AI for Machines, Power Electronics and Industrial Assets

Introduction

Electrical machines, power converters, transformers, batteries, and industrial processes are increasingly instrumented with sensors and connected to supervisory control systems. This creates opportunities for artificial intelligence (AI) to estimate hidden parameters, improve efficiency, diagnose faults, predict remaining useful life (RUL), and schedule maintenance before failure occurs.

AI is most valuable when it is combined with engineering knowledge. A purely data-driven model may perform well under familiar operating conditions but fail when load, temperature, speed, topology, or materials change. Physics-informed and hybrid models reduce this risk by incorporating conservation laws, circuit equations, machine models, degradation mechanisms, and operating constraints.

This chapter develops a practical framework for applying AI to electrical assets. It covers motor drives, power electronic converters, transformers, batteries, and industrial processes, with emphasis on model selection, data quality, validation, safety, and deployment.

Learning Objectives

After studying this chapter, a student should be able to:

1. Explain the roles of data-driven, physics-based, and hybrid AI models in electrical engineering.
2. Formulate parameter-estimation problems for machines, converters, transformers, and batteries.
3. Select suitable measurements and features for efficiency optimization and fault diagnosis.
4. Distinguish fault detection, fault classification, anomaly detection, and RUL prediction.
5. Evaluate AI models using technically meaningful metrics and physically valid tests.
6. Design a predictive-maintenance workflow that includes uncertainty, human oversight, and safety constraints.

1. Asset Intelligence: From Measurements to Decisions

An industrial asset can be viewed as a dynamical system. At time k, its internal state depends on its previous state, inputs, disturbances, and degradation:

Here, xₖ is the internal state, uₖ is the commanded input, dₖ represents disturbances such as load or ambient temperature, θ is a parameter vector, and wₖ represents process uncertainty. Measurements are described by:

where yₖ is the measured output and vₖ is measurement noise. The functions f and h may be known from physics, learned from data, or constructed as hybrids.

AI tasks can be organized as follows:

A complete system therefore includes sensing, data conditioning, a model, decision logic, and an action. Prediction alone does not constitute predictive maintenance.

2. Data-Driven, Physics-Based and Hybrid Models

2.1 Data-driven models

Data-driven models infer relationships from examples. Linear regression, decision trees, random forests, support-vector machines, artificial neural networks, convolutional neural networks, recurrent models, and autoencoders are common choices.

For supervised regression, a model estimates ŷ from feature vector z:

where w denotes learned parameters. A common training objective is mean squared error:

N is the number of samples. MSE has the squared units of the target and is sensitive to large errors. Mean absolute error, MAE = (1/N) ∑|y_i - ŷ_i|, has the same units as the target and is more robust to isolated large errors.

Data-driven models are effective when representative labelled data are available. Their weaknesses include sensitivity to distribution shift, unexplained decisions, and poor extrapolation beyond the training range.

2.2 Physics-based models

Physics-based models use known relationships. For a three-phase motor, electrical equations, magnetic relationships, mechanical dynamics, and thermal equations can be combined. For a converter, switching states, inductor current, capacitor voltage, and semiconductor losses define the model. For a battery, equivalent-circuit or electrochemical equations describe voltage and state evolution.

Physics provides interpretability and constraints, but parameters may be uncertain and models may be computationally expensive. Simplifications can also omit ageing, nonlinear saturation, parasitic effects, and manufacturing variation.

2.3 Hybrid and physics-informed models

A hybrid model may use physics to predict the main behavior and AI to learn residual error:

Alternatively, a neural network may be trained with a loss that includes both measurement error and violation of a physical relationship:

λ is a weighting factor. It must be selected carefully: an excessively large value can prevent the model from fitting valid but unmodeled effects, whereas a very small value provides little physical protection.

Examples of physical constraints include non-negative resistance, conservation of energy, bounded temperature, monotonic capacity loss with ageing, and current or voltage limits. Hybrid modeling is especially useful when failure data are scarce but normal-operation data and engineering equations are available.

3. AI for Motor Drives

3.1 Machine variables and parameter estimation

Motor-drive performance depends on parameters such as stator resistance, rotor resistance, inductance, inertia, friction, and flux linkage. These parameters vary with temperature, saturation, speed, and ageing.

