Four-Channel Vibration Dataset for Cross-Speed Rotating-Machinery Fault Diagnosis at 600, 900, and 1200 RPM
Description
This dataset contains 3,545 synchronized one-second vibration samples acquired from an experimental rotating system under normal, imbalance, and misalignment conditions at rotational speeds of 600, 900, and 1200 RPM. Each sample contains four radial vibration channels sampled at 25.6 kHz, comprising two orthogonal measurement directions at each of two sensing locations. The dataset includes 1,200 imbalance samples, 1,176 misalignment samples, and 1,169 normal samples. Each MATLAB file contains the four-channel vibration array together with the sampling frequency, operating speed, condition label, acquisition-level identifier, and within-acquisition second identifier. The release is organized into nine condition-speed archives. It includes a complete sample manifest, class-speed distribution, archive inventory, SHA-256 checksums, and predefined acquisition-disjoint partitions for three multi-source leave-one-speed-out scenarios and six single-source cross-speed scenarios. These partitions prevent samples originating from the same acquisition from appearing in both source training and validation sets. The dataset supports research on rotating-machinery fault diagnosis, cross-speed domain generalization, tacholess diagnosis, multisensor vibration analysis, interpretable machine learning, signal processing, and predictive maintenance. Reference implementation and evaluation code are available at https://github.com/yakoopqasim/SRO-Net.
Files
Steps to reproduce
1. Download the complete dataset folder and verify the downloaded files using SHA256SUMS.txt. 2. Select and extract the required condition-speed archives from the data folder. The nine archives correspond to three conditions—imbalance, misalignment, and normal—and three rotational speeds: 600, 900, and 1200 RPM. 3. Each extracted MATLAB file contains: - signals: four-channel vibration array of size 4 x 25600; - fs: sampling frequency of 25600 Hz; - speed_rpm: rotational speed of 600, 900, or 1200 RPM; - condition: imbalance, misalignment, or normal; - original_file: acquisition-level grouping identifier; - second_id: one-second identifier within the acquisition. 4. Use dataset_manifest.csv to associate every public sample identifier with its archive, condition, speed, acquisition group, fold assignment, and SHA-256 hashes. 5. Use recording_disjoint_partitions.csv to reproduce the predefined cross-speed experiments. It contains the training, validation, and unseen-target assignments for three multi-source and six single-source scenarios. The publication protocol uses fold 0, with acquisition identity enforced as the grouping constraint. 6. Target-speed samples must not be used for training, normalization, validation, checkpoint selection, or hyperparameter selection. 7. The SRO-Net implementation and baseline code are available at: https://github.com/yakoopqasim/SRO-Net The repository documents the required Python environment and experiment commands.
Institutions
- King Fahd University of Petroleum and MineralsEastern Province, Dhahran