1. Introduction
The three-phase induction motors (IMs) are the main electrical rotating machine installed in many industrial environments. Usually, the induction motors are capable of driving various types of load, thus this kind of machine is widely used around the world [
1,
2,
3,
4]. Although IMs have important advantages, such as robustness, simplicity and lower cost, when compared to other rotating machines (synchronous machines and DC motors, for example) [
5], they are subjected to some mechanical and electrical faults, particularly in stator windings, bearings and rotor cage [
1,
6].
Many researchers and engineers have investigated broken rotor bars and bearing failures in other applications, including commercial cases and industrial plants. Typically, broken rotor bars and cracked end-ring faults share for 5–10% of induction machine failures, but, as cited in [
7], these events are a key issue. As described by [
8], for medium-voltage (MV) motors the rotor cage fault is even more common than that of small machines due to the extensive thermal stresses on the rotor. A partial or a fully broken bar increases the machine vibration [
9], the current in the rotor bars adjacent to the faulty one and the temperature rise in the motor [
7].
The rotor cage fault can also lead to bearing failures and air gap eccentricity [
8]. Other researchers, for example, have been discussed the fault in the rotor related to adjacent and nonadjacent broken bars [
10,
11]. Some cases with nonadjacent broken bars are also observed for frequently started motor applications. In [
12], nonadjacent broken bars for 6.6 kV, 500 kW, ten-pole induction motor has been reported and although the fault frequency component increases with the number of broken bars, this component is quite difficult to observe if the damaged bars are 90
(electrical) apart since the asymmetry in the rotor is canceled out [
10,
11,
13].
Particularly for medium voltage (MV) induction motors, the costs related to machine shutdown are much more significant when compared to small motor applications. In [
12], an example of a 6.6-kV 2400 kW eight-pole induction motor was described for rotor under inspection due to a false positive (FP) indication. In this case, an unexpected failure of the motor would result in an estimated financial loss of US 2.5 to US 4.0 million (loss of revenue and repair cost) and a period of 8 to 12 weeks to recover normal plant operation. In other MV applications (Offshore Oil Production Platform), a 11 kV 1950 kW squirrel cage induction motor inspection has been related for broken rotor bars evaluation and a motor shutdown could take three months of downtime from start to finish and a total financial loss of approximately US 2 million [
13]. As cited in [
4], for larger motors, longer downtime per failure is related to induction motors starting more than once per day or in cases wherein applications of pulsating load or direct online startups. As aforementioned, in MV motors, the rotor cage fault is even more common than that of small machines due to the extensive thermal stresses on the rotor. In addition, large motors, with hundreds or thousands of kilowats, usually have more rotor bars when compared to small motors; thus, as described by [
8], a severe failure is necessary in those rotor cages for successfully applying the traditional motor current signature analysis (MCSA) technique, for rotor conditioning evaluation.
The MCSA has been used in the past two decades to detect broken rotor bars, particularly due to its noninvasive approach [
14], by applying fast Fourier Transform (FFT) in only one current phase of the motor [
15,
16].
The aforementioned MCSA technique is a comprehensive approach for broken rotor bars detection and recognized by researchers and electrical engineers as an efficient diagnosis tool, since this solution is capable of identifying failures in the rotor cage in many cases and applications using only one phase current.
However, MCSA has some limitations and drawbacks well known in literature such as the spectral leakage detected when the motor is working at very low slip, for example, since the amplitude of frequency components, i.e., the left sideband component responsible for an index fault is close to the line frequency [
17,
18,
19,
20]. It should be mentioned that, in a closed-loop control structure, for example, the harmonics cannot be directly applied, as cited by [
21].
The work published by [
17] proposed the use of a Hilbert Transform and FFT to extract a fault frequency index for motor running at low slip. However, this approach also requires a long measurement time for current signal processing. It should be noted that MCSA requires an acquisition time of 100 s for current signal processing to extract the sidebands for failure evaluation. For both cases, it is not possible to detect the fault in transient condition, but only in steady-state, since the slip may change in 100 s.
