Motor Current Signature Analysis (MCSA): A Comprehensive Guide
The reliability and efficiency of machinery are of paramount importance in modern industry. As a result, online condition monitoring and fault diagnosis (FD) of electromechanical systems have become crucial tasks in many industrial applications. Maintenance strategies such as condition-based maintenance (CBM) and predictive maintenance (PM) have therefore gained increasing importance, especially with the introduction of Industry 4.0.
Condition monitoring, upon which CBM and PM are based, involves tracking the system state of health to detect fault occurrences. Any condition monitoring (CM) procedure requires sensor measurements and analysis of signals (such as vibration, speed, current, and temperature) that are relevant to the performance state of the system’s components.
Although vibration analysis remains the most widely used method, motor current signature analysis (MCSA) has gained increasing attention in recent years. MCSA is a non-intrusive and cost-effective approach when compared to vibration-based techniques, as it relies only on the motor phase currents, which are already used in motor control. Other advantages over vibration analysis include simpler sensor installation, reduced sensitivity to sensor installation location, and less sensitivity to external factors such as background noise.
Initially, this technique was adopted only to monitor electric motor components (e.g., windings, rotor, bearings, resistance); this allows the tracking of the motor’s internal state of health, but it does not provide insight into the condition of the mechanical systems driven by the motor. In the last years, MCSA has also been applied to fault diagnosis of components in the mechanisms attached to the electric motor, such as gears, bearings, belt conveyor systems, and rotate vector reducers.
Figure 1. Time Domain Format
Fundamentals of MCSA
Motor current signature analysis (MCSA) is an interesting noninvasive alternative to vibration analysis for the condition monitoring and fault diagnosis of mechanical systems driven by electric motors. The MCSA approach is based on the premise that faults in the mechanical load driven by the motor manifest as changes in the motor’s current behavior.
Motor Current Analysis (MCA) is a diagnostic technique used to evaluate the condition of electric motors and the equipment they drive. A non-intrusive diagnostic and monitoring technique that analyzes the electrical current signatures of motors to detect mechanical and electrical issues in both the motor and the driven equipment. Different failure mechanisms can impact the current sine wave, and each type of fault, such as pump cavitation or bearing damage, displays unique patterns.
The motor current signature is recorded in a time domain format. The current is represented in a graph form with the amplitude shown on the “Y” axis and the time on the “X” axis. In order to analyze the data, a Fast Fourier Transform (FFT) is performed. An FFT is a mathematical operation designed to extract the frequency information from the time domain and transform it into the frequency domain. While the FFT spectrum is a great source for identification of rotor bar problems in motors, it proved difficult to analyze most other frequencies.
Figure 2. Demodulated Spectrum
Demodulated Current Spectrum
In recent years, one of the most exciting advancements in PdM technologies is the demodulated current spectrum. Modulation is when lower frequencies are merged on top of a higher frequency. In other words, lower frequencies ride on the higher frequency signal. This makes the carrier frequency the dominant peak in the FFT spectrum, and most of the information is lost in the noise floor of the spectrum.
Demodulation is simply the process of taking the carrier frequency out of the spectrum. In this case, the carrier frequency is the fundamental electrical frequency being used. The fundamental frequency in the United States is 60 Hz. In many other countries, it is 50 Hz. Frequencies such as speed, pole pass, belt pass, vane pass, gears and bearing frequencies can be identified and trended in the demodulated current spectrum.
Due to the technology being relatively new, trending remains the most accurate method of identifying a problem in a machine. The ability to have baseline data when the machine is in good health is ideal, while comparing data to similar machines is also very effective.
Figure 3. Belted machine current spectrum
The basic current signature of a belt-driven machine is shown in Figure 3. It shows the fundamental 60 Hz frequency being the dominant peak. Note the sideband peaks on each side of 60 Hz that are labeled with their frequencies. This type of signature has the potential of being misconstrued as a potential rotor bar problem. In this example, the sidebands are actually related to belt pass frequency. The mechanical frequency of interest in this case would then be calculated by multiplying 6.5 Hz by 60. The result would be 390 RPM, which is the belt pass frequency of this machine. Rotor bar problems show up at pole pass frequency, with the peaks much closer to the fundamental frequency.
