An Adaptive EMD Technique for Induction Motor Fault Detection

Reliable induction motor (IM) fault detection techniques are very useful in industries to diagnose IM defects and improve operational performance. An adaptive empirical mode decomposition (EMD) technology is proposed in this paper for rotor bar fault detection in IMs. As the characteristic fault frequency will change with operating conditions related to load and speed, the proposed adaptive EMD technique correlates fault features over different frequency bands and intrinsic mode function (IMF) sidebands. The adaptive EMD technique uses the first IMF to detect the fault type and the second IMF as an indicator to predict the fault severity. It can overcome the problems of the sensitivity of sideband frequencies related to the speed and load oscillations. The effectiveness of the proposed adaptive EMD technique is verified by experimental tests under different motor conditions.


Introduction
Induction motors (IMs) are commonly used in various industrial applications such as electric vehicles, machine tools and robots; in addition, IMs account for about 40% of the annual global electricity consumption [1]. Accordingly, reliable IM health condition monitoring is very useful in industries to identify IM failures to maintain motor power efficiency and avoid unexpected failures that may result in large productive and economic losses. IM defects are related to imperfections in stators, bearings, and rotors [2]. This work focuses on broken rotor bar (BRB) faults. The causes of rotor bar faults include excessive dynamic load be based on MCSA.
Several MCSA-based techniques have been proposed in literature for BRB fault detection, based on spectral analysis of sidebands of the power frequency or its harmonics using methods such as multiresolution Taylor-Kalman approach [5], low-frequency load torque oscillations [6], auto-regression analysis [7], autoregressive-spectrum analysis [8] and high-resolution parameter estimation [9]. However, similar sidebands can also occur in spectral maps of healthy motors due to rotor asymmetries caused by manufacturing/assembly tolerances and connection coupling. BRB fault can also be detected based on signal processing techniques such as the wavelet transform [10], mathematical morphology [11] and orthogonal matching pursuit [12]. In using this wavelet transform-based analysis, the estimation of instantaneous frequency is not always efficient due to frequency smoothing imperfections.
Empirical mode decomposition (EMD) can solve several of these problems [13]. The EMD decomposes the signal into different frequency bands, known as intrinsic mode function (IMF), to process non-stationary signals for IM fault detection [14] [15]. However, when the signal length becomes longer, EMD generates extra number of IMFs due to the spline interpolation, which will not only increase processing time [16] [17], but also result in some false spectral components in the frequency bands [18]. In addition, our investigation shows that the IMF magnitude also depends on the hardware setup of data acquisition (DAQ) system.
Startup transient can be used as a means to detect BRB faults [13] [19]. The efficiency of these startup-based methods depends on the length of the transient (i.e., being long enough), which is difficult to apply in applications using softstarted inverter or variable-frequency-drive -fed motors [20]. Such startup-based methods are also not suitable to monitor IM conditions in actively running operation. On the one hand, another challenge in MCSA is how to estimate the severity of the BRB fault, as the current techniques depend on an accurate estimation of the magnitude of the sidebands [21] [22], which requires very accurate DAQ systems [23].
To tackle the aforementioned problems, an adaptive EMD technique is proposed in this work for BRB fault detection in IMs. It is new in the following aspects: 1) it can extract representative features for BRB fault detection, without the need of calculating side band frequencies.
2) This new technique enables the EMD to process signals with long length and diagnose the fault feature in the time domain, which is less sensitive to fault frequency variations.
The remainder of this paper is organized as follows. Section 2 discusses the M. Mahmud, W. Wang Journal of Signal and Information Processing proposed adaptive EMD technique. Its effectiveness of in IM fault detection is verified in Section 3. Some concluding remarks are summarized in Section 4.

Overview
In general, spectral magnitude and fault sidebands are used to estimate the severity of BRB faults, but it is prone to generating false spectral components. As false frequency components do not inversely affect time series data, the proposed adaptive EMD technique will be undertaken in the time domain.
The relationship between two time series signals can be estimated by correlation coefficients of two stationary signals, but not for non-stationary signals whose statistical quantities are time varying. Fault related features from IMs could be non-stationery [23]. Different from general EMD processing, in the proposed adaptive EMD technique, the sliding window is determined adaptively by the estimation of the instantaneous period of the IMF; this provides the intrinsic relationship between the related signatures. In addition, the value of correlation is used to select the fault represented frequencies. Once the correlation value goes below to a certain threshold, the process will stop in calculating next IMF [20]. In this technique, only fault represented IMFs (which also have higher correlation coefficient) are calculated. Hence, it can process signals with any length. Figure 1 shows the block diagram of proposed adaptive EMD technique.

