TinyML Powered Real Time Rotating Machinery Fault Detection Using Vibration Spectral Analysis on ESP32

This research makes a significant contribution to the field of predictive maintenance (PdM) by successfully bridging the diagnostic gap between high-end industrial monitoring frameworks and low-cost consumer electronics. The primary contribution of this study is the development and physical deployment of a fully autonomous, real-time edge computing solution utilizing Tiny Machine Learning (TinyML) on a highly resource-constrained ESP32 microcontroller. By relying entirely on continuous high-frequency data from a single MPU6050 MEMS accelerometer, this architecture completely eliminates the requirement for expensive multi-sensor arrays and bypasses cloud connectivity. This direct-to-edge approach effectively resolves systemic IoT limitations, including unpredictable network transport latencies, bandwidth saturation, and elevated power consumption. The technical novelty of this work lies in its highly optimized inference engine, which seamlessly integrates a robust Digital Signal Processing (DSP) pipeline—featuring a 256-point Fast Fourier Transform (FFT)—with an 8-bit integer quantized (int8) Feed-Forward Neural Network. Empirical profiling validates remarkable computational efficiency; the entire spectral processing and classification cycle requires a deterministic execution latency of just 2 ms while consuming a minimal 11.4 KB of Peak RAM and 42.0 KB of Flash memory. Despite operating under extreme sub-milliwatt power and memory restrictions, the system demonstrates high diagnostic robustness, achieving an 80.27% overall accuracy on completely unseen physical test data with zero false positives for structural fault states. Ultimately, this research substantiates that advanced, non-stationary vibration classification can function precisely and reliably at the extreme edge, providing a highly scalable and cost-effective predictive maintenance paradigm for domestic rotating machinery.