Abstract
This article presents a comprehensive dataset acquired from two fault diagnosis environments: (1) industrial AC motors operating under various real-world conditions, and (2) belt-loosening scenarios in HVAC air handling units. The dataset was collected to support the development and validation of data-driven fault detection methods across diverse mechanical and electrical systems. For the AC motor dataset, faults were deliberately introduced to simulate common degradation modes, including coil winding faults, inter-phase short circuits, misalignment, and bearing-related issues such as rolling-element and journal bearing faults. Each fault was replicated with multiple severity levels. Data were collected under randomized speed fluctuations (6 % and 16 %) using variable frequency drive. Data were also collected under variable load conditions, and different motor capacity. The recorded sensor signals include three-phase current data (R-, S-, T-phase), vibration data (z-axis), and torque data. The HVAC dataset focuses on belt-loosening faults within air handling units and includes vibration data (x-, y-axis), current data (R-, S-, T-phase), and RPM (motor part, fan part) under varying belt tension levels. The dataset comprises over 60 GB of raw signals with current sampled at 100 kHz, vibration and torque at 25.6 kHz, and RPM at 100 kHz. Each test scenario ranges from 120 to 300 s, resulting in various of labeled data segments suitable for training and benchmarking machine learning models. Unlike existing public datasets that often assume constant speed or isolated fault types, this dataset uniquely incorporates multi-fault, multi-severity conditions under randomized speed/load variations, filling critical gaps in real-world applicability for robust fault diagnosis algorithms. The dataset enables robust evaluation of machine learning models and signal processing algorithms for fault detection, condition monitoring, and predictive maintenance in rotating machinery. The inclusion of multi-fault, multi-severity, and variable-condition data makes it especially suitable for training generalizable diagnostic algorithms in both academic and industrial contexts. Metadata and labeling for fault type, severity, and operating conditions are provided to facilitate supervised learning applications.