TinyML-Based Predictive Maintenance System for Low-Cost Industrial Equipment

Abstract

Small and medium enterprises run large numbers of low-cost motors, pumps, fans and compressors that are maintained only when they break down, because commercial condition-monitoring systems cost more than the machines themselves. This paper presents the design of a TinyML-based predictive maintenance system that runs the complete analysis on a microcontroller attached to each machine. An ESP32-S3 samples a three-axis ADXL345 accelerometer at 3200 Hz, a clamp-on current sensor and a casing temperature sensor, computes 113 time- and frequency-domain features every 0.32 s, and scores them with an 8-bit quantized autoencoder trained only on data from the healthy machine. The anomaly score is smoothed into a health index that drives green, amber and red alerts on the device, while only a 100-byte summary is sent over MQTT every 10 s; a velocity check against ISO 20816 zones and an optional fault classifier run in parallel. A review of 17 related studies shows that most TinyML fault detectors are trained and tested on laboratory datasets with labelled faults, which are rarely available for a specific low-cost machine. Design analysis shows that the autoencoder has 7,929 parameters and needs about 7,700 multiply–accumulate operations per window, that the signal buffers and model fit in under 30 KB of the ESP32-S3\'s 512 KB SRAM, and that on-device inference cuts the data sent from 1.66 GB to 0.86 MB per machine per day. A validation protocol using a motor test rig with seeded faults and public bearing datasets is provided.

Introduction

This research focuses on developing a low-cost TinyML-based predictive maintenance system for industrial machines using embedded sensors and machine learning. Traditional maintenance approaches, such as run-to-failure and scheduled maintenance, may result in unexpected breakdowns or unnecessary component replacements. Predictive maintenance (PdM) addresses these problems by monitoring machine conditions and identifying abnormal behaviour before equipment failure occurs.

Conventional predictive maintenance systems often depend on cloud servers or powerful computers to process large amounts of vibration data. These systems can be expensive for small and medium-sized industries. Therefore, the proposed research uses TinyML technology to perform machine-condition monitoring directly on low-cost microcontrollers.

The proposed system uses an ESP32-S3 microcontroller with three sensors:

• Vibration sensor (ADXL345): Measures machine vibrations to identify mechanical abnormalities.

• Current sensor (SCT-013): Monitors motor current to identify changes in electrical and operating conditions.

• Temperature sensor (DS18B20): Measures machine casing temperature to detect abnormal thermal conditions.

The system collects sensor measurements and extracts 113 features from every 0.32-second data window. These features include vibration frequency-band energies, RMS values, peak values, crest factor, kurtosis, motor-current characteristics, and temperature measurements.

The main machine-learning technique is an INT8-quantized autoencoder trained using only healthy machine data. Unlike traditional supervised learning methods, this approach does not require labelled fault data during initial training. It learns the normal operating behaviour of an individual machine and identifies abnormalities by measuring reconstruction errors. An optional fault-classification model can be added when labelled fault data become available.

The literature review examines 17 related studies covering predictive maintenance, TinyML, vibration analysis, embedded machine learning, and anomaly detection. It identifies several research gaps:

1. Most existing systems require labelled fault datasets, which are difficult to obtain for individual industrial machines.

2. Model performance may decrease when deployed on real hardware compared with laboratory testing.

3. Low-frequency sampling may fail to capture important bearing-fault characteristics.

4. Many existing systems rely mainly on vibration measurements rather than combining multiple sensor inputs.

5. Machine-learning predictions are not always connected to practical maintenance decisions.

The proposed methodology addresses these limitations through machine-specific healthy-data training, higher-frequency vibration sampling at 3200 Hz, combined vibration/current/temperature monitoring, and a continuous model-retraining process.

The system generates a health index and provides three alert levels:

• Green: Machine operating normally.

• Amber: Abnormal behaviour detected; inspection should be planned.

• Red: Serious abnormality detected; urgent inspection is recommended.

The system is also designed to operate independently of an internet connection. It displays alerts locally using an LED and buzzer while sending summarized monitoring information and warning notifications to a gateway or cloud dashboard through MQTT.

The proposed framework will be evaluated using public datasets such as CWRU and MAFAULDA, along with a laboratory motor test rig. Different faults, including imbalance, misalignment, mechanical looseness, and bearing defects, will be introduced to evaluate the system.

Performance will be assessed using AUC, precision, recall, F1-score, false alarms, detection delay, inference time, memory consumption, and power usage. The proposed targets include an AUC of at least 0.95, an F1-score of at least 0.90, and real-time processing within 320 milliseconds.

The expected outcome is an affordable, energy-efficient, and intelligent predictive maintenance system capable of monitoring individual machines, detecting abnormal operating conditions, and providing early warnings without requiring continuous cloud connectivity or large labelled fault datasets.

Overall, the research aims to develop a practical TinyML-based predictive maintenance solution that combines vibration, motor-current, and temperature measurements with unsupervised anomaly detection to improve machine reliability, reduce maintenance costs, and make intelligent condition monitoring accessible to small and medium-sized industries.

Conclusion

This paper presented the design of a TinyML-based predictive maintenance system for low-cost industrial equipment. An ESP32-S3 sensor node extracts 113 vibration, current and temperature features every 0.32 s and scores them with an INT8 autoencoder trained only on the healthy machine, raising green, amber and red alerts locally and reporting a short summary over MQTT. The autoencoder needs 7,929 parameters and about 7,700 MACs per window, the whole pipeline fits in under 30 KB of RAM, and on-device inference reduces the data sent by a factor of about 1,900 compared with raw streaming. A review of 17 studies showed that this machine-specific, unsupervised approach addresses the lack of labelled fault data and the loss of accuracy after deployment seen in earlier TinyML work.

