1. Introduction
Significant attention has been paid to the detection of broken rotor bar (BRB) faults in induction motors [
1,
2,
3,
4,
5,
6,
7,
8,
9,
10,
11,
12,
13,
14,
15,
16,
17,
18,
19]. The spectral analysis of the stator current has been widely applied to detect BRB faults [
1,
2,
3,
4,
5,
6,
7,
8]. This technique consists of monitoring of the stator current components at (1 ± 2
s)
fs (where
s is the motor slip and
fs is the supply frequency), which are related to BRB fault.
Besides the stator current, several alternative signals, such as the Instantaneous Reactive Power (IRP) and Instantaneous Active Power (IAP) [
9,
10,
11,
12,
13], the instantaneous active and reactive currents [
14] and the instantaneous active and reactive components of the current space vector [
15], were also analyzed to detect BRB fault. These techniques consist of monitoring the components at 2
sfs (twice the slip frequency) superimposed by BRB fault in the IRP, IAP and the other signals.
It has been widely recognized that load oscillation, of which the frequency
fL is next to 2
sfs, produces similar effects as BRB fault on the stator current and thus confuses the BRB fault detection based on the spectral analysis of the stator current [
10,
11,
12,
13,
14,
15,
16,
17]. Therefore, the issue of discerning BRB fault and load oscillation must be addressed. Some methods have been presented in References [
10,
11,
12,
13,
14,
15] based on the spectral analysis of the IRP, IAP and their derived signals such as the instantaneous power factor. However, all these methods perhaps fail in the adverse case that BRB fault and load oscillation are simultaneously present and especially, perforce fail in the most adverse case that
fL is absolutely equal to 2
sfs, as reported in Reference [
16].
To tackle this problem, the fifth- and seventh-order components of the stator current were used in Reference [
16]. It is well known that these components are definitely trivial compared to the
fs component and thus, an acquisition system with an extremely-high precision is required. Time-frequency analysis of the startup current was also used in Reference [
16], requiring a motor startup that should not be too short. Similarly, offline methods, for example, at standstill and manual rotation of the rotor, were recommended in Reference [
15,
17]. Certainly, all these methods are lack of practicability from the engineering point of view.
In the light of this, an improved method for discerning BRB fault and load oscillation is proposed in this paper.
It is noted that the theoretical basis for discerning BRB fault and load oscillation has so far been limited: these two phenomena are individually considered rather than simultaneously when deducing the analytical expressions of the IRP, IAP and their derived signals [
10,
15]. Consequently, there is a lack of theoretical support for discerning the simultaneously-present BRB fault and load oscillation. Thus, the analytical expressions of the IRP and IAP in the adverse case that BRB fault and load oscillation are simultaneously present are deduced in this paper. By comparing them with those previously deduced in References [
10,
13], some important findings on the spectra of the IRP and IAP are obtained. Such work extends the theoretical basis for discerning BRB fault and load oscillation and moreover, yields a novel strategy, which can correctly discern the simultaneously-present BRB fault and load oscillation, even in the most adverse case that
fL is absolutely equal to 2
sfs.
In addition, all the methods presented in References [
10,
15] adopt Fast Fourier Transform (FFT) as the benchmark technique for spectral analysis. Thus, they all suffer the intrinsic limitations of FFT: (1) it is only effective for stationary signals; (2) it requires a long acquisition time to get a high frequency-resolution (these two parameters are inversely proportional). Such problem has been widely reported and addressed [
4,
5,
6,
7,
8,
18]. As an example, Estimation of Signal Parameters via Rotational Invariance Technique (ESPRIT) was adopted to replace FFT in References [
6,
7,
18] for the BRB fault detection, yielding significant improvements. This ascribes to the fact that ESPRIT can easily achieve a high frequency resolution even with a short acquisition time and thus, is definitely superior over FFT. Therefore, ESPRIT is adopted in this paper to analyze the IRP and IAP.
The above work yields an improved method for discerning BRB fault and load oscillation.
3. An Improved Method for Discerning BRB fault and Load Oscillation
In order to overcome the existing drawback stated above, an improved method for discerning BRB fault and load oscillation is proposed in this paper.