For a simplified direct-current motor, the armature equation is:

where Vₐ is armature voltage in volts, Rₐ is resistance in ohms, iₐ is current in amperes, Lₐ is inductance in henries, and e_b is back electromotive force in volts. The back EMF is approximately:

where kₑ is the back-EMF constant and ω is angular speed in radians per second. Electromagnetic torque is approximately:

where kₜ is the torque constant in newton-metres per ampere. These simplified relations assume an unsaturated machine and consistent SI units.

AI can estimate parameters from voltage, current, speed, and torque data. Recursive least squares is useful when parameters change gradually. A Kalman filter or extended Kalman filter can estimate unmeasured states, provided that a state-space model and approximate noise statistics are available. Neural estimators may be used when nonlinearities are substantial, but they require careful validation across speed and load ranges.

3.2 Efficiency optimization

Drive efficiency is:

where Pₒᵤₜ and Pᵢₙ are output and input power in watts. For an AC drive, input real power may be calculated as:

for a balanced three-phase system, where V_LL is line-to-line RMS voltage, I_L is line RMS current, and cosφ is power factor. Efficiency optimization must include motor copper loss, core loss, mechanical loss, inverter switching loss, conduction loss, and auxiliary power.

An AI optimizer can select switching frequency, flux level, speed trajectory, or reactive-power setpoint. It must operate within constraints such as current, voltage, thermal limits, torque demand, acoustic noise, and stability margins. Optimizing only an electrical efficiency estimate may increase temperature or reduce service life.

3.3 Fault diagnosis

Common motor faults include bearing defects, stator inter-turn short circuits, rotor-bar damage, eccentricity, insulation deterioration, and cooling failure. Useful signals include phase currents, vibration, acoustic emission, shaft speed, temperature, and common-mode voltage.

Feature examples include RMS current, crest factor, spectral amplitudes, negative-sequence current, temperature rise, and current residuals. A vibration spectrum may reveal characteristic bearing frequencies, but frequency interpretation must account for speed and sensor location.

A classifier should report confidence and allow an “unknown” or “insufficient evidence” outcome. In safety-critical drives, automatic shutdown should be based on validated protection logic rather than an unverified classifier alone.

4. AI for Power Electronic Converters

Converters operate with rapid switching and contain semiconductor, magnetic, dielectric, and thermal failure mechanisms. Important measurements include gate signals, phase currents, DC-link voltage, switching-node voltage, heat-sink temperature, junction-temperature estimates, and output quality.

4.1 Parameter and state estimation

A capacitor’s ideal relation is:

where i_C is capacitor current in amperes, C is capacitance in farads, and v_C is capacitor voltage in volts. A reduction in estimated capacitance can indicate ageing. For an inductor:

These equations are idealized; equivalent series resistance, parasitic inductance, switching ripple, sampling delay, and sensor noise must be considered in implementation.

Observers and neural networks can estimate junction temperature, capacitor equivalent series resistance, or unmeasured current. Estimation is often preferable to adding expensive sensors, but the model must be calibrated against reference measurements.

4.2 Loss and thermal optimization

Conduction loss in a simplified semiconductor model may be represented by:

where V₀ is an equivalent threshold voltage, r_d is differential on-state resistance, and currents are evaluated over the relevant interval. Switching loss depends on voltage, current, switching frequency, and device transition energy. The equation is an approximation, not a universal device law.

AI can select a switching strategy that reduces losses while maintaining output ripple and thermal limits. The optimization objective should include both immediate energy use and long-term degradation. A constrained formulation is:

subject to current, voltage, temperature, ripple, and control constraints. α, β, and γ express engineering priorities; R_risk is a risk indicator rather than a universally defined physical quantity.

5. Transformers and Industrial Assets

Transformer condition monitoring uses oil temperature, winding temperature, dissolved-gas analysis, moisture, partial discharge, bushing capacitance, load current, and vibration. Hydrogen, methane, ethane, ethylene, and acetylene concentrations can provide evidence of thermal or electrical faults, but gas interpretation requires sampling quality, historical trends, loading context, and accepted diagnostic standards.

Anomaly detection is often more realistic than direct fault classification because confirmed failure examples are rare. An autoencoder can learn a normal operating representation and calculate reconstruction error:

where z is the feature vector, f is an encoder, g is a decoder, and A is an anomaly score. A threshold should be established using validation data and operating regimes, not selected solely to maximize laboratory accuracy.