References

[1] C. Krupitzer, T. Wagenhals, M. Züfle, V. Lesch, D. Schäfer, A. Mozaffarin, J. Edinger, C. Becker, and S. Kounev, “A survey on predictive maintenance for Industry 4.0,” arXiv:2002.08224, 2020. [2] P. P. Ray, “A review on TinyML: State-of-the-art and prospects,” Journal of King Saud University – Computer and Information Sciences, vol. 34, no. 4, pp. 1595–1623, 2022. [3] R. David, J. Duke, A. Jain, V. Janapa Reddi, N. Jeffries et al., “TensorFlow Lite Micro: Embedded machine learning for TinyML systems,” in Proc. Machine Learning and Systems (MLSys), vol. 3, pp. 800–811, 2021. [4] C. Banbury, V. Janapa Reddi, P. Torelli, N. Jeffries, C. Kiraly et al., “MLPerf Tiny benchmark,” in Proc. NeurIPS Track on Datasets and Benchmarks, vol. 1, 2021 (arXiv:2106.07597). [5] 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, 2015, doi: 10.1016/j.ymssp.2015.04.021. [6] H. Purohit, R. Tanabe, T. Ichige, T. Endo, Y. Nikaido, K. Suefusa, and Y. Kawaguchi, “MIMII dataset: Sound dataset for malfunctioning industrial machine investigation and inspection,” in Proc. Detection and Classification of Acoustic Scenes and Events Workshop (DCASE), 2019, pp. 209–213. [7] T. Ince, S. Kiranyaz, L. Eren, M. Askar, and M. Gabbouj, “Real-time motor fault detection by 1-D convolutional neural networks,” IEEE Transactions on Industrial Electronics, vol. 63, no. 11, pp. 7067–7075, 2016, doi: 10.1109/TIE.2016.2582729. [8] S. Givnan, C. Chalmers, P. Fergus, S. Ortega-Martorell, and T. Whalley, “Anomaly detection using autoencoder reconstruction upon industrial motors,” Sensors, vol. 22, no. 9, Art. no. 3166, 2022, doi: 10.3390/s22093166. [9] H. Ren, D. Anicic, and T. A. Runkler, “TinyOL: TinyML with online-learning on microcontrollers,” arXiv:2103.08295, 2021. [10] K. S. Charan, “An auto-encoder based TinyML approach for real-time anomaly detection,” SAE International Journal of Advances and Current Practices in Mobility, vol. 5, no. 4, pp. 1496–1501, 2023, doi: 10.4271/2022-28-0406. [11] W. Liao, “Real time bearing fault diagnosis based on convolutional neural network and STM32 microcontroller,” arXiv:2304.09100, 2023. [12] Y. Yoo, H. Jo, and S.-W. Ban, “Lite and efficient deep learning model for bearing fault diagnosis using the CWRU dataset,” Sensors, vol. 23, no. 6, Art. no. 3157, 2023, doi: 10.3390/s23063157. [13] C.-C. Chung, Y.-P. Liang, and H.-J. Jiang, “CNN hardware accelerator for real-time bearing fault diagnosis,” Sensors, vol. 23, no. 13, Art. no. 5897, 2023, doi: 10.3390/s23135897. [14] H. Al-Libawy, “Efficient implementation of a tiny deep learning classifier based on vibration-related fault detection using limited resources hardware,” Journal of University of Babylon for Engineering Sciences, vol. 33, no. 1, pp. 88–105, 2025, doi: 10.29196/770r2493. [15] S. Arciniegas, D. Rivero, J. Piñan, E. Diaz, and F. Rivas, “IoT device for detecting abnormal vibrations in motors using TinyML,” Discover Internet of Things, vol. 5, Art. no. 41, 2025, doi: 10.1007/s43926-025-00142-4. [16] W. Ayadi, A. Amouri, S. Ali, N. Alzahmi, and M. N. Khalid, “Leveraging lightweight AI for anomaly detection in mechanical power transmission at the edge,” EPJ Web of Conferences, vol. 343, Art. no. 05014, 2025, doi: 10.1051/epjconf/202534305014. [17] C. A. C. de Lucena Filho and F. V. de Almeida, “TinyML for fault detection in rotating machinery: A post-training quantization assessment on the MAFAULDA dataset,” in Proc. Brazilian e-Science Workshop (BRESCI), 2026, pp. 33–40, doi: 10.5753/bresci.2026.249387. [18] B. Jacob, S. Kligys, B. Chen, M. Zhu, M. Tang, A. Howard, H. Adam, and D. Kalenichenko, “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” in Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 2704–2713. [19] J. W. Cooley and J. W. Tukey, “An algorithm for the machine calculation of complex Fourier series,” Mathematics of Computation, vol. 19, no. 90, pp. 297–301, 1965. [20] ISO 20816-1:2016, “Mechanical vibration — Measurement and evaluation of machine vibration — Part 1: General guidelines,” International Organization for Standardization, 2016. [21] Espressif Systems, “ESP32-S3 series datasheet,” technical documentation. [22] Analog Devices, “ADXL345 3-axis, ±2 g/±4 g/±8 g/±16 g digital accelerometer,” datasheet. [23] YHDC, “SCT-013 split core current transformer,” datasheet. [24] Analog Devices (Maxim Integrated), “DS18B20 programmable resolution 1-Wire digital thermometer,” datasheet. [25] OASIS, “MQTT version 3.1.1,” OASIS Standard, A. Banks and R. Gupta, Eds., Oct. 2014.

Copyright

Copyright © 2026 Sudhanshu Shekhar, Sachin Kumar, Ashish Kumar, Priyanshu Kumar. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.