3.1. The Extension of Theoretical Basis
In the case that BRB fault and load oscillation are simultaneously present, the complex stator current vector can be expressed as:
where
,
,
,
and
are given by (1), (3), (4), (7) and (8), respectively.
With a sinusoidal supply voltage given by (10), the IRP and IAP are deduced as (18) and (19), respectively, using the similar procedure shown in
Section 2.
The most adverse case for discerning BRB fault and load oscillation is that
fL is absolutely equal to 2
sfs. In such a case, the expression of the IAP can be yielded from (19) as (20), where
Icom and
εcom are given by (21) and (22), respectively.
It is noted that (18) is identical with (13), leading to the following finding:
Finding (a): The IRP and especially, Aq is mainly dependent on BRB fault and thus, largely independent on load oscillation.
Therefore, the
Aq based BRB fault detection method is actually immune to load oscillation. In other words, this kind of method is always effective no matter load oscillation is present or not. Certainly, an appropriate index should be established to evaluate the severity of BRB fault. This index should be sensitive to BRB fault, while robust to other factors such as load level. In Reference [
10], a severity factor of BRB fault was established as the relation between
Aq and the average active power
Pav:
As revealed in Reference [
10],
SFq is robust to load level. In the meantime, as can be inferred from the above finding,
SFq is independent on load oscillation. Moreover, it is worth mentioning that
SFq is in nature related to the severity of BRB fault. According to (18) and (19),
SFq can be further deduced as:
As demonstrated in Reference [
19], the severity of BRB fault can be quantitatively evaluated according to the following relationship:
where
n is the number of the broken rotor bars;
N is the number of the total rotor bars.
Substituting (25) into (24) yields:
To sum up, SFq is an appropriate index to evaluate the severity of BRB fault. The SFq based BRB fault detection method can thus be expected to produce a satisfactory result. Therefore, SFq is readily used to identify the presence of BRB fault in this paper.
Finally yet importantly, an appropriate threshold should be correspondingly pre-determined for diagnosing BRB fault is present or not.
According to (26), the threshold can be set as:
which corresponds to the case of half broken rotor bar.
However, it should be noted that (25) only holds for large motors operating with about full load, as stated in Reference [
19]. Therefore, the threshold set according to (27) may be inappropriate for other applications, for example, a small motor operating with half load and thus, should be refined. As a common practice, this threshold can be refined by means of some a priori knowledge of the investigated motor. As an example, the relationship of
SFq versus the number of broken rotor bars and the load level was successfully used in Reference [
10] to obtain the appropriate threshold. This kind of practice is readily adopted to deal with a small induction motor investigated in this paper.
Finding (b): The IAP and especially, Ap is mainly dependent on load oscillation and thus, largely independent on BRB fault.
As demonstrated in Reference [
10], in the case of BRB fault, it stands that
Aq >>
Ap. Therefore, a numerical relationship can be deduced from (13) and (14) as:
Actually, such relationship has already been demonstrated in Reference [
14] (see (15), (16) and the corresponding statements in Reference [
14]).
Substituting (28) into (21) yields
Substituting (29) into (20) and comparing it with (16) lead to the above finding. In fact, this finding can also be obtained by substituting (28) into (20) and comparing it with (16).
This finding tells that the presence of load oscillation can be identified by using Ap and moreover, this kind of identification is immune to BRB fault. Certainly, an appropriate criterion should be established for diagnosing load oscillation is present or not.
The ratio of
Aq and
Ap can be obtained from (18)–(21) and (28)–(29) as:
Observing (4) and (8) yields
Substituting
fL =
2sfs into (31) yields
Substituting (32) into (30) yields
Note that and TP in (33) are the oscillating torque amplitudes arising from BRB fault and load oscillation, respectively.
Here assumes that is not much greater than TP. Such assumption is valid due to the following reasons.
(1) is usually small, for example, a few percentages of the average torque, especially at the early stage of BRB fault. Thus, if a certain load oscillation gives rise to an oscillating torque, whose amplitude TP is much smaller than , this kind of load oscillation is practically negligible.
