Multimodal Induction-Motor Dataset for Operating-State Classification under a Bandwidth-Limited STM32 Edge Node

This dataset contains multimodal time-series recordings of a VFD-driven induction motor acquired with a single STM32-based embedded edge node, intended for machine-learning research on operating-state classification and condition monitoring under realistic embedded resource constraints. Motor under test. Baldor-Reliance IDNM3538 inverter-duty induction motor (0.5 HP, 4 poles, nominal 1725 rpm at 60 Hz), run unloaded and driven by a variable-frequency drive (VFD) programmed with nine preset electrical frequencies. Acquisition hardware. STMicroelectronics B-U585I-IOT02A Discovery board (STM32U585AI Cortex-M33 at 160 MHz) running FreeRTOS. Onboard sensors: ISM330DHCX three-axis accelerometer and gyroscope, IIS2MDC three-axis magnetometer, HTS221 temperature/humidity sensor, and MP23DB01HP PDM microphone. An SSD1306 OLED display and serial logging ran concurrently as periodic RTOS tasks, which constrained the sustainable sampling rate. Operating states (11). S0b (VFD off), S0a (VFD energized, motor stopped), and S1–S9 with electrical frequencies fe ∈ {5, 8, 14, 17, 23, 29, 33, 36, 45} Hz. Each state was held for approximately 300 s; total acquisition ≈ 51.8 min. Data format. 311 883 samples at an effective sampling rate Fs ≈ 100 Hz (measured timing jitter ≈ 0.19 ms), organized as a CSV file with 47 columns, one row per sample. Columns include a timestamp, the state label, three-axis accelerometer (g), three-axis gyroscope (dps), three-axis magnetometer (G or µT — specify), temperature, humidity, and [a 32-band log-magnitude audio representation / audio features]. Intended use. Supervised classification of operating state, time-series regression of electrical frequency, and unsupervised/anomaly-detection benchmarking on resource-constrained edge hardware. Because the acquisition is monotonic in speed, users should be aware of the temporal collinearity between the independent variable and time, and are advised to use group-aware cross-validation. Known limitations. Single motor, unloaded operation, no induced faults; magnetic and inertial channels sampled at 100 Hz (below the band required for bearing-defect diagnostics); audio provided only as a reduced/decimated representation, not raw waveform.