, and receives a reward. It may be used for adaptive sensing, active experiment design, or control-oriented signal processing. The objective is generally a discounted return

where r is reward and 0 ≤ γ < 1 is the discount factor. Safety constraints are essential when actions affect physical equipment.

4. Principal AI Models for Signal Analysis

4.1 Convolutional Neural Networks

A one-dimensional convolutional layer applies learnable kernels to local signal neighborhoods. For channel c and output feature map q, a simplified discrete form is

where w_qc[k] is a kernel coefficient and b_q is a bias. Convolution shares weights across time, giving translation tolerance and parameter efficiency. Pooling or striding reduces resolution but may remove precise timing information.

Two-dimensional CNNs are often applied to spectrograms. This can be effective when local time-frequency patterns are informative, but a spectrogram is not an ordinary image: its axes have physical meaning and transformations must preserve that meaning.

4.2 Recurrent and Temporal Models

Recurrent neural networks process sequences using an internal state. Long short-term memory (LSTM) and gated recurrent unit architectures reduce, but do not eliminate, difficulties associated with long-range dependencies and vanishing gradients. They are useful when event order and variable-duration dynamics matter.

Temporal convolutional networks use causal or dilated convolutions. A causal model ensures that output at time t depends only on present and past samples, which is necessary for real-time inference. Bidirectional models can use future samples and are unsuitable for strict online prediction unless a delay is acceptable.

4.3 Attention and Transformers

Attention assigns data-dependent weights to interactions among sequence elements. For query Q, key K, and value V, scaled dot-product attention is

where dₖ is the key-vector dimension. Transformers can model long-range relationships, but their memory and computational requirements may be substantial for high-rate signals. Patchwise tokenization, downsampling, efficient attention, and hybrid convolution-attention models are common engineering responses.

4.4 Autoencoders and Anomaly Detection

An autoencoder maps an input to a latent vector and reconstructs it:

The reconstruction loss may be ||x − x̂||². Training primarily on normal data allows a high reconstruction error to indicate an anomaly. However, a powerful autoencoder may reconstruct abnormal signals well; reconstruction error is therefore not automatically a calibrated probability of failure. Thresholds should be selected using validation data and an explicit false-alarm/missed-detection trade-off.

5. A Rigorous AI Signal-Analysis Workflow

5.1 Define the Engineering Task

The first question is not “Which neural network should be used?” It is “What decision is required?” A system may detect an event, identify its type, estimate a parameter, forecast a future value, or rank maintenance actions. The required latency, false-alarm cost, interpretability, and operating envelope should be specified before model selection.

5.2 Data Acquisition and Quality

Sampling frequency, sensor bandwidth, anti-alias filtering, synchronization, quantization, calibration, missing data, saturation, and electromagnetic interference affect model validity. Training data should cover normal variation, fault variation, environmental conditions, sensor units, and realistic deployment modes. A large dataset with poor coverage is not necessarily informative.

5.3 Segmentation and Labelling

Signals are often divided into overlapping windows. Overlap increases the number of training examples but creates dependence between windows. Randomly splitting overlapping windows can place nearly identical samples in training and testing sets, causing leakage. Splits should usually be performed by subject, machine, experiment, or time period before windowing.

5.4 Normalization

Standardization is commonly written as

where μ_train and σ_train are calculated only from the training set. Applying test-set statistics during preprocessing leaks information. For deployment, the same stored training statistics must be used, unless an explicitly designed adaptive normalization scheme is validated.

5.5 Evaluation

For binary classification, let TP, TN, FP, and FN denote true positives, true negatives, false positives, and false negatives. Then

These definitions require nonzero denominators. Accuracy can be misleading for rare faults. Receiver operating characteristic and precision–recall curves show threshold behaviour, while calibration assesses whether predicted probabilities correspond to observed frequencies. Confidence intervals, repeated trials, and external validation are preferable to reporting one favourable split.

For regression, useful measures include MAE, RMSE, bias, and prediction-interval coverage. The metric must reflect engineering consequences. A one-degree temperature error and a one-millimetre displacement error cannot be compared without a meaningful scale or cost model.