(2) Now that a certain load oscillation can affect the BRB fault detection, it should give rise to an oscillating torque whose amplitude TP is at least comparable to .
With the above assumption, it can be easily seen from (33) that Aq will not be much greater than Ap if load oscillation is present.
Thus, the criterion for diagnosing load oscillation is present or not can be established as: if Aq >> Ap, then load oscillation is negligible; otherwise, load oscillation is present.
The above work yields a novel strategy for correctly discerning BRB fault and load oscillation even simultaneously present, outlined in
Figure 1.
3.2. The Use of ESPRIT
As mentioned above, ESPRIT can easily achieve a high frequency resolution even with a short acquisition time and thus, is definitely superior over FFT, which was taken in Reference [
10,
11,
12,
13,
14,
15] as the benchmark technique for spectral analysis. Therefore, ESPRIT is adopted in this paper as the benchmark technique rather than FFT to conduct the spectral analysis of the IRP and IAP for a certain improvement compared to the existing methods presented in Reference [
10,
11,
12,
13,
14,
15].
In fact, ESPRIT has been successfully adopted to improve the BRB fault detection [
6,
7,
18]. Especially, by adopting ESPRIT rather than FFT to conduct the spectral analysis of the Hilbert modulus, significant improvements have been noted in Reference [
18]. Note that the IRP and IAP both share the same frequency-composition with the Hilbert modulus. As an example, they all contain the DC component and the 2
sfs component, which is related to BRB fault. The background in this paper is also same as that in Reference [
18] in terms of monitoring the 2
sfs components and thus, diagnosing BRB fault is present or not. It can thus be inferred that a certain improvement can be expected by adopting ESPRIT rather than FFT to conduct the spectral analysis of the IRP and IAP for discerning BRB fault and load oscillation.
The basics of ESPRIT have been established in Reference [
20,
21,
22]. The implementation procedure of ESPRIT for detecting BRB fault has been described in Reference [
6,
7,
18]. Refer to them for detailed information about ESPRIT.
It is worth mentioning that, besides ESPRIT, quite a few advanced signal-processing techniques have been adopted for improving the BRB fault detection, such as ZFFT in Reference [
4], MUSIC and ZSC in Reference [
5] and Taylor-Kalman filters in Reference [
8]. A comprehensive review on this can be found in Reference [
23]. In the future work, all these techniques will be successively tried.
3.3. The Improved Method for Discerning BRB Fault and Load Oscillation
Based on the above work, an improved method for discerning BRB fault and load oscillation is proposed. The basic steps used are as follows:
- (1)
Calculate the IRP q and remove the DC component, yielding the pulsating component Δq.
- (2)
Calculate the IAP p and remove the DC component, yielding the pulsating component Δp.
- (3)
Conduct ESPRIT on Δq, yielding the ESPRIT spectrum ΔqESPRIT.
- (4)
Conduct ESPRIT on Δp, yielding the ESPRIT spectrum ΔpESPRIT.
- (5)
Investigate ΔqESPRIT to identify the possible 2sfs and fL peaks, labeled as ΔqBRB and ΔqLO, respectively. Note that the former should be greater in amplitude than the latter if they are both present.
- (6)
Investigate ΔpESPRIT to identify the possible fL and 2sfs peaks, labeled as ΔpLO and ΔpBRB, respectively. Note that the former should be greater in amplitude than the latter if they are both present.
- (7)
If Δ
qBRB and Δ
qLO can be clearly discerned, as well as Δ
pLO and Δ
pBRB, it indicates that there exists a discernible gap between
fL and 2
sfs, the procedure in Reference [
10,
11,
12,
13] can thus be followed.
- (8)
Otherwise, it means that
fL is approximately or absolutely equal to 2
sfs, the strategy outlined in
Figure 1 is then recommended.
As demonstrated by Finding (a), the first decision in
Figure 1 will satisfactorily achieve the diagnosis on the presence of BRB fault, no matter load oscillation is present or not. Moreover, as demonstrated by Finding (b), the second decision in
Figure 1 will satisfactorily achieve the diagnosis on the presence of load oscillation. Thus, the correct discernment of BRB fault and load oscillation will be achieved.
