Motor Fault Bench

Test rig

Fig. 1 of the paper, redrawn · hover or focus a component for its role · the part carrying the selected fault is highlighted

Fault location

Diagnosis

MLP prediction for this window

—

Class probabilities · top 6

Raw channels

First 100 ms of the window, sensor volts converted to engineering units

Key-phase and torque

V · Nm

key-phase (V)torque (Nm)

Drive-end vibration

horizontalaxialvertical

Stator currents

phase Aphase Bphase C

Spectra with order-tracked fault markers

Hover for frequency and amplitude · markers follow the measured fr and fe

Vibration spectrum, horizontal

g · 0–500 Hz

Envelope spectrum, 2–6 kHz band

g · 0–12 ×fr

Current spectrum, phase average

dB re 1 A · 0–8 ×fe

Profile timeline

Press “Run profile” to play a full operating profile

speed (rpm)torque (Nm)■ correct window■ missed window■ skipped (too slow)

Neural network

Multi-layer perceptron, lit with the activations of the current window · hover a cell for its name and value

input below training meaninput above mean / active unit / class probabilityzero (ReLU off)edges: the 150 strongest weights per layer, blue positive, copper negative

Training hyperparameters

Training loss

cross-entropy · per epoch

Validation accuracy

% · per epoch

On a microcontroller

Memory and latency budget of features + MLP · analytic estimate, calibrate on the board

Target

Window length (what-if; the model was trained on 16 384)

Model

RAM per window

Latency per window

Feature vector fed to the MLP (54 values, standardized)

Feature	Raw value	Model input	z-score vs training set	Group

About this simulator

Architecture54 → 128 → 64 → 22

Test accuracy—

Splitrun-level

Training windows23 672

The signals are synthetic. They reproduce the acquisition layout and fault catalog of the dataset described in Chen, Liu, Li, Zou, He, and Zhou, “Multi-mode fault diagnosis datasets of three-phase asynchronous motor under variable working conditions,” Data in Brief 65 (2026), with textbook fault signatures: bearing impulses at BPFO, BPFI, and 2×BSF, broken-bar sidebands at fe(1 ± 2s), unbalanced magnetic pull at 2fe, negative-sequence current for winding shorts and voltage unbalance. The real dataset (9.5 GB, Mendeley Data) shares the same 9-column CSV layout, so the same feature extractor can be pointed at it.

The classifier is a multi-layer perceptron trained in PyTorch on 54 order-tracked features. Every spectral feature is read at a multiple of the measured shaft or supply frequency, which is what lets one model work from 200 to 3000 rpm. Confusions you will see most often mirror the evaluation report: light versus high severity of the same fault, and light bearing defects at low speed drifting toward “healthy.”

Code, training scripts, the MCU budget script, and the paper-based fault catalog are in the motor-fault-bench repository on GitHub. Built by Marcelo Rovai with Claude Fable 5.1 (Anthropic), which generated the simulator, model, and this page under his direction.