This dataset was acquired through measurements on a condition monitoring demonstrator consisting of an AC induction motor powered by 230 V 50 Hz single phase AC using a motor capacitator to start and operate the motor. The motor has a removable fan housing and its original fan is replaced by different 3D printed fans both similar to the original fan and with modifications, such as missing fan blades. The motor is connected to an air-compressor using a metal shaft with an attached 3D printed attachment that allows to fasten a grub screw in order to create an unbalance. The aluminum profiles holding the motor and compressor are uncoupled from each other and the main frame using polymer springs. An LSM9DS1 sensor is mounted on the motor profile and used to measure 3D accelerations at 400 Hz sample. The sensor is connected to an ESP32 microcontroller reading the measurements at <= 3 ms sample time. The condition monitoring demonstrator allows to configure a multitude of operating conditions, of which we select the eight conditions described below for this paper's dataset.

Each condition is labeled with an ID and an abbreviated label and a short description is given. We recommend to also view the video documentation of the machine conditions at https://t1p.de/ai4i2021video

For each condition 10 seconds of structure-borne sound data is collected using the accelerometer. Accelerometer data showed a jitter with sample times between 2 to 3 ms. The data was then harmonized to regular time intervals at a sampling rate of 300 Hz using cubic spline interpolation. Time series data is available in the folder 'Time Series Data'. There, both raw (_raw.csv) and harmonized data (_hrm.csv) are available. Acceleration values are represented in mg (=10^(-3) g)

Air-borne sound data is recorded using a microphone, which to a small extent contains background noises, although much less than could be expected in may real industrial settings. Microphone data was collected at 48000 Hz and 16 bit resolution and is stored as (*_audio.wav) files in the ‘Time Series Data Folder’.

To train a condition monitoring classifier, we recommend to use the frequency features in the folder 'Frequency Features'. There, a short-time Fourier transform using a 200 ms rectangular window is performed on both structure-borne and air-borne sound data. To acquire more observations, windows overlap by 80 %. Structure-borne sound data is transformed to 10, 15, ..., 120 Hz and air-borne sound data to 25, 50, ..., 2500 Hz frequency amplitude values.

This results in 250 observations per condition, each with 3 x 23 = 69 structure-borne, and 100 air-borne frequency features. The resulting feature dataset of 8 x 250 = 2000 observations is labeled with corresponding IDs and labels, contains the time-stamp at which the STFT window started and the 169 frequency features. The table’s heading denotes the respective acceleration direction and frequency (e.g. xAcc0085Hz, zAcc0015Hz) or the air-borne sound (e.g. snd0075Hz, snd1225Hz).

Stephan Matzka, HTW Berlin, stephan.matzka@htw-berlin.de

This dataset is part of a publication, please cite. S. Matzka, J. Pilz and A. Franke, "Structure-borne and Air-borne Sound Data for Condition Monitoring Applications," 2021 4th International Conference on Artificial Intelligence for Industries (AI4I), 2021, pp. 1-4, doi: 10.1109/AI4I51902.2021.00009

Table of G'0 (low frequency elastic plateau) measurements in Pa, relative distances between the centre of beads and the edge of the cell cluster in microns and volume of cell clusters in micron3.