For industrial processes, AI may model quality, energy consumption, equipment health, or production throughput. Process variables are correlated and may be affected by operator actions, raw-material variation, and control-loop interactions. Causal reasoning, process constraints, and change-point detection are therefore important. A model that predicts a quality variable accurately but recommends an infeasible control action is not an effective engineering solution.

6. Batteries: State, Health and Remaining Life

Battery management systems estimate state of charge (SOC), state of health (SOH), power capability, and temperature. A coulomb-counting SOC estimate is:

where C_n is nominal capacity in ampere-hours, i is signed current, and η_c is coulombic efficiency. The sign convention must be defined; here positive i denotes discharge. The estimate accumulates sensor bias and requires periodic correction using voltage, a reference model, or another measurement.

SOH can be defined in several ways. A capacity-based definition is:

where C_available is measured usable capacity and C_rated is rated capacity under specified test conditions. Resistance-based SOH is also used. These definitions are not interchangeable.

Equivalent-circuit models commonly include an open-circuit-voltage source, ohmic resistance, and one or more RC polarization branches. AI can estimate model parameters online, infer SOH from voltage-current-temperature histories, and predict RUL. However, battery ageing depends on temperature, depth of discharge, charge rate, rest time, calendar time, and chemistry. A model trained on one chemistry or duty cycle should not be assumed valid for another.

RUL is preferably expressed as a distribution or interval rather than a single number. If failure threshold is Q_limit and predicted capacity trajectory is Q(t), the nominal crossing time satisfies:

Uncertainty in future load, temperature, measurements, and model parameters should be propagated to the predicted crossing time.

7. Fault Diagnosis and Predictive Maintenance Workflow

A robust workflow contains the following stages:

1. Define the decision: for example, derate, inspect, replace, or continue operation.
2. Identify failure modes: use failure-mode-and-effects analysis, maintenance records, and engineering knowledge.
3. Select measurements: choose sensors whose signals are sensitive to the failure mechanism and practical to maintain.
4. Synchronize and clean data: handle missing values, sensor drift, timestamp errors, outliers, and operating-mode changes.
5. Construct features or representations: retain physically meaningful quantities where possible.
6. Split data by asset and time: random row-level splitting can cause leakage when adjacent samples from the same event appear in both training and test sets.
7. Train and calibrate the model: include class imbalance and uncertainty.
8. Validate physically: test across loads, temperatures, speeds, and asset identities.
9. Integrate with protection and maintenance systems: define alarms, escalation, and human review.
10. Monitor after deployment: detect model drift and recalibrate under controlled procedures.

Useful diagnostic metrics include precision, recall, F1 score, false-alarm rate, missed-detection rate, detection delay, and calibration of confidence. Accuracy alone is misleading for rare faults. For RUL, MAE may be supplemented by prediction-interval coverage and penalty functions that distinguish early from late predictions.

Maintenance decisions should include consequence and cost. A simple expected-cost comparison is:

where probabilities and costs correspond to the chosen decision scenario. This expression is a decision aid, not a complete reliability model; dependencies, downtime, safety consequences, and spare-parts logistics may be necessary.

8. Worked Example: Motor-Drive Efficiency and Diagnosis

Consider a balanced three-phase motor drive operating at V_LL = 400 V RMS, I_L = 12 A RMS, and power factor 0.86. Suppose mechanical output power is 6.0 kW.

Input real power is:

Therefore:

The estimated efficiency is approximately 84.0%. If a thermal sensor simultaneously indicates a rising heat-sink temperature and an AI residual model detects excess input power for the same torque, possible causes include inverter switching loss, cooling degradation, current-sensor bias, or motor inefficiency. The AI result should initiate a verification sequence: check sensor calibration, compare phase-current balance, inspect cooling airflow, and review operating history. It should not immediately identify a single failed component without corroborating evidence.

9. Illustrative Case Study: Predictive Maintenance for a Pumping Station

Background

This illustrative case concerns a municipal pumping station using induction motors driven by variable-frequency drives. Unexpected bearing failures caused unplanned outages and emergency maintenance.

Problem

The station had current, speed, power, vibration, and bearing-temperature data, but no large labelled database of failures. Operating speed varied substantially with demand, making fixed vibration thresholds unreliable.

Analysis

Engineers first grouped data by speed and load. They calculated order-tracked vibration features, phase-current imbalance, temperature rise, and a physics-based estimate of expected input power. Maintenance records were aligned with sensor histories. The dataset contained missing records and several sensor replacements, which were explicitly marked.