It is important to highlight that large motors usually operate at low slip even at rated load. In [
18], an improvement of the Hilbert method via estimation of signal parameters with rotational invariance technique (ESPRIT) has been proposed to detect broken bars in induction motors running at low slip. In this case, a measurement time equal to 10 s was used for current signal processing and the motor was line-fed.
Recently, some typical root causes of false positive and false negative indications have been reported using MCSA based rotor fault detection in MV induction motors [
12]. The false negative conditions are related to nonadjacent broken bars, load variation and an incorrect speed estimation. The MCSA technique requires an accurate value of the rotor speed (slip) for reliable detection of rotor faults. The work published by [
12] still disclosed that, for MV motors, it is necessary an acquisition time of at least 30–60 s for a correct broken bar diagnosis using MCSA.
In [
22], for example, it was described that, in many applications, the rotational speed and the demanded torque of the machines change significantly over time; thus, it is very difficult or almost impossible to have a long enough steady period of time, such as required by MCSA to detect broken rotor bars. The work proposed by [
23] requires the slip information to extract some fault frequencies from vibration and sound of the machine, but the authors highlighted that, in field applications, motor speed is not always available. The typical root causes of false positive (FP) and false negative (FN) indications produced by MCSA-based rotor fault detection are: load variation (FN), incorrect speed estimation (FN) and low frequency load oscillations (FP), as cited by [
12,
24,
25]. As mentioned by [
8], load oscillation can also induce current harmonics at the same frequencies in the stator current, such as those found in damaged cage rotors.
Recently, some researchers have also investigated the broken rotor bars in induction motors fed by an inverter. As cited by [
26,
27,
28], the rotor fault detection in converter-fed induction motors has some drawbacks when compared to a sinusoidal supply situation, particularly due to the additional harmonics that will be induced in the current spectrum. Other works have also reported difficulties in detecting rotor faults for the motor fed by an inverter, such as in [
22,
29,
30,
31]. In [
32], a method for broken rotor bars detection relies on monitoring certain statistical parameters estimated from the analysis of the start-up stator current envelope. In this case, the simulations results were carried out for motor running under direct online start and inverter-fed modes, but this research did not address cases of load variation and/or load oscillation.
In [
5], the analysis of the startup current (ATCSA) of the motor was also proposed to detect broken rotor bars. This approach applies time-frequency (T-F) transforms for the continuous analyses, i.e., a Short Time Fourier Transform (STFT) and a Discrete Wavelet Tranform (DWT) for the computation of a fault severity indicator in line-fed motors. This paper shows an interesting application of the ATCSA method in four induction motors operating in mining facilities, but this technique requires an specialist interpretation of a V-shaped image responsible for failure evaluation and the presence of several breakages in the rotor cage for a relevant fault severity indicator (DWT should lead to a value lower than the threshold level 50 dB). This work did not address the use inverter-fed mode. The authors disclosed that the method is able to detect the rotor failure with accuracy severity levels around one broken bar out of 28, but large induction motors usually have more than four or five tens of bars.
Table 1 shows the specifications for some induction motors with lots of rotor bars.
It is important to mention that the work published by [
36] has shown a survey of existing broken rotor bar fault detection techniques based on fault signatures analysis and the authors pointed out that time-frequency representation using Wavelet Transform is a powerful tool, but it suffers from some drawbacks, as the need for optimum selection of the mother wavelet and the overlap between adjacent frequency bands. The same work also described that other researchers are investigating the limitations of the detection based on MCSA, especially the load variation, since it produces frequency components in the current spectrum close to the broken bar fault indicator.
In addition to the processing methods, other works also employed different approaches to identify faults in rotational electrical machines, such as computational intelligence algorithms and/or machine learning techniques. In [
37], for example, novel insights have been discussed for the classifier evaluation in the field of electric machine diagnosis using Decision Tree (DT) and support vector machine (SVM). In this case, the authors pointed out that the choice of a correct classifier depends enormously on the accurate evaluation of its performance, in order to reduce the occurence of false diagnosis in predictive maintenance. A case study was presented in [
37] and a rotor condition of a small induction motor (0.75 kW) has been tested using the MCSA method. However, this approach did not consider the load variation for validation purposes.
