Back to Selected Work

Project 03

Machines Before Failure

A predictive-maintenance workflow using industrial telemetry, rolling-window features, and model comparison.

  • COMPLETED INTERNSHIP PROJECT
Category
Industrial AI / Predictive Maintenance
Role
Electrical & Instrumentation Intern
Maturity
Completed internship project
Year
2024-2025

30-second proof

What this proves

Problem
Industrial motor telemetry needed to be transformed into interpretable maintenance-risk signals.
Operational or business context
Bahrain Steel internship project using real industrial telemetry.
My contribution
I engineered rolling-window sensor features, compared Random Forest and XGBoost, evaluated the normalized time-to-failure target, and presented the findings.
Constraints
Multivariate sensor signals · Time-dependent patterns · Normalized prediction target · Stakeholder interpretation
Project maturity
Completed internship project
Contextual technical stack
Python · pandas · scikit-learn · XGBoost · Time-series feature engineering
100k+telemetry records
~200sensors
~10motor systems
0.0146MAE on normalized time-to-failure target
0.0680RMSE on normalized time-to-failure target

Measurement noteMAE and RMSE refer to the normalized target and must not be described as physical time units.

Ownership

Designed

  • I designed rolling-window features to represent recent sensor behavior.

Implemented

  • I implemented the model comparison workflow for Random Forest and XGBoost.

Collaborative context

  • I presented the findings in an internship stakeholder context.
Overview

Overview

A predictive-maintenance workflow using industrial telemetry, rolling-window features, and model comparison.

The problem

The problem

Raw industrial telemetry needs careful feature preparation before it can support predictive-maintenance analysis.

Context

Context

The project used industrial motor telemetry in an internship context.

My role

My role

I worked on this project as an Electrical & Instrumentation Intern.

What I built

What I built

I built a workflow using rolling-window features and model comparison for predictive maintenance.

How it works

How it works

Telemetry records are shaped into rolling-window features and compared through predictive models against a normalized time-to-failure target.

  1. Telemetry
  2. Rolling-window features
  3. Model comparison
  4. Risk estimate
  5. Maintenance insight

Decisions

  • I used rolling-window features to represent recent sensor behavior.
  • I compared two tree-based model approaches.
  • I evaluated performance on the normalized target.
Evidence and results

Evidence and results

Verified evidence includes telemetry scale, sensor scale, motor-system count, and normalized-target error values.

The MAE and RMSE values describe error on a normalized target and should not be interpreted as physical time units.

MAE and RMSE refer to the normalized target and must not be described as physical time units.

Challenges and limitations

Challenges and limitations

Public details are limited to the verified summary and aggregate evidence listed here.

What I would improve next

What I would improve next

I would add sanitized model-comparison diagrams and clearer operational interpretation.

Related work