4. Simulation and Experimental Validation
Simulation and experimental results are provided in this section to validate the proposed method. Herein, the adverse case that BRB fault and load oscillation are simultaneously present is addressed.
The test motor is a small induction motor, typed Y100L-2 and rated at 3 kW, 380 V, 50 Hz and 2880 r/m.
The results were sampled up to 10,000 hits with a sampling frequency of 1000 Hz and then, analyzed by means of FFT and ESPRIT, respectively.
The simulation is based on the multiple-coupled-circuit theory, in which the motor is assumed ideal with symmetrical construction, sinusoidal distributed windings and so forth. Detailed information can be referred to [
1,
3].
Initially, the simulation is set under the condition that load oscillation is not present to obtain the relationship of
Aq versus the number of broken rotor bars and the load level, as shown in
Figure 2a. Meanwhile, the relationship between
SFq and these two factors is obtained, as shown in
Figure 2b. Pay attention that both of
Aq and
SFq are definitely zero in the case of healthy motor here assumed ideal.
The simulation is then set under the condition that the motor operates with one broken rotor bar, 50% rated load and an oscillating load complying with (6), where TP is fixed at 2.5% rated load, fL and (αT + δ) are set to be 1.56 Hz and π/2, respectively. Note that 2sfs is 1.56 Hz under this condition and thus, fL is actually set to be absolutely equal to 2sfs.
Figure 3 shows the corresponding results. The FFT spectra of the IRP and IAP are given in
Figure 3a,b, respectively. In the meantime, the ESPRIT spectra of the IRP and IAP are given in
Figure 3c,d, respectively.
Note that the DC components in the IRP and IAP have been pre-filtered out to highlight the interested 2sfs and fL components.
As can be seen from
Figure 3,
Aq ≈ 57 Var,
Ap ≈ 71 W and thus, it does not stand that
Aq >>
Ap, the criterion used in Reference [
10,
11,
12,
13] to indicate the presence of BRB fault. Therefore, the existing methods presented in Reference [
10,
11,
12,
13] will give a false diagnosis that BRB fault is not present.
In addition,
Pq −
Pp is calculated to be about π/9 for the simulated condition. This is significantly different from the criteria of |
Pq −
Pp| ≈ π/2 and
Pq −
Pp ≈ π to indicate the presence of BRB fault and load oscillation, respectively. Thus, although the phase information is used as an auxiliary, the method presented in Reference [
11] fails.
In fact, as can be seen from (18), (19) and (20), Pq − Pp is stochastic since δ is actually stochastic. Therefore, in the authors’ opinion, the phase information is not suitable for discerning BRB fault and load oscillation.
This instance demonstrates the aforementioned drawback: the existing methods for discerning BRB fault and load oscillation could fail in the adverse case that BRB fault and load oscillation are simultaneously present.
Now evaluate the improved method proposed in this paper. According to
Figure 2b, the threshold for the BRB fault detection can be set as 2.0%. Note that such a threshold is noticeably larger than the
SFq in the case of healthy motor and in the meantime, noticeably smaller than the
SFq in the case of one broken rotor bar. Thus, it is appropriate to guarantee the sensitivity and reliability of the BRB fault detection. Here
SFq =
Aq/
Pav × 100% = 57.14/1421.2 × 100% = 4.02% is greater than the threshold of 2.0%. Thus, the first decision in
Figure 1 will reach the correct diagnosis that BRB fault is present. Meanwhile, the second decision in
Figure 1 will correctly indicate the presence of load oscillation since
Aq (≈57 Var) is not greater than
Ap (≈71 W). Therefore, the proposed method achieves the correct discernment of BRB fault and load oscillation, which are simultaneously present, even in the most adverse case that
fL is absolutely equal to 2
sfs. Certainly, this is a great improvement compared to the existing methods presented in Reference [
10,
11,
12,
13,
14,
15].
In addition, as can be seen from
Figure 3, the ESPRIT spectra exhibit a clearer visibility than the FFT spectra, demonstrating that ESPRIT is preferable to FFT in terms of frequency resolution.