Approach and solution

A hybrid anomaly detector learned normal residual behavior after subtracting expected power and temperature trends. A separate supervised classifier was trained only on verified bearing-fault events. The system used three levels:

• advisory: persistent anomaly requiring review;
• inspection: anomaly supported by vibration and temperature evidence;
• protection: independent hardware or control-system limit requiring immediate action.

The model was evaluated by leaving entire motors out of the training data, preventing the model from memorizing individual assets. Maintenance staff reviewed explanations showing which signals and operating periods contributed to an alarm.

Outcome

The system did not claim to predict every failure. Instead, it provided earlier inspection opportunities and reduced nuisance alarms by accounting for speed. The principal benefit was improved decision timing, not merely a higher classification score.

Learning points

1. Operating context is essential for rotating-equipment diagnosis.
2. A small number of verified failures can be combined with abundant normal data.
3. Independent protection must not be replaced by an experimental AI model.
4. Maintenance personnel and data engineers must jointly define useful alarms.

10. Limitations, Governance and Professional Practice

AI systems may fail because of sensor degradation, data leakage, concept drift, unrepresented operating conditions, adversarial or accidental data manipulation, and incorrect labels. Cybersecurity is important because manipulated measurements can cause unsafe recommendations. Access control, signed firmware, network segmentation, event logging, and secure model-update procedures are relevant controls.

Safety requires clear allocation of responsibility. AI may recommend inspection or control changes, but protection functions should remain deterministic, validated, and compliant with applicable electrical and functional-safety requirements. Engineers should document model scope, training data, assumptions, validation conditions, confidence limits, and override procedures.

Fairness is less often discussed for industrial AI than for social applications, but analogous concerns arise across assets: a model may perform well on newer machines and poorly on older machines, or well in one plant and poorly in another. Asset populations, maintenance priorities, and resource allocation should therefore be audited.

Best practice includes:

• begin with a measurable maintenance or energy decision;
• use domain knowledge to select signals and constraints;
• preserve raw data and metadata;
• validate by asset, time, and operating regime;
• report uncertainty and unknown conditions;
• compare AI against simple engineering baselines;
• test failure consequences before automation;
• monitor drift and retrain under change control;
• retain human accountability for high-consequence decisions.

11. Contemporary Developments

Digital twins combine asset models, live measurements, and historical data. Edge AI places inference near the machine, reducing latency and network dependence. Self-supervised learning can exploit large quantities of unlabelled operating data. Federated learning may allow organizations to improve models without centralizing sensitive plant data, although differences in assets and operating regimes remain challenging. Physics-guided neural operators, uncertainty-aware deep learning, and foundation models for industrial time series are active research areas.

These developments do not remove the need for instrumentation, calibration, engineering validation, or safe control design. The central trend is toward systems that combine physical structure, learned representations, probabilistic predictions, and decision-aware maintenance.

Chapter Summary

AI applications in electrical engineering are strongest when data and physics are combined. Data-driven models can estimate parameters, classify faults, detect anomalies, optimize efficiency, and predict degradation, while physical models impose interpretability, constraints, and extrapolation safeguards. Motor drives require attention to electrical, mechanical, magnetic, and thermal behavior. Converters require analysis of switching, conduction, capacitor, magnetic, and thermal stresses. Transformers and industrial processes benefit from contextual anomaly detection, trend analysis, and hybrid models. Battery applications depend on clearly defined SOC, SOH, ageing conditions, and uncertainty in RUL.

A professional predictive-maintenance system begins with a decision and failure mechanism, not with an algorithm. It requires reliable sensors, leakage-free validation, confidence estimates, human review, cybersecurity, independent protection, and continuous monitoring after deployment.

Key Terms

Asset intelligence: Use of data and models to support operation and maintenance of physical assets.

Anomaly detection: Identification of behavior that differs from an established normal pattern.

Data drift: Change in the statistical distribution of operational data over time.

Digital twin: A digital representation of a physical asset updated with measurements and models.

Fault diagnosis: Determination of the likely cause or class of abnormal behavior.

Hybrid model: Model combining physical equations with data-driven components.

Mean absolute error: Average absolute difference between predicted and measured values.

Parameter estimation: Inference of unknown model parameters from observations.

