Motor Current Analytics
Motor Current Analytics is a current-based maintenance-screening and evidence-explorer prototype for three-phase induction motors.
It helps an engineer inspect what motor current can reveal about healthy operation and selected fault patterns—while making the measurement limits explicit instead of hiding uncertainty behind a confident prediction.
Inspiration
I am an electrical engineer, and I kept returning to a practical industrial question: induction motors run everywhere, often for long hours and under changing operating conditions. Small losses in efficiency, undetected degradation, and unplanned maintenance can scale into major operational cost.
Condition monitoring is valuable, but the most useful signals are not always the easiest to deploy everywhere. Three-phase current is often already accessible around a motor. I wanted to investigate whether current-derived signal representations could help maintenance teams screen for fault patterns, produce useful evidence, and know when the right answer is to escalate to inspection.
What it does
Motor Current Analytics opens a verified reference recording and lets an engineer inspect:
- current-derived STFT, wavelet, and envelope representations;
- the fixed signal-processing and ensemble-evaluation pipeline;
- stored held-out evaluation results;
- configuration comparisons, per-class performance, and confusion matrices; and
- the sensing boundary that determines when current-only screening should escalate to human inspection.
This public prototype is intentionally an evidence explorer, not a fake browser upload demo. It does not claim to run live inference on a newly uploaded plant file.
How I built it with Codex and GPT-5.6
I began with pre-existing motor-current research code and stored evaluation artifacts. During OpenAI Build Week, I used Codex with GPT-5.6 to meaningfully extend that work into a public, runnable product experience.
Codex helped me turn a research package into an inspectable application: tracing scripts and outputs, organising the evidence bundle, building the frontend flow, improving the interaction design, reconciling results with documentation, reviewing scope and reproducibility boundaries, and preparing the public project materials.
The result combines Python signal-processing and model-evaluation workflows with a React-based browser interface. The product uses stored, verified artifacts rather than inventing dashboard values or pretending that a static demo is a live deployment.
What I learned
The strongest result was not simply a high score.
The archived broad held-out baseline reached 96.71% window-level accuracy with 95.60% macro-F1. A clearly labelled retrospective scoped analysis reached 99.85% window-level accuracy and 99.75% macro-F1 across 23 held-out source recordings.
Those windows are overlapping segments, not 1,311 independent motors. More importantly, one bearing-fault operating condition was not sufficiently separable from current in this evaluation.
That became the central engineering finding:
A CNN cannot recover a fault signature that the current measurement never exposes.
Instead of treating that as an inconvenience to hide, Motor Current Analytics treats it as a decision boundary: when the signal is outside the validated sensing scope, the correct action is to escalate to inspection or an additional sensing modality.
Challenges
The hard part was making the project honest as well as polished.
Real measured signals are messy, and a strong aggregate metric does not automatically prove broad field readiness. I had to distinguish window-level performance from independent motor-level validation, preserve the broad baseline alongside the scoped analysis, and avoid presenting a browser UI as a deployed industrial inference system.
I also had to turn a research workflow into something a maintenance engineer could understand without removing the evidence that makes it credible.
Where this goes next
The next step is a pilot-stage current-based screening workflow: validate acquisition quality and operating conditions, generate an evidence packet for a maintenance engineer, flag risk, and escalate uncertain cases to inspection.
The goal is not blind automation. It is practical maintenance intelligence that makes useful evidence visible—and makes its limits visible too.
Log in or sign up for Devpost to join the conversation.