Tiny Machine Learning at the Edge: A Framework for Predictive Maintenance in Smart Manufacturing Environments

Abstract

Unplanned downtime remains one of the most persistent sources of financial loss in discrete and process manufacturing, and the industry has responded over the past decade by migrating from reactive and calendar-based maintenance toward predictive maintenance grounded in continuous condition monitoring. Conventional predictive maintenance architectures, however, depend on the transmission of raw sensor streams to cloud or on-premise servers, a dependency that introduces latency, bandwidth cost, connectivity fragility, and privacy exposure that many factory floors cannot tolerate. Tiny Machine Learning, the discipline of compressing and executing neural inference directly on microcontroller class hardware operating within milliwatt power envelopes, offers a structurally different alternative in which anomaly detection and fault classification occur at the sensor itself. This paper undertakes a theoretically grounded, qualitative synthesis of the scholarly and technical literature on TinyML enabled predictive maintenance, examining the hardware, model compression, and deployment layers that constitute the TinyML stack and situating them within the broader trajectory of Industry 4.0. Drawing on a structured review of primary studies published between 2019 and 2026, the paper develops a conceptual framework that links sensing modality, feature representation, model architecture, and compression strategy to measurable outcomes in accuracy, latency, energy consumption, and maintainability. The findings indicate that quantised convolutional and recurrent architectures deployed through frameworks such as TensorFlow Lite for Microcontrollers and CMSIS-NN can approach the diagnostic accuracy of server-hosted models while consuming a fraction of the energy and eliminating continuous connectivity requirements, although challenges of dataset scarcity, on-device retraining, and standardised benchmarking persist. The paper contributes an integrative typology of TinyML predictive maintenance research and offers methodological and managerial implications for manufacturers considering edge intelligence adoption.