2.2. Experimental Testbed and Dataset Acquisition

The dataset used in this work is obtained from the Mechanical Engineering Construction and Drive Technology (KAt) Research datacenter of Paderborn University, Germany. The current signal dataset is produced from a test bed as shown in Figure 2, which consists of a 425 W permanent magnet synchronous motor controlled by a standard frequency inverter of switching frequency 16 kHz. The detailed experimental procedure is provided in [47]. Here, the test rig was operated under various working conditions to confirm its robustness as well as to investigate the impact of the operating parameters. To make the data more reliable, both real and artificially damaged bearings were used. Among the 32 different bearings used, six were healthy, 14 were naturally damaged, and in the remaining 12, the damage was created artificially. Regarding bearing damage categorization, the specific criteria were set by the researchers in [47] that allowed for grouping the bearings into four categories, among them the first three categories described the bearing properties, whereas the last category that was called “damage” was prepared according to ISO 15243. Furthermore, the appropriate directions of geometrical sizes of the cracks in bearings were assigned according to VD1 3832 (2013). Moreover, in the original study, the measured data was verified by means of envelope analysis and machine learning-based classification.

Schematic diagram of signal measurement.

The sets of bearings used in this dataset can be split into three groups based on their health state: healthy bearings, bearings with inner race faults, and bearings with an outer race fault. Four different operating conditions were created by varying the rotational speed, load torque, and radial force, as shown in Table 1.

Operating Conditions.

The motor phase currents were measured with a current transducer (LEM CKSR 15-NP) at two different phases denoted as Current Signal 1 (CS1) and Current Signal 2 (CS2). First, 20 measurements, each over a duration of 4 seconds, were taken for every bearing set presented in Table 2. Then, the collected signals were filtered using a 25 kHz low pass filter and finally sampled at a rate of 64 kHz. For each bearing set, 20 measurements were performed to determine the force, CS1, CS2, speed, torque, and vibration signal. Although the current signals from 32 bearing sets were recorded in the original dataset, the final dataset used in this study was constructed using the signals from 17 bearing sets picked from the three different health state classes. The bearing codes and types used in this research are provided in Table 2.

Bearing sets for experiments.

The final dataset can be expressed as a 136000 × 5118 matrix that contains two current signals from different phases (CS1 and CS2) and a label placed in the last column to differentiate the health states of the bearings.

The plot of current signals collected for three bearing states (healthy, inner fault, and outer fault) is provided in Figure 3. It can be seen that there are very subtle differences among the signals in time domain representation.

Motor current signal for three different conditions of bearing mentioned in Table 2.

For demonstrating the bearing fault characteristics frequencies, the most common approach is to convert the current signal from time domain to frequency domain. The current signal is sampled at a rate of 64 kHz. After combining all the data, we performed the fast Fourier Transform (FFT) on the instances of three classes. To get better frequency resolution, the initial length of the FFT function is selected equal to the length of the data signal; however, for the representation purposes, the “mirrored” part of frequency spectrum has been removed. The obtained frequency responses are provided in Figure 4. Here, the characteristic frequencies are masked because of the external noise and presence of distributed damages. Therefore, with only characteristics frequency approaches, it is very difficult to detect fault. In such cases, denoising techniques and the machine learning models can be applied to classify different damages.

Frequency spectrum from MCS for healthy bearing, inner ring damage, and outer ring damage.