Predictive Maintenance from Machine Vibration Data

The problem: machines that fail without warning

In a textile mill, a flour mill, a cold store or a water pumping station, a failed motor bearing means stopped production, spoiled stock and an emergency repair at the worst time. Large plants use condition monitoring; small and mid-sized factories usually run to failure or replace parts on a calendar, which wastes good parts and still misses early failures.

The signals that precede a bearing failure are usually present in vibration data days or weeks in advance. This is a well-studied problem with good public datasets, which makes it a reliable choice for a project that must produce results on a schedule.

What already exists and the gap

• Siemens Senseye Predictive Maintenance offers predictive maintenance at scale for manufacturers.
• Augury combines wireless vibration sensors with AI-based machine health diagnostics as a service.
• SKF sells condition monitoring sensors and software, drawing on its bearing expertise.
• Public datasets such as the Case Western Reserve University bearing data and NASA IMS run-to-failure data are standard research benchmarks.

These products target large industrial customers with sensor contracts and enterprise pricing. The gap is a low-cost, understandable system for a workshop with ten critical motors: cheap sensors, transparent features an engineer can check, and honest numbers about false alarms. Research code exists, but much of it reports inflated accuracy from leaky data splits.

What you build

Two users: the maintenance engineer who acts on alerts, and the plant owner who sees machine health at a glance.

• Feature pipeline over raw vibration: RMS, kurtosis, crest factor, FFT bands, envelope spectrum.
• Fault classifier and a remaining useful life estimator.
• Streaming path: sensor or replayed data into InfluxDB, scored continuously.
• Grafana dashboard with health state and alerts.

The technically interesting part

The physics does most of the work. Bearing faults produce characteristic frequencies determined by bearing geometry and shaft speed. Computing those explicitly, and looking at the envelope spectrum around them, gives a small model strong, explainable performance. Throwing raw signals at a large network is a common student mistake; your comparison between the two approaches should show why.

Generalisation across conditions. A model trained at one load and speed often fails at another. Train on some conditions, test on held-out ones, and report both.

How to evaluate it

• Leak-free splits: splitting one continuous recording randomly into train and test leaks information and produces near-perfect accuracy that means nothing. Split by run, machine or operating condition, and state which.
• Classification: accuracy and per-class recall on held-out conditions.
• Early warning: hours before failure at which detection becomes reliable, and false alarms per week at that threshold.
• Remaining useful life: mean absolute error on run-to-failure data.
• Rig results (if built): detection on faults you induced yourself, which count for far more than another benchmark run.

Suggested plan

1. Weeks 1 to 2: read on bearing fault frequencies; download CWRU and IMS; explore.
2. Weeks 3 to 5: feature pipeline and leak-free splits.
3. Weeks 6 to 8: classifier, remaining useful life, deep model comparison.
4. Weeks 9 to 11: streaming path, InfluxDB and Grafana; build the rig if you can.
5. Weeks 12 to 14: early warning analysis, rig tests, report.

Path to a product

The first customer is a small factory where a family member or professor has contacts, with a handful of critical motors. Offer a pilot: install cheap sensors on three machines for a month and send a weekly health report. The model is a monthly fee per monitored machine, including sensors, with alerts by SMS or WhatsApp. Scale through maintenance contractors who already visit many factories and can resell it.

Scope warning

• One component type. Bearings are the sensible choice because the physics is documented and the datasets exist.
• Do not build a full maintenance management system with work orders and spare parts.
• Do not report accuracy from random splits of continuous recordings.
• Do not work near running industrial machinery without supervision and permission.

Learn what you need on Sefism

• Machine learning roadmap : features, classical models and validation strategy.
• Machine learning cheat sheet : metrics and cross-validation at a glance.
• Google Colab : enough compute for feature extraction and small models.
• Kaggle Datasets : mirrors of bearing and run-to-failure datasets.
• Data analytics roadmap : exploration and dashboard skills for the Grafana side.

Get help with this idea

Most of the marks here come from leak-free evaluation and a clear early-warning analysis. Tauseef can help you plan the splits, decide whether a physical rig is worth it for your team, and scope the dashboard so it supports the results rather than replacing them. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session