Predictive maintenance: Maintenance planned using condition or predicted future health.

Remaining useful life: Predicted time or usage remaining before a defined end-of-life condition.

Residual: Difference between a measured value and a model-predicted value.

Root mean squared error: Square root of the mean squared prediction error.

State of charge: Estimated available charge relative to a defined battery capacity.

State of health: Condition measure such as available capacity or resistance relative to a reference.

Thermal derating: Reduction of allowable operating level to limit temperature.

Uncertainty quantification: Characterization of confidence or variability in model outputs.

Physics-informed learning: Learning constrained or guided by physical laws or engineering relationships.

Concept shift: Change in the relationship between inputs and outputs.

Review Questions

Short Answer Questions

1. Distinguish a data-driven model from a physics-based model.
2. Why should motor-fault features be interpreted with speed and load information?
3. Define SOC and explain one limitation of coulomb counting.
4. What is the purpose of an anomaly score?
5. Why is accuracy alone inadequate for rare-fault diagnosis?

Long Answer Questions

1. Explain how a hybrid model can improve parameter estimation for a motor drive.
2. Discuss the measurements, failure mechanisms, and AI methods relevant to power-converter health monitoring.
3. Describe a complete predictive-maintenance workflow for a transformer.
4. Explain how uncertainty should be incorporated into battery RUL prediction.
5. Evaluate the governance and safety requirements for deploying AI in industrial control systems.

Analytical / Application Questions

1. A drive operates at 400 V line-to-line, 12 A line current, and power factor 0.86 while producing 6.0 kW mechanical output. Calculate input power and efficiency, then identify two additional checks needed before diagnosing an efficiency fault.
2. Design a sensor and modeling strategy for detecting cooling degradation in a variable-speed motor system. Explain how you would avoid false alarms caused by changing load.
3. A battery-health model performs well on laboratory cycling data but poorly in field vehicles. Analyze at least four possible causes and propose a validation plan.

Multiple Choice Questions with Answers

Question 1

Which model structure explicitly combines a physical prediction with a learned residual?

A. y = y_physics + r_AI

B. y = random noise only

C. y = constant regardless of input

D. y = label frequency only

Correct Answer: A. y = y_physics + r_AI

Question 2

For a balanced three-phase system, which expression gives real input power using line-to-line RMS voltage?

A. P = V_LL I_L

B. P = √3 V_LL I_L cosφ

C. P = 3 V_LL I_L sinφ

D. P = V_LL/I_L

Correct Answer: B. P = √3 V_LL I_L cosφ

Question 3

Which measurement is especially useful for detecting bearing-related motor faults?

A. Vibration

B. Paint color

C. Cabinet label length

D. Room lighting level

Correct Answer: A. Vibration

Question 4

What is a major limitation of coulomb-counting SOC estimation?

A. It cannot use current measurements

B. Sensor bias accumulates over time

C. It always measures temperature directly

D. It is independent of battery capacity

Correct Answer: B. Sensor bias accumulates over time

Question 5

Why should industrial time-series data be split by asset and time during validation?

A. To increase timestamp errors

B. To cause data leakage

C. To test generalization and avoid memorization

D. To remove all normal data

Correct Answer: C. To test generalization and avoid memorization

Question 6

In predictive maintenance, an anomaly score usually indicates:

A. The exact repair cost

B. Deviation from learned normal behavior

C. Guaranteed component failure

D. Nominal rated voltage

Correct Answer: B. Deviation from learned normal behavior

Question 7

If input power is 7.14 kW and output power is 6.0 kW, the efficiency is approximately:

A. 16.0%

B. 71.4%

C. 84.0%

D. 119.0%

Correct Answer: C. 84.0%

Question 8

Which is an appropriate role for independent deterministic protection?

A. Replacing all sensors with guesses

B. Providing validated response to hazardous limits

C. Training only on unlabelled data

D. Maximizing model complexity

Correct Answer: B. Providing validated response to hazardous limits

Question 9

A capacity-based battery SOH definition commonly compares available capacity with:

A. Rated capacity

B. Ambient humidity only

C. Switching frequency

D. Transformer oil pressure

Correct Answer: A. Rated capacity

Question 10

Which metric is particularly important when a missed fault is much more serious than a false alarm?

A. Missed-detection rate

B. File size

C. Number of neural-network layers

D. Display resolution

Correct Answer: A. Missed-detection rate