5.6 Generalization and Robustness

A model may perform well on an independent test set but fail under a new sensor, load, temperature, speed, or firmware version. Robustness tests should deliberately vary nuisance conditions. Domain adaptation, transfer learning, invariant representations, augmentation, and recalibration can help, but each introduces assumptions that must be verified.

6. Worked Example: Window Size and Frequency Resolution

Suppose a vibration signal is sampled at fₛ = 12,000 Hz and analyzed using N = 2,048 samples. The record duration is

The DFT bin spacing is

Thus, two sinusoids separated by substantially less than 5.86 Hz will not generally appear as clearly separated DFT-bin peaks, especially after windowing. Increasing N improves nominal frequency spacing but may reduce temporal responsiveness if the system changes during the longer window. This example illustrates why window selection is a signal-analysis decision, not merely a software parameter.

7. Detailed Applied Case: Illustrative Induction-Motor Fault Monitoring

Background

Consider an illustrative industrial induction motor instrumented with three-phase current sensors and an accelerometer. The objective is to detect healthy operation, bearing degradation, rotor-bar faults, and severe misalignment early enough to schedule maintenance.

Problem

The plant has many healthy recordings but relatively few confirmed fault recordings. Motor speed and load vary, sensors differ slightly among machines, and maintenance labels are recorded at event level rather than exact fault-onset time. A random window-level split would overestimate performance because adjacent windows from one operating episode are highly correlated.

Analysis and Approach

First, the engineering team defines the deployment task as machine-level alerting over a five-minute observation interval. Missing an incipient fault is more costly than producing an occasional inspection request, but excessive false alarms reduce operator trust.

Signals are anti-alias filtered, synchronized, quality-checked, and divided into windows long enough to contain relevant mechanical cycles. Current and vibration channels are retained both as raw waveforms and as STFT magnitudes. The training, validation, and test partitions are made by motor and calendar period, not by individual windows. This tests transfer to unseen machines and later operating conditions.

A hybrid model is selected. A one-dimensional CNN learns local waveform patterns, while order-related spectral features are supplied as additional inputs. Motor speed is included as a conditioning variable because fault frequencies may shift with speed. Class-weighted cross-entropy addresses imbalance, but the alert threshold is selected using a validation cost that assigns higher weight to missed severe faults.

The model is compared with a physics-informed baseline based on manually defined sideband and vibration features, and with a simple anomaly detector trained only on healthy data. Performance is reported by motor, not only by pooled windows. Calibration plots and confusion matrices are examined, and alerts are reviewed by maintenance engineers.

Outcome

The hybrid approach improves detection of known fault patterns while reducing false alarms under load changes. However, performance decreases for a sensor type absent from training. Recalibration with a small local dataset and sensor-specific normalization partly restores performance. The final system therefore includes a data-quality gate, an uncertainty or abstention state, periodic monitoring for drift, and human confirmation before a costly intervention.

Learning Points

1. Data partitioning can matter more than architecture choice.
2. Operating variables should be modelled rather than treated as irrelevant noise.
3. A model score is not automatically an operational decision.
4. Baselines and ablation studies reveal whether AI adds real value.
5. Deployment requires monitoring, recalibration, and governance after training.

8. Interpretability, Uncertainty, and Trust

Interpretability has several meanings. A feature may be physically interpretable, a model may be structurally transparent, or a post hoc method may explain a particular output. Saliency maps, integrated gradients, occlusion tests, and feature permutation can identify influential regions, but they do not prove causal reasoning. Explanations should be checked against synthetic signals, known fault frequencies, and controlled perturbations.

Uncertainty should be distinguished from ordinary class probability. Aleatoric uncertainty arises from irreducible measurement variability; epistemic uncertainty arises from limited knowledge or data coverage. Ensembles, Bayesian approximations, conformal prediction, and calibrated probabilistic models can provide useful uncertainty estimates, but their validity depends on assumptions such as exchangeability. In safety-critical applications, abstention and human review may be preferable to forced classification.

9. Challenges and Limitations

Data Scarcity and Label Quality

Fault data are rare, labels may be delayed or ambiguous, and healthy data may contain undocumented degradation. Synthetic data and simulation can augment training, but a simulation-to-reality gap is common.

Nonstationarity

Industrial systems change due to ageing, repairs, software updates, and environmental conditions. A static model can become obsolete. Drift detection should monitor both input distributions and performance where labels become available.

