Motor Current Signature Analysis for Induction Motor Fault Detection on a Public Dataset

Motor Current Signature Analysis (MCSA) detects induction-motor faults from the stator current alone — no accelerometers and no machine access. A broken rotor bar modulates the current with sidebands at frequencies f_s(1 ± 2ks/p) around the supply component, where s is slip, k = 1, 2, … and p is pole pairs. This dissertation turns that classical physics into a rigorous, fully declared measurement study on public, CC-BY-licensed data: the LIAS Laboratory's MCSA-DC test-bench corpus (a 400 V, 50 Hz, 1.1 kW motor with healthy, one-broken-bar and two-broken-bar rotors recorded at three load levels, with currents, speed and rotor position) plus the Mendeley Data current-signature corpus for generalisation.

The study has two detection layers and compares them honestly. Layer 1 — physics-based: Welch spectra of the current signals locate the 2sf sidebands; sideband-to-fundamental amplitude ratios in dB become the fault indicator, with per-load decision thresholds established on a declared subset and slip estimated from the recorded speed channel, not guessed. Layer 2 — feature-based: time-domain and spectral features feed interpretable classifiers (k-NN, SVM, random forest) from scikit-learn under a leakage-free, grouped cross-validation protocol where all recordings from one acquisition campaign stay on the same side of the split.

Every step carries a declared protocol: FFT resolution justified against the smallest sideband spacing, metrics reported per class with confusion matrices. The result is a lab-safe, zero-hardware dissertation a viva panel can interrogate down to individual spectra — the research-grade standard of Project Work Phase-2, relevant to India's industrial motor fleets.

Broken rotor bars are among the most consequential induction-motor faults in Indian industry — pumps, compressors and textile machinery run continuously, and an undetected bar failure escalates into stalling, rotor-to-stator damage and unplanned downtime. MCSA promises non-invasive detection from stator current alone, and the literature reports high accuracies, but three methodological weaknesses make those numbers untrustworthy for practice: spectral parameters (resolution, window length) are rarely justified against the physics of the smallest sideband spacing; decision thresholds are typically quoted at rated load while real Indian plants run motors across a load range; and classifier validation commonly splits recordings randomly, leaking near-duplicate segments between train and test sets and inflating accuracy.

There is a need for an MCSA detection study with a fully declared protocol: slip estimated from measured speed rather than assumed, FFT parameters derived from sideband physics, per-load thresholds built on a declared calibration subset, leakage-free grouped cross-validation, per-class metrics with confusion matrices, and an explicit measurement of how much the leakage error inflates reported performance. Built entirely on public CC-BY datasets, such a study is reproducible by any college lab and yields guidance an industrial user can actually apply.

Abstract

Induction motors drive most of Indian industry's pumps, compressors and conveyors, and their dominant rotor fault — the broken rotor bar — develops silently until stalling or secondary damage. MCSA detects it from the stator current: broken bars introduce sidebands at f_s(1 ± 2ks/p) whose amplitude relative to the supply component indicates fault severity. This dissertation builds a complete, protocol-declared MCSA detection study on public CC-BY datasets: physics-based sideband analysis with per-load thresholds, and an interpretable feature-based classifier comparison, both validated under leakage-free cross-validation with per-class metrics.

Introduction

Condition monitoring programs stall on cost: vibration sensors and data acquisition per machine are expensive across a large fleet, and small factories rarely retrofit them. MCSA's appeal is that the current transformer already exists in every panel, so detection needs only the signal acquisition already present or cheap to add. What stands between that promise and practice is methodological: thresholds that quietly assume full load, spectrum parameters nobody justifies, and classifier results inflated by data leakage between nearly identical recordings. This study is designed so that none of those failure modes can hide.

Existing work and the gap

• The classical reference (Thomson and Fenger, IEEE Industry Applications Magazine 2001) established sideband interpretation and threshold folklore, and decades of papers report detection accuracies above 95 percent.
• Reported accuracies are frequently optimistic: recordings from the same continuous acquisition are split randomly between training and test sets, so near-duplicate segments leak across the split, and load dependence of the sideband indicator is often averaged away.
• Public machine-current corpora now exist (LIAS MCSA-DC; the Mendeley current-signature dataset), yet few published studies declare their FFT parameters, threshold construction and validation split together. The gap: a fully declared, leakage-free MCSA study that also quantifies how much the leakage mistake inflates reported accuracy.

Proposed study

• Physics layer: estimate slip from the recorded speed channel, compute Welch spectra with declared resolution and window, extract the ±2sf sideband ratios in dB, and establish per-load decision thresholds on a declared calibration subset of the LIAS corpus; report detection outcomes per condition and load.
• Feature layer: compute time-domain features (RMS, crest, kurtosis) and band-energy features; train k-NN, SVM and random-forest classifiers under grouped cross-validation, and repeat the split randomly to measure how much accuracy leakage inflates — a methodological finding in its own right.
• Generalisation check: apply the frozen physics-layer thresholds to the Mendeley corpus and report how transfer degrades, which is the honest version of a results table.