The Methodology
Due to the rotor asymmetry and imbalance in cage winding, there is a backward rotating field at slip frequency analogous to the forward rotating rotor [1]. As a result, current and induced electromagnetic signal are found at different slip frequencies. In addition, due to rotor oscillation, there exits an upper sideband current component in the stator winding [14]. Considering all these effects, sideband frequencies sb f due to broken rotor bars will be ( ) where 1 f is the power line frequency in Hz and s is the slip. As the EMD is based on the calculation of the envelope of maxima and minima of the input signal [18], fault related envelope can be independent of the 1 f . The line cur- where A I is the amplitude of the current, and α is the phase angle of the sinusoidal line current.
Consider the sideband components with phase angle, then Equation (2) be- where lk I and rk In the EMD process, the signal is decomposed into a set of IMFs. An IMF is a function that satisfies the following two requirements: 1) In the whole data set, the difference between the number of extrema and the number of zero-crossings points is either 0 or 1.
2) The mean value of the envelope due to the local maxima and local minima is zero.
Variable amplitude and frequency are represented as functions of time in the M. Mahmud, W. Wang Journal of Signal and Information Processing IMF. Spline interpolation can be used to determine the extreme points and form the negative and positive envelopes. The mean value of these two envelopes is subtracted from the original signal and the residual is investigated to see whether it satisfies these two criteria for IMF or not. After calculating each IMF, the correlation will be checked. If the correlation is above some threshold (usually determined by trials and errors) then the next IMF is computed.
On the other hand, from our systematic investigation, it is observed that the dot product between two IMFs is not zero. That means that EMD cannot strictly decompose IMF orthogonally. As a result, we cannot predict that fault frequency is particularly strict in one IMF, which can also be verified by correlation analysis as discussed in Section 3.
In processing, the Hilbert transform is applied to each of the IMFs such that The residual is a monotonic function: To calculate the IMF, the mean value ( ) Then the input signal for the next IMF is calculated as, It is noted that applying MCSA in the IMF is highly dependent on the bandwidth of the DAQ system. The proof of the use of multiple IMFs to detect the BRB faults and operating points will be discussed in the following section.

Experimental Setup
The experimental setup used in this test is shown in Figure 2. The tested motor (model 056T34F5301 from Marathon Electric) has 34 rotor bars and 24 stator slots. The speed of these IMs is controlled by a variable frequency drive (VFD) (from Delta Electronics Inc.). In this experiment, 50 Hz and 60 Hz supply frequencies are used, switching at 15 kHz. The IM output shaft is coupled to a gearbox (from Boston Gear) to adjust the speed ratio of the dynamometer. A magnetic particle clutch (PHC-50 from Placid Industries) is used as a dynamometer to provide variable load via an external DC power source. Current signals are collected by the use of a wireless smart current sensor DAQ system, developed in our research team [24], which can measure 2 -3 phase current signals.
The vibration signals corrected by using the 3D vibration sensor are used to estimate the motor shaft rotating speed, which is then used to compute the slip [24].

Experimental Results
The effectiveness of the proposed adaptive EMD technique is evaluated under a series of tests corresponding to different load and speed operating conditions.   IMs are tested in this work: two healthy IMs, and three faulted IMs with one, two, and three simulated BRBs, respectively. Figure 3 shows an example of an IM with two simulated BRBs.
At first, EMD is applied to the stator current. Figure 4(a) illustrates the first three IMFs of healthy motor and Figure 4(b) shows the IMFs of faulty motor using current signals at full load state. A comparison of Figure 4(a) and Figure   4(b) indicates that BRB can increase the oscillation of the IMF 2 amplitude. In this case, only the first three IMFs are calculated, as after these the correlation coefficients are almost negligible. Figure 5 shows the correlation between the signal and IMFs. It is seen that the first IMF, or IMF 1 , has the highest correlation value, as it has higher energy    IMF before doing spectral analysis. It is also noted that a new frequency pattern of ( ) To further investigate the severity of the fault under different operating conditions, Figure 7(a) shows the fluctuation of IMF 2 at 3600 RPM for 1 and 3 BRBs at the full load condition. It is seen that the magnitude of IMF 2 with 3 BRBs is greater than the magnitude of 1 BRB. Figure 7(b) illustrates the results of IMF 2 for a motor with 3 BRBs at two respective speeds of 3600 RPM and 3000 RPM, at full load condition. Figure 8 presents the results of IMF 2 at medium load state (i.e., 67% of full load) for different fault and speed conditions. It is seen that the magnitude of IMF 2 at 3600 RPM is much higher than that at 3000 RPM. Therefore, the distinction for different motor speed is clear, even though the distinction for different number of BRBs becomes less clear in this case.

Discussion of the Results
From the above result analysis, it is seen that in a 34-bar rotor, if one bar is broken it influences 1/34 times of the total operation. But if 3 bars are broken it will affect 3/34 times of the total operation. Increasing the number of BRBs will result in asymmetry and saturation in the breakage area. The adjacent bars of broken bars will draw more current in this case. As a result, local flux density and Journal of Signal and Information Processing