In [
30], four different learning machine techniques were investigated for broken rotor bars detection in a three-phase induction motor fed by an inverter. In this case, a fuzzy ARTMAP network, SVM, a k-nearest neighbour (KNN) and a multilayer perpetron network (MLP) have been tested for broken bars evaluation using the motor current as a source signal for an acquistion time duration of 6 s. This work has shown the potential of the machine learning approaches for failure detection at different speeds, but the load variation and load oscillation did not address.
In [
38], a neuro-fuzzy approach was applied to locate broken rotor bars in induction motors running at very low slip. In this work, a Fast Fourier Transform was used to extract the features from a Hall effect sensor installed inside the machine and it was used an acquisition time duration of 4 s. The processed signal was the magnetic flux density measured by the Hall sensor. This research also did not address the rotor fault detection in variation or oscillating loads, only in steady-state condition.
In [
39], three learning machine techniques (MLP, KNN and SVM) were used to detect and classify rotor faults also running at low slip. In this case, some frequency and statistical features have been used to evaluate the rotor condition in an acquistion time of 4 s for motor running at steady-state. In [
40], an investigation of vibration and current monitoring for effective fault prediction in an induction motor has been proposed using a multiclass support vector machine algorithm and in [
41], and a lot of signal processing and feature extraction techniques were described for fault diagnosis in rotating machines, such as time-domain and frequency domain approaches, as well as the use of KNN, SVM and Fuzzy c-Means classifiers, among others.
Although the works published by [
39,
40,
41,
42] have successfully used machine learning methods for electrical and/or mechanical failures diagnosis, they did not address the problem related to load variation or load oscillation for broken rotor bars detection.
As can be seen, today, a wide time-domain and frequency-domain signal processing techniques are used for feature extraction and the rotor condition monitoring systems are also strongly related to the use of computational intelligence techniques or machine learning approaches, due to a huge amount of data acquired by sensors.
This paper presents the use of a fuzzy inference system (FIS) to extract the features of a Hall sensor signal, installed inside the motor, according to the rotor condition. The FIS is able to quantify crisp values from a physical environment, i.e., by using the uncertainty representation of the magnetic flux disturbances. As described by [
43], the vast majority of the related studies is focused on noninvasive techniques, such as the measurement of stator current and vibration, among others. These approaches are understandable from a cost perspective, since an external sensor is easy to install; thus, it means that a small extra cost is necessary compared to the cost of a standard induction motor. However, [
43] states that there are special motors designed for particular applications and these machines are usually expensive and not easy to replace. In this case, large induction motors (tens or hundreds of kilowatts) often operate in extreme dynamic environments under increased stresses. Therefore, a new apparatus that helps preventing an unexpected downtime is an acceptable, or interesting, option.
The analysis of broken bars is taken into account for time-domain purposes, including the signal processing of the signal and the data classification using an SVM classifier. The reason for using this machine learning classifier is justified with earlier studies of multi-class classifiers [
40,
41].
Based on the aforementioned literature, this paper proposes a method to detect fully broken rotor bars in large IMs, since this kind of motor usually has lots of rotor bars and also reduces the false positive and false negative indications for rotor cage condition evaluation. For these purposes, the present approach is focused on the following contributions:
- (i)
It is not necessary to estimate rotor slip, as required by the MCSA method and other techniques (related to FN indication);
- (ii)
It is possible to detect broken bars during load variation (related to FN indication);
- (iii)
It is possible to detect broken bars during low frequency load oscillations (related to FP indication);
- (iv)
It is possible to classify the severity of rotor faults;
- (v)
It is possible to detect broken rotor bars for motor running at low slip (related to FN indication); and
- (vi)
The method allows the rotor fault detection for motor fed by an inverter and also fed by sinusoidal power supply.
The present approach has been validated by using simulation results of two IMs (1200 kW and 7.5 kW) and experimental tests have been performed using an induction motor of 7.5 kW, 220 V and 38 rotor bars, fed by a sinusoidal supply and by an inverter. The methodology and the results are presented in the following subsections.