**3. Experimentation**

The electrical current signal from the startup transient of a SCIM is used for identifying and classifying a healthy motor (HLT) or a motor with a faulty condition from those treated in this study, i.e., one broken rotor bar (1BRB), two broken rotor bars (2BRB), outer-race faulty bearing (BRN), and mechanical unbalance (UNB). Figure 2 shows the testbench configuration, which employs distinct 1-HP SCIM (model WEG 00136APE48T) for assessing the feasibility of using the introduced procedure to identify and classify distinct operational states.

**Figure 2.** Testbench used for assessing the proposed method for detecting and classifying distinct faults in SCIM.

The motors under test received an electric power supply of 220 V ac at 60 Hz. They had 28 bars in the rotor, two poles, and the used mechanical load was an ordinary alternator that was equivalent to one-quarter of the SCIM nominal load. One phase of the three-phase electric current supply signal was collected through an i200s, ac current clamp from Fluke. The data acquisition system (DAS) used a 16-bit analog-to-digital converter (ADS7809). The instrumentation system used a sampling frequency *f* <sup>0</sup> of 1.5 kHz, obtaining 4096 samples in 2.7 s during the induction motor startup transient.

For this study, one motor was kept in healthy condition, i.e., in good physical condition, to be used as benchmark. The 1BRB and 2BRB conditions were generated in a synthetic way by drilling one hole and two holes, respectively, with a diameter of 7.938 mm, without damaging the rotor shaft, as shown in Figure 3a,b, respectively. On the other hand, the bearing was synthetically harmed on its outer race by boring a 1.191-mm hole utilizing a tungsten drill bit. Figure 3c displays the bearing model 6203-2ZNR that was synthetically damaged to carry out the experimentation. The mechanical unbalance fault was generated by adding a mass in one of the pulley arms. In a drilled hole with 8 mm of diameter, a two-sided screw was placed and secured using female screws on both sides of the pulley arm, as depicted in Figure 3d. A total of 100 trials were executed for each motor condition.

**Figure 3.** Artificially generated faults. (**a**) One broken rotor bar (1BRB); (**b**) two broken rotor bars (2BRB); (**c**) bearing with outer race damaged (BRN), and (**d**) mechanical unbalance.

Figure 4 depicts the proposed methodology for multiple fault diagnosis and classification. The electrical current signal of the startup transient was obtained by the current clamp; then, it was adjusted and analog-to-digital transformed in the DAS. The resulting discrete current signal was treated for obtaining the desired features, homogeneity and kurtosis, which were used as entries to the artificial neural network, a multilayer perceptron with a feed-forward architecture.

**Figure 4.** Proposed approach for distinct fault detection and classification.

Homogeneity and kurtosis values were obtained and normalized for each motor condition by the definitions (5) and (6), respectively. For each motor condition, a statistical analysis was performed. The mean (μ) and standard deviation (σ) of the homogeneity and kurtosis values for a motor without harm (HLT), a motor with one separated rotor bar (1BRB), a motor with two shattered rotor bars (2BRB), a motor with outer-race bearing damage (BRN), and a motor with unbalance (UNB) show that the respective probability density functions (PDF) partially cover each other in some degree, impeding a direct classification. Figure 5 shows the PDF of homogeneity and kurtosis features where the overlap among some of the treated conditions is evident. Therefore, a neural network classification was used for refining the proper operation that allowed a precise identification of multiple operational conditions.

**Figure 5.** Homogeneity and kurtosis PDF values for the different motor states: HLT, healthy; 1BRB, one broken rotor bar; 2BRB, two broken rotor bars; BRN, outer-race bearing damage; and UNB, unbalance.

The classification was carried out utilizing the homogeneity and kurtosis data as entries to the ANN. The artificial neural network was a multilayer perceptron with a feedforward architecture with two inputs (kurtosis values and homogeneity values) only, one hidden layer, and one output. A Levenberg–Marquardt backpropagation algorithm was used for training the ANN and the mean-square-error index was used for performance assessment. Inputs to the ANN were the feature vectors composed of the homogeneity and kurtosis data, where each motor condition had a total of 20 values in its validation data set, given as a result of feature vectors of length 100.
