Project synopsis: Bearing fault detection for small induction motors on a low-cost edge device

Department: B.Tech Electronics and Communication Engineering, final year · Author: Aditya Kulkarni · Date: 5 Aug 2026

Abstract

Small workshops and pump houses run 1 to 5 HP induction motors until they fail, because vibration analysers cost more than the motors. Bearing faults cause a large share of these failures and show up in vibration weeks before a breakdown. We propose a clip-on sensor node built on an ESP32 and a MEMS accelerometer that samples vibration, extracts frequency features on the device and classifies the bearing as healthy, inner race fault or outer race fault with a small neural network. Alerts go to the owner's phone over Wi-Fi. We will train on the public Case Western Reserve University bearing dataset, then test on a 1 HP motor in the department's machines lab with seeded faults. We expect at least 90% accuracy on the lab motor at a hardware cost under ₹2,500 per node.

Problem statement

Motor owners in small industries find out a bearing has failed when the motor stops. An unplanned stop costs a day of production plus the rewinding bill when a seized bearing burns the winding. Industrial condition monitoring systems exist but cost well over ₹50,000 per machine and need a trained analyst. There is no affordable device that tells a non-expert "this bearing needs attention" early enough to plan the repair.

Objectives

1. Build a sensor node that samples three-axis vibration at 3.2 kHz for under ₹2,500 in parts.
2. Train a classifier that separates healthy, inner race and outer race faults with at least 95% accuracy on the CWRU dataset.
3. Run the classifier on the ESP32 in under 200 ms per decision.
4. Reach at least 90% accuracy on a real 1 HP motor with seeded bearing faults.
5. Send a phone alert within one minute of a fault being detected.

Literature survey

The gap: methods that work on a PC with a lab dataset, and tools for tiny models, but no tested low-cost device that combines them and is checked on a real motor outside the dataset's rig.

Proposed methodology

flowchart LR
    A["ADXL345 accelerometer"] --> B["ESP32 sampling at 3.2 kHz"]
    B --> C["FFT and band energy features"]
    C --> D["Quantised neural network"]
    D --> E{"Fault?"}
    E -->|Yes| F["Phone alert over Wi-Fi"]
    E -->|No| B

Data and features

We will split the CWRU drive end recordings into one second windows and compute the energy in frequency bands around the shaft speed and the bearing defect frequencies, plus the RMS and kurtosis of each window. This gives about 40 features per window.

Model

A two layer dense network trained in TensorFlow, then quantised to 8-bit integers so it fits in the ESP32's memory. We will compare it with a random forest on the same features to check the network is worth its complexity.

Lab validation

The machines lab will lend us a 1 HP motor. We will seed faults by spark-eroding a 0.5 mm notch into the inner and outer race of spare 6205 bearings, then record 30 minutes of each condition at three loads.

Tools and resources

Total per node: ₹1,500, plus ₹900 of bearings for testing.

Timeline

Reviews: Review 1 in week 4, Review 2 in week 10, final review in week 15.

Team roles

Expected outcomes

• A working clip-on node and its firmware, published as open source.
• Accuracy of at least 95% on CWRU and 90% on the lab motor, with a confusion matrix for each.
• A cost and power budget for one node.

Risks

References

1. R. B. Randall and J. Antoni, "Rolling element bearing diagnostics: A tutorial," Mechanical Systems and Signal Processing, vol. 25, no. 2, pp. 485 to 520, 2011.
2. W. A. Smith and R. B. Randall, "Rolling element bearing diagnostics using the Case Western Reserve University data: A benchmark study," Mechanical Systems and Signal Processing, vol. 64 to 65, pp. 100 to 131, 2015.
3. Case Western Reserve University Bearing Data Center, bearing vibration dataset.
4. P. Warden and D. Situnayake, TinyML, O'Reilly Media, 2019.

Approval

• Guide approval (Dr. R. Subramanian, 8 Aug 2026)
• Panel approval (Review 1, week 4)