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

Department: B.Tech Electronics and Communication Engineering, final year project report · Author: Aditya Kulkarni · Date: 24 Apr 2027

Abstract

Small workshops run induction motors until they fail, because vibration monitoring equipment costs more than the motors. We built a clip-on sensor node from an ESP32 and an ADXL345 accelerometer that classifies a motor's bearing as healthy, inner race fault or outer race fault, entirely on the device, and sends a phone alert when it finds a fault. The node computes 40 frequency band and statistical features from each one second window of vibration and runs an 8-bit quantised neural network of 11 KB. Trained on the Case Western Reserve University (CWRU) bearing dataset, the model reached 97.8% accuracy on held-out CWRU data. On a 1 HP motor in our machines lab, with faults seeded into spare bearings, accuracy dropped to 88.4%, and fine-tuning on 20% of our own recordings raised it to 93.1%, above the 90% target. Each decision takes 38 ms on the ESP32, an alert reaches the phone in 22 seconds on average, and a node costs ₹1,500 in parts. The results show that useful bearing fault detection is possible at a price a small workshop can pay, provided the model is tuned on the motor it will watch.

1. Introduction

Bearing faults are one of the most common causes of induction motor failure, and they show up in vibration well before a motor seizes. Industrial monitoring systems detect them but cost tens of thousands of rupees per machine. This project set out to detect the common bearing faults with a device costing under ₹2,500.

1.1 Objectives

2. Literature review

Envelope analysis and spectral methods are the standard way to diagnose bearings, and Randall and Antoni's tutorial [1] describes them in detail, but they assume a trained analyst reads the spectra. The CWRU dataset [3] made machine learning approaches easy to compare, and Smith and Randall's benchmark [2] showed that some CWRU recordings are much harder to diagnose than others, a warning against trusting accuracy figures from that dataset alone. Work on running small neural networks on microcontrollers [4] showed that quantised models of tens of kilobytes are practical. No work we found combined these into a low-cost device and tested it on a motor other than the CWRU rig, which is the gap this project fills.

3. System design

flowchart TB
    subgraph node1 [Sensor node on the motor]
        A["ADXL345 accelerometer, 3.2 kHz"] --> B["ESP32: windowing and FFT"]
        B --> C["40 features per window"]
        C --> D["Quantised neural network, 11 KB"]
    end
    D --> E["Wi-Fi alert service"]
    E --> F["Owner's phone"]

3.1 Sensing

The ADXL345 is held to the motor's drive end bearing housing by a magnetic mount and samples three axes at 3.2 kHz over SPI. The ESP32 collects one second windows with 50% overlap.

3.2 Features and model

For each window the ESP32 computes an FFT and the energy in bands around the shaft speed and the bearing defect frequencies, plus the RMS and kurtosis of each axis, 40 features in all. A dense network with two hidden layers of 32 and 16 units classifies the window. A fault is reported only when five windows in a row agree, which removed almost all false alarms in testing.

4. Implementation

5. Results

Faults were seeded by spark-eroding a 0.5 mm notch into the inner race of one spare 6205 bearing and the outer race of another. We recorded 30 minutes of each condition at no load, half load and full load. 20% of the recordings were used for fine-tuning, and 600 windows per condition were drawn from the rest as the test set.

Confusion matrix on the lab motor after fine-tuning (rows are the true condition):

6. Discussion

The drop from 97.8% on CWRU to 88.4% on our motor is the main finding. Our motor runs at a different speed and load from the CWRU rig, so the defect frequencies move and the band energies the model learned shift with them. Fine-tuning on 20% of our own recordings recovered most of the loss. Most remaining errors are between inner and outer race faults at no load, where the vibration is weakest. The node cannot yet tell how severe a fault is, only that one is present.

7. Conclusion

We built and tested a ₹1,500 sensor node that detects inner and outer race bearing faults on a real motor with 93.1% accuracy, deciding in 38 ms on the device and alerting the owner in about 22 seconds. All five objectives were met, with the caveat that the model must be fine-tuned on recordings from the motor it watches.

8. Future work

• Normalise features by shaft speed so one model works across motors without fine-tuning.
• Estimate fault severity from notch sizes of 0.2, 0.5 and 1.0 mm.
• Run the node from a battery and measure how long it lasts at one decision per minute.

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.

Appendix A: Individual contributions