Datasets Comparison

This processed dataset provides spectrogram images created from raw accelerometer and microphone data for a ball-bearing test rig operating under stable loads and speeds. This dataset provides a platform to assess the effectiveness of spectrogram-based fault diagnosis techniques under constant conditions. Moreover, the dataset holds potential in facilitating the application and refinement of deep learning methods. With this resource, researchers can conduct thorough testing to enhance the accuracy and reliability of machine learning methods. This dataset includes spectrograms processed from the UODS-VAFDC datasets using a Hanning window and the short-time Fourier transform (STFT) method. These spectrograms are made available to allow researchers to have access to pre-processed data, saving time in transforming the raw data. Each dataset is collected under a constant speed and load. A total of 20 bearings are tested, where each bearing has three associated datasets for three different conditions: healthy, fault developing, and faulty. Therefore, a total of 60 datasets are included. The signal length used for creating the spectrograms is 512 and each dataset contains 400 images. Therefore, a total of 24,000 images are provided (12,000 accelerometer-based images and 12,000 microphone-based images). 20 healthy datasets are provided, and the remainder are fault developing and faulty. In all cases, the data is sampled at 42,000 Hz, the sampling duration is 10s, and the nominal rotational speed of the bearing is 1,750 RPM. All data collected for healthy bearings, and bearings with inner, outer, or cage faults are collected under a nominal constant load of 400 N. Ball fault data is collected under no load. The datasets are labelled as {Letter}-{Number}-{Number}-{Number}, where: • The letter represents the bearing’s final condition: H = healthy, I = inner race fault, O = outer race fault, B = ball fault, C = cage fault. • The first number identifies the bearing that is tested. • The second number represents the bearing’s health condition: 0 = healthy, 1 = fault developing, 2 = faulty. • The third number represents the image number. • Examples of fault dataset labelling: o I-2-1-1 means the first image of an inner race fault for bearing 2 with a fault developing. o I-2-2-47 means the 47th image of an inner race fault for bearing 2 that is faulty. o O-6-1-201 means the 201st image of an outer race fault for bearing 6 with a fault developing. o O-7-2-323 means the 323rd image of an outer race fault for bearing 7 that is faulty.

The python code to convert the raw data to spectrograms is provided in the link below

University of Ottawa

Acoustic Imaging, Mechanical Vibration

Creative Commons Attribution 4.0 International

Spectrograms are created from the raw data by using the short-time Fourier transform (STFT), while using a Hanning window to convert the accelerometer and microphone values into 2D images. For each raw data file, 400 spectrogram images are created with a signal length of 512. A total of 24,000 spectrogram images are created, 12,000 images for each of the accelerometer and microphone data files. The processed data are provided in the zip files as portable network graphics files (.png).

The Python code to convert the raw data to spectrograms is provided in the first link below. The second link provides the pre-print of the article related to this dataset.

University of Ottawa

Acoustic Imaging, Mechanical Vibration

Creative Commons Attribution 4.0 International