Spurious Correlation

A model may learn machine identity, sensor mounting, background noise, or test-procedure artefacts instead of the target phenomenon. Cross-site validation and controlled experiments are essential.

Computational Constraints

Edge deployment requires attention to latency, memory, energy, quantization, and fail-safe behaviour. Compression methods such as pruning and knowledge distillation must be validated for their effect on rare but important events.

Reproducibility

A credible study should document data provenance, preprocessing, windowing, split strategy, random seeds where relevant, hyperparameter selection, excluded observations, and software versions. Reproducibility is particularly important because small methodological changes can produce large apparent gains.

10. Ethical, Safety, and Governance Issues

Signal-analysis systems can affect workers, patients, infrastructure, and financial decisions. Data collection should respect privacy, ownership, consent, and cybersecurity requirements. Health and biometric signals require especially careful access control and retention policies. In industrial settings, an AI recommendation must not silently override protective interlocks.

Governance should define who is accountable for model approval, threshold changes, incident response, and retirement. Records of training data, model versions, validation evidence, and post-deployment changes support auditability. Fairness is relevant whenever performance differs systematically across devices, populations, sites, or operating regimes. Regulatory classification depends on the application and jurisdiction; engineers should consult applicable safety, quality, privacy, and sector-specific standards rather than assuming that technical accuracy alone establishes compliance.

11. Contemporary Developments and Best Practice

Current research includes foundation models for multichannel time series, masked signal modelling, graph neural networks for sensor networks, neural operators for physical systems, federated learning, and physics-informed machine learning. These developments are promising but should not be adopted solely because of novelty. The central tests remain: Does the method improve the defined engineering decision? Is it robust outside the development distribution? Can its failure modes be detected?

Recommended practice is to:

• begin with a clear physical and operational baseline;
• preserve raw data and metadata, including units and sensor configuration;
• split data by the deployment-relevant independent unit;
• prevent preprocessing and feature-selection leakage;
• compare simple and complex models;
• report uncertainty, calibration, subgroup or site performance, and computational cost;
• test under realistic drift and adversarial or corrupted inputs;
• include an abstention or escalation path where errors are consequential;
• monitor performance after deployment and document model changes.

Chapter Summary

Data-driven signal analysis learns useful mappings, representations, or structures from measured data, while classical signal processing supplies essential physical and mathematical foundations. Supervised learning supports classification and regression; unsupervised and self-supervised methods address limited labels; reinforcement learning can support adaptive sensing and control. Time-domain, frequency-domain, time-frequency, and learned representations expose different signal properties. CNNs, recurrent models, attention mechanisms, and autoencoders provide complementary modelling capabilities.

A reliable system requires more than a high test score. It requires appropriate acquisition, labelling, leakage-resistant partitioning, normalization, task-relevant metrics, robustness testing, calibration, uncertainty handling, and post-deployment monitoring. Hybrid physical–data-driven models are often attractive because they combine interpretability and prior knowledge with the flexibility of learning. Ethical, safety, privacy, cybersecurity, and accountability considerations must be included from design through operation.

Key Terms

Data-driven analysis: Signal analysis in which patterns or mappings are learned from observations.

Model-based analysis: Analysis founded primarily on an explicit mathematical or physical model.

Hybrid model: A method combining physical knowledge with learned components.

Feature: A measured or computed attribute used by an analytical model.

Representation: The form in which signal information is provided to a model.

DFT: Finite-sample transform that represents a discrete sequence in frequency bins.

STFT: Windowed Fourier representation showing how spectral content changes with time.

Supervised learning: Learning from input–target pairs.

Self-supervised learning: Learning from targets generated from the data itself.

Convolutional neural network: Neural architecture using shared local filters.

Attention: Data-dependent weighting of interactions among sequence elements.

Autoencoder: Network that encodes and reconstructs an input.

Data leakage: Unintended access to information unavailable for genuine prediction.

Concept drift: Change in the signal-to-target relationship over time.

Calibration: Agreement between predicted probabilities and observed frequencies.

Uncertainty: Quantified lack of confidence arising from noise or limited knowledge.

Domain shift: Difference between development and deployment data distributions.