Feasibility

• Technical: both datasets are public and CC-BY-licensed; the entire pipeline is Python (NumPy, SciPy, pandas, scikit-learn) on a standard laptop; no lab visit is required.
• Economic: zero cost — no hardware, no sensors, no licences.
• Risk: misinterpreting dataset metadata (sampling rate, units, load labels) is the main hazard; the protocol therefore begins with a metadata audit and a physics sanity check (supply frequency and slip must be recoverable from the raw spectra before any further step is trusted).

In scope

• Rotor electrical faults (broken rotor bars) in three-phase induction motors, detected from stator current signals in two public CC-BY corpora (LIAS MCSA-DC primary; Mendeley current-signature dataset for generalisation).
• Physics-based detection: Welch spectra, sideband ratios at f_s(1 ± 2ks/p), per-load dB thresholds.
• Feature-based detection: interpretable time-domain and band-energy features with k-NN, SVM and random-forest classifiers under leakage-free grouped cross-validation.
• Declared spectral parameters, calibration subset, split protocol and per-class metrics (precision, recall, F1, confusion matrices).

Out of scope

• Mechanical faults detectable by vibration (bearing outer-race analysis) except as the Mendeley corpus's bearing classes appear in the generalisation check, reported as transferred performance only.
• Mixed eccentricity, inter-turn stator winding faults and online surge testing; the study stays on the rotor-bar physics.
• Deep-learning architectures; the interpretable classical models are deliberate, both for viva defensibility and because the datasets are small.
• Any new laboratory acquisition, motor disassembly or instrumentation — the study is fully offline on public data.

The study follows a quantitative, experimental research design with a gated pipeline: metadata audit → physics sanity gate → spectral analysis → physics layer → feature layer → generalisation check → analysis. Every gate must pass in writing before the next stage runs, and all results regenerate from version-controlled notebooks over immutable downloaded data.

Protocol skeleton: datasets downloaded once, hashed, and read-only thereafter; the calibration subset (threshold construction) and evaluation set are fixed before any threshold is computed; the grouped cross-validation key (acquisition-campaign id per recording) is declared before model selection; the random-split replication that measures leakage inflation is run with 10 fixed seeds.

Key derivations before analysis: the smallest sideband spacing 2sf sets the minimum record length (for example, a 2 Hz spacing requires a resolution of at most 1 Hz, hence records of at least 1 s and a Hann window of documented trade-off between leakage and amplitude accuracy); the dB margin between the faulty and healthy sideband ratios on the calibration subset sets each per-load threshold, with the margin printed on the threshold chart. Validation: the physics sanity gate is itself the first validation; classifier metrics are per-class with confusion matrices; the leakage experiment is its own control. Reproducibility: fixed seeds, hashed data, scripted figure and table regeneration.

The study is a signal-processing pipeline with two parallel detection layers over a common, frozen data foundation. The data layer ingests and hashes the public corpora and enforces the metadata audit; the signal layer turns raw current samples into declared-resolution spectra and feature vectors; the physics layer and feature layer each produce detection decisions that are finally compared against each other and against the leakage-replication control.

flowchart TD
  A["Public corpora
LIAS MCSA-DC + Mendeley dataset"] --> B["Data layer
hashes, metadata audit, read-only"]
  B --> C["Physics sanity gate
slip from speed vs spectra"]
  C --> D["Signal layer
Welch spectra, declared resolution"]
  D --> E["Physics layer
sideband dB ratios, per-load thresholds"]
  D --> F["Feature layer
time and band features, grouped CV classifiers"]
  F --> G["Leakage control
random-split replication, 10 seeds"]
  E --> H["Comparison and analysis
per-class metrics, confusion matrices"]
  G --> H
  H --> I["Generalisation check
frozen thresholds on Mendeley corpus"]
  I --> J["Dissertation results and practitioner guidance"]

Two design decisions carry the scientific weight. First, the calibration/evaluation split is fixed before any threshold exists, so the physics layer's detection table is honest out-of-sample evidence, not a fit to its own data. Second, the leakage control runs the same classifiers under a deliberately wrong random split, quantifying inaccuracy inflation — turning a common mistake in the literature into a measured methodological result. Every figure regenerates from saved arrays with fixed seeds, so a panel can ask for any spectrum, threshold or confusion matrix to be re-derived on the spot.

• Extend the study to eccentricity and bearing-fault signatures with dedicated public corpora, completing a multi-fault MCSA suite under the same declared-protocol discipline.
• Add current-signal simulation (winding-function or multi-harmonic models) to generate synthetic intermediate severity levels that public data lack, and test threshold sensitivity against them.
• Pilot the frozen thresholds on a real college-lab motor rig with a clamp meter and a sound card or low-cost acquisition board, closing the loop from public data to own measurement.
• Package the audited corpus index, spectral engine and evaluation harness as an open condition-monitoring teaching toolkit for UG and PG labs.