Experiments were conducted to further evaluate the proposed method. The test bench is composed of the test motor and a DC generator as the load, as shown in
Figure 4a. BRB fault is emulated by drilling off the rotor bar in the region close to the end-ring joint where BRB fault usually starts, as depicted in
Figure 4b,c. The load level can be set by adjusting the resistor bank connected to the armature terminals and the field voltage of the DC generator. In the experiments, the needed oscillating load is generated by a customized control system. It is composed of two adjustable resistors (
R1 and
R2) and an electronic switch (MOSFET) controlled by the digital output pulse of a DSP (OMAP-L138), as shown in
Figure 4d. The needed
Tp and
fL can be obtained by adjusting
R1 in accordance with
R2 and the MOSFET switching frequency, respectively.
The relationship of
Aq versus the number of broken rotor bars and the load level is shown in
Figure 5a. Meanwhile, the relationship between
SFq and these two factors is shown in
Figure 5b.
According to
Figure 5b, the threshold for the BRB fault detection can be set as 0.5%, noticeably larger than that in the case of no broken rotor bar and in the meantime, noticeably smaller than that in the case of one broken rotor bar. Such a threshold can guarantee the sensitivity and reliability of the BRB fault detection.
The test was set under the condition that the motor operated with one broken rotor bar, 50% rated load. An oscillating load complying with (6) was introduced, where the amplitude TP was fixed at 3.0% of rated torque and the oscillating frequency fL was set as 1.53 Hz. Under this condition, 2sfs was approximately 1.50 Hz.
Figure 6 shows the experimental results. The FFT spectra of the IRP and IAP are given in
Figure 6a,b, respectively. For comparison, the ESPRIT spectra of the IRP and IAP are given in
Figure 6c,d, respectively.
As can be seen from
Figure 6,
Aq ≈ 26 Var,
Ap ≈ 63 W and thus,
Aq >>
Ap used in Reference [
10,
11,
12,
13] to indicate the presence of BRB fault does not stand. Therefore, the existing methods presented in Reference [
10,
11,
12,
13] fails to give the correct diagnosis that BRB fault is present.
Again, the aforementioned drawback is demonstrated: the existing methods for discerning BRB fault and load oscillation could fail in the adverse case that BRB fault and load oscillation are simultaneously present.
Certainly, the improved method proposed in this paper succeeds. Here
SFq =
Aq/
Pav × 100% = 25.55/1640.4 × 100% = 1.56% is greater than the pre-set threshold of 0.5%. Thus, the first decision in
Figure 1 will reach the correct diagnosis that BRB fault is present. Meanwhile, the second decision in
Figure 1 will correctly indicate the presence of load oscillation since
Aq (≈26 Var) is not greater than
Ap (≈63 W). Therefore, the proposed method achieves the correct discernment of BRB fault and load oscillation, which are simultaneously present, even
fL is approximately equal to 2
sfs. Certainly, this is a great improvement compared to the existing methods presented in Reference [
10,
11,
12,
13].
In addition, as can be seen from
Figure 6a,b, the FFT spectra fail to discern the
fL and 2
sfs peaks. However, as can be seen from
Figure 6c,d, the ESPRIT spectra achieve to correctly discern them. Herein,
Aq >>
Ap at 2
sfs (1.49 Hz), indicating the presence of BRB fault. Meanwhile,
Ap >>
Aq at
fL (1.53 Hz), indicating the presence of load oscillation. Thus, the correct discernment of BRB fault and load oscillation is achieved. This ascribes to the superiority of ESPRIT over FFT in terms of frequency resolution.
A series of experimental results demonstrate that the ESPRIT spectra of the IRP and IAP can always correctly discern the simultaneously-present 2
sfs and
fL peaks, as long as the gap between them is not lesser than 0.04 Hz. This means a certain improvement compared to the FFT spectra. However, this kind of improvement is limited, since the ESPRIT spectra still fail to discern the
fL and 2
sfs peaks in the most adverse case that
fL is absolutely equal to 2
sfs, as can be seen from the simulation results,
Figure 3c,d. Certainly, such a case is seldom encountered from the engineering point of view and thus, ESPRIT is still preferable to FFT.