An MEMS Gyroscope IoT Platform for Three-Phase Induction Motors: Within-Recording Phase-Loss Diagnosis, Drive-Setpoint Estimation, and Confound Auditing

This paper presents an Internet-of-Things (IoT) prototype that uses a single six-axis MEMS inertial sensor mounted on the motor housing to monitor three-phase induction motors, estimate their operating frequency, and diagnose phase-loss faults. The sensor data is acquired by an STM32F103-based node and transmitted simultaneously via Bluetooth and Wi-Fi to an Android application, where the motor condition, estimated operating point, and vibration level are displayed. This approach eliminates the additional cost and wiring typically required for current sensors, tachometers, and industrial accelerometers. Both datasets were collected at 150 Hz using all six inertial channels, with five independent trials conducted for each condition. The sensor was removed and remounted between trials. The drive-setpoint dataset comprised 270,000 samples collected at twelve setpoints ranging from 5 to 60 Hz, while the phase-loss dataset contained 45,000 samples labelled as healthy or as one of three single-phase-loss faults. Die temperature was recorded as a seventh channel but was not used in any model. Windowed time- and frequency-domain features were compared with raw-sample inputs using six classical machine learning model families and four deep learning architectures. The evaluation included a buffered chronological split, blocked five-fold cross-validation, and leave-one-trial-out cross-validation. In the final protocol, neither the held-out recording nor data from the corresponding sensor mounting were included in training. The repeated trials made this evaluation possible, and all headline results are reported under this protocol. When evaluated on held-out trials, four-class phase-loss diagnosis achieved 96.2% accuracy across 450 windows, with a cluster-bootstrap interval of 89.0–99.8% calculated over the twenty held-out recordings. The commanded drive setpoint was estimated with an RMSE of 1.50 Hz for setpoints spaced 5 Hz apart. Because a four-class score averages two tasks of very different difficulty, the two are also scored separately: detecting a fault transfers exactly, at 450 out of 450 windows for four of the five models, while naming which of the three phases was lost falls from 99.1–100% within recordings to 86.2–96.9% across them. The within-recording protocols report 100% and 0.36 Hz on the same data, and that gap is the central methodological result: a leakage-safe split inside a recording is not a substitute for a held-out acquisition. Diagnostic tests clarify the limits of the system. Accelerometer amplitude consistently tracked the drive setpoint across five sensor remountings and remained informative after accounting for temperature, indicating that the model estimates the setpoint from signal level rather than directly measuring frequency. The accelerometer carried more relevant information than the gyroscope. Because setpoint estimates became unreliable during faults, the application now suppresses them whenever a fault is detected. The system is therefore presented as a feasibility prototype