⚡ Key Takeaways
- This dataset contains real thermographic inspections of industrial bearings and is designed to generalize to field conditions.
- Each record is synchronized across radiometric thermal image + visible image + pixel-wise temperature CSV.
- Data is pre-labeled Healthy/Faulty, with healthy bearings around ~44°C and faulty bearings around ~74–82°C.
- The structure supports both supervised classification and unsupervised anomaly detection workflows.
- Produced by Dira Reliability S.L. and distributed via Kuinbee for research and commercial predictive maintenance use cases.
The Bearing Failure Problem: Why Industry Needs Better Data
Bearings are critical and failure-prone components across rotating machinery. When they fail unexpectedly, operations face downtime, production loss, and potential safety impact.
Most organizations still run time-based maintenance schedules, which often replace healthy components too early while still missing fast-degrading components. Predictive maintenance depends on high-quality condition signals and robust training data.
The difference between a bearing replaced on schedule and a bearing replaced on condition is the difference between a maintenance cost and a production crisis. That difference is data.
Why Thermal Data Is a Key Signal
Thermography captures early heat signatures from rising friction, lubrication breakdown, contamination, or mechanical wear—often before vibration or audible anomalies are obvious.
The challenge is model robustness under real factory variability. That requires real labeled thermographic data from actual industrial environments.
Dataset Overview: What Is Included
Dataset Specifications — Industrial Thermography for Bearing Fault Detection
| Dimension | Value |
|---|---|
| Data Type | Real industrial (non-synthetic) |
| Modalities | 3 synchronized modalities per record |
| Labels | Healthy / Faulty |
| Image Types | Radiometric IR + Visible spectrum |
| Numerical Data | Pixel-wise temperature matrices (CSV) |
| ML Readiness | Supervised and unsupervised workflows |
The dataset follows a monthly folder structure and keeps one-to-one correspondence between thermal image, visible image, and CSV matrix for each inspection record.
Healthy vs. Faulty Thermal Signature
A strong separation exists between healthy and faulty examples. Healthy samples are around 43.9–44.4°C, while faulty samples are around 73.1–82.3°C, creating an approximate 38°C differential.
Bearing Temperature Profile: Healthy vs. Faulty
💡 Original Insight
Beyond binary class labels, thermal gradients in faulty bearings can encode severity and progression patterns. This enables models to move from simple fault detection toward condition-stage estimation.
Real Industrial Data vs. Synthetic Alternatives
Real industrial thermal data includes load variability, ambient shifts, reflective effects, and operational noise that synthetic or controlled lab datasets usually miss.
Real Industrial vs. Synthetic Thermography Data
| Dimension | This Dataset | Synthetic / Lab |
|---|---|---|
| Environmental variability | Captured in real conditions | Mostly absent |
| Load variation | Operationally present | Fixed or scripted |
| Fault behavior realism | Natural progression | Artificial induction |
| Model generalizability | Higher deployment relevance | Needs adaptation |
| Multimodal synchronization | IR + Visible + CSV | Often single modality |
| Label reliability | Inspection-driven | Programmatic/simulated |
AI and Machine Learning Applications
CNN Classification
Train image models on radiometric thermography for Healthy/Faulty classification.
Anomaly Detection
Train normal-state models and flag thermal deviations from healthy distributions.
Multimodal Fusion
Fuse IR images, visible images, and temperature matrices for improved robustness.
Transfer Learning
Fine-tune pretrained vision backbones for industrial thermal fault detection.
Temporal Analysis
Use monthly organization to study progression and model fault timelines.
Deployment Pipelines
Build end-to-end predictive maintenance workflows from camera input to alerting.
Predictive maintenance adoption continues to accelerate as organizations combine better sensor coverage, lower compute cost, and high-quality labeled industrial datasets.
— Industry market analyses and enterprise predictive maintenance research, 2025–2026
Access the Bearing Thermography Dataset
Real industrial data. Multimodal. Pre-labeled. Structured for machine learning workflows.
Explore on KuinbeeHow Kuinbee Supports Access and Integration
- Preview and evaluation: Assess structure, synchronization, and quality before full dataset licensing.
- Full commercial access: High-resolution radiometric data and complete matrix coverage for production ML.
- Pipeline compatibility: Formats suitable for major ML frameworks and ingestion workflows.
- Custom collection: Commission additional thermographic data for specific machinery or conditions.
- Enterprise licensing: Licensing options for AI product teams and industrial OEM use cases.
Frequently Asked Questions
What is included in each record?
Each record pairs a radiometric thermal image, a visible image, and a pixel-wise temperature CSV for the same bearing at the same inspection moment.
Why is real industrial data important for AI performance?
It captures real operational variability and noise, improving model generalization for production deployment.
What architectures work well on this dataset?
CNN/ViT classifiers, anomaly detection models, and multimodal fusion pipelines are common strong baselines.
How can this be used commercially?
Teams use it for predictive maintenance products, reliability analytics, and condition-monitoring alert systems.
How can I access licensing options?
Kuinbee provides preview, commercial, and enterprise licensing paths, with support for custom collection.
The Bottom Line
In industrial AI, real labeled data is the constraint that matters most. This dataset provides practical, deployment-relevant thermographic signal for bearing fault detection workflows.
With strong thermal separation, synchronized multimodal records, and practical labeling, it is a strong foundation for predictive maintenance model development.
Start with Kuinbee
Access industrial thermography datasets and build predictive maintenance AI on real-world data.
Visit Kuinbee.com