Abstention: Deliberate refusal to provide a potentially unreliable prediction.

Review Questions

Short Answer Questions

1. Distinguish model-based, data-driven, and hybrid signal analysis.
2. Why does window length create a time–frequency resolution trade-off?
3. Define data leakage and give one signal-processing example.
4. State the difference between precision and recall.
5. Why should normalization statistics normally be computed from training data only?

Long Answer Questions

1. Explain a complete workflow for developing an AI-enabled vibration-monitoring system, from acquisition to deployment.
2. Compare one-dimensional CNNs, recurrent networks, and attention-based models for multichannel signals.
3. Discuss supervised, unsupervised, and self-supervised learning in the context of scarce fault labels.
4. Explain how distribution shift and concept drift affect signal-analysis systems and how they can be managed.
5. Critically evaluate the role of physics-informed and hybrid models in postgraduate engineering applications.

Analytical / Application Questions

1. A signal is sampled at 8 kHz using 1,024 samples per analysis window. Calculate the window duration and nominal DFT bin spacing. Explain the implications for transient detection and frequency discrimination.
2. A fault classifier produces TP = 72, FP = 8, and FN = 18. Calculate precision, recall, and F1 score, and explain which operational cost would influence threshold selection.
3. Design a leakage-resistant experimental protocol for evaluating a model trained on overlapping windows from multiple machines operating under changing loads. Justify the unit of partitioning, metrics, and robustness tests.

Multiple Choice Questions with Answers

Question 1

Which expression represents the normalization procedure that avoids using test-set statistics during training?

A. x′ = (x − μ_train)/σ_train

B. x′ = (x − μ_test)/σ_test

C. x′ = x/(μ_train + σ_train)

D. x′ = (x + μ_train)σ_train

Correct Answer: A. x′ = (x − μ_train)/σ_train

Question 2

For a sampling frequency of 12,000 Hz and an analysis length of 2,048 samples, the nominal DFT bin spacing is approximately:

A. 0.171 Hz

B. 5.86 Hz

C. 24 Hz

D. 24,576 Hz

Correct Answer: B. 5.86 Hz

Question 3

Which situation is the clearest example of data leakage?

A. Applying an anti-alias filter before sampling

B. Training on data from several operating loads

C. Randomly splitting highly overlapping windows from one event across train and test sets

D. Using a held-out machine for final evaluation

Correct Answer: C. Randomly splitting highly overlapping windows from one event across train and test sets

Question 4

In binary classification, recall is defined as:

A. TP/(TP + FP)

B. TN/(TN + FN)

C. (TP + TN)/(TP + TN + FP + FN)

D. TP/(TP + FN)

Correct Answer: D. TP/(TP + FN)

Question 5

What is a principal limitation of using a long analysis window?

A. It necessarily eliminates spectral information

B. It can reduce temporal responsiveness to changing signals

C. It always increases aliasing

D. It prevents all forms of normalization

Correct Answer: B. It can reduce temporal responsiveness to changing signals

Question 6

Which model property makes a causal temporal network suitable for strict online inference?

A. It uses future samples to improve every prediction

B. It depends only on present and past samples

C. It requires labels during inference

D. It has no learned parameters

Correct Answer: B. It depends only on present and past samples

Question 7

An autoencoder trained on healthy signals is commonly used for anomaly detection by examining:

A. The sampling frequency only

B. The reconstruction error

C. The number of classes in the dataset

D. The DFT bin index alone

Correct Answer: B. The reconstruction error

Question 8

If TP = 72 and FP = 8, the precision is:

A. 0.80

B. 0.818

C. 0.90

D. 0.98

Correct Answer: C. 0.90

Question 9

Which practice best tests transfer to previously unseen machines?

A. Randomly splitting adjacent windows from every machine

B. Partitioning data by machine before windowing

C. Using test labels to select preprocessing parameters

D. Duplicating rare windows in the test set

Correct Answer: B. Partitioning data by machine before windowing

Question 10

What is a major reason to combine physical knowledge with a learned model?

A. To guarantee zero prediction error

B. To remove the need for validation

C. To improve use of prior structure, interpretability, or extrapolation

D. To ensure that labels are unnecessary in every task

Correct Answer: C. To improve use of prior structure, interpretability, or extrapolation