Although the mechanical peaks in Figure 3 are prevalent, this is often not the case. Usually, the mechanical peaks will be lost in the noise floor of this type of spectrum. Figure 4 shows the demodulated spectrum derived from the current signature in Figure 3. Notice how much cleaner and easier to read this example appears. The sideband frequencies shown in Figure 3 now show up in Figure 4 at 6.5 Hz along with a 2x and 3x belt pass frequency in the demodulated spectrum.
Looking closely at the current spectrum, the 2x and 3x frequencies are present but not as easily identified. Both drive and driven speeds will also be found in the spectrum if there is a problem. As in vibration, a 1x rotational speed will signify an unbalanced effect on the machine. In Figure 4, the fan speed shows up at just over 25 Hz. Normally, this peak will be at or near the noise floor of the spectrum. The amplitude response on the spectrum is remarkably sensitive. A noticeable amplitude increase will occur with the incorrect key length in the hub.
Figure 4. Pump Assembly with Problem
Figure 5. Follow-up Test
An example of a brittle flexible coupling with stress cracking is shown in Figure 5. Figure 6 shows the follow-up test completed after a new coupler and laser alignment was performed.
Another common frequency found in a demodulated current spectrum is the pump vane pass and fan blade pass frequency. This will help to trend and locate problems with the impeller or flow restrictions.
With all that demodulated MCSA can do, there are still many questions as to how it will fit into and benefit a PdM program. Frequently asked questions might include: Why do I care about finding mechanical faults with a demodulated current signal if I can find them with technologies like vibration analysis or infrared thermography? How reliable is the data that is generated and can it take the place of vibration?
With any condition monitoring technology, there are strengths and weaknesses. Each technology applied will give a more complete view of the health of the equipment. For best results, it is recommended to complete MCSA at least quarterly. If a program is testing less frequently than this, the overall results of the motor testing program will be compromised.
As for MCSA technology looking for mechanical faults, there are many reasons why this can benefit a PdM program. For example, when it comes to belt and coupler problems, demodulation will give an earlier and often more accurate fault indication than vibration analysis. The amount of energy created by the early stages of this type of fault is relatively low. A demodulated current spectrum has the ability to detect the fault early enough to provide plenty of time to plan and schedule the repairs.
However, demodulated MCSA is not intended to take the place of a vibration program. An added benefit of this technology would be in remote equipment locations or areas where equipment is not accessible during normal operations. On this type of equipment, visual inspections can be difficult, and the ability to perform vibration analysis is limited. Depending on the risk assessment, remote wiring transducers for vibration may be too costly.
Part of any strong PdM program is having the ability to verify a fault with more than one technology. This not only ensures the validity of the fault but also helps make a more accurate and precise repair recommendation. The importance of verification with a second technology is never more evident than on a critical piece of equipment that requires plant outages for repair.
Monitoring motors using advanced techniques like Motor Current Signature Analysis (MCSA) is crucial for condition monitoring of induction motors, especially in challenging environments such as nuclear power plants. The core principle of MCSA is that the electrical current drawn by a motor is directly influenced by its operating condition.
To implement MCSA, sensors are strategically installed inside the motor control cabinet (MCC). The collected data includes current and voltage measurements, which are then subjected to advanced algorithms designed to detect anomalies. Fault detection and classification are important aspects of motor current analysis.
MCSA is adept at identifying faults, including broken rotor bars, bearings, and stators. Once an anomaly is detected, MCSA categorizes the fault based on its unique characteristics. This classification process involves analyzing the frequency and amplitude of the fault signal, which provides crucial information about the nature and severity of the issue. Current spectrum analysis is a fundamental component of motor current analysis.
Current spectrum analysis allows for a precise and detailed understanding of the motor’s operating condition.
One of the primary challenges is the need for a substantial amount of data to achieve accurate and reliable results. Instead, it is a complementary technology that provides additional insights into motor conditions. Despite these challenges, motor current analysis has substantial benefits.
This is particularly valuable in applications with impractical traditional condition monitoring techniques, such as nuclear power plants or other hazardous environments. As the condition monitoring industry continues to evolve, motor current analysis is expected to grow, driven by its ability to provide early fault detection and support predictive maintenance strategies.
Advanced Techniques in MCSA
To extract features for FD from the raw current signals frequency domain, time-frequency domain, or combinations of time and frequency domain techniques have been proposed in the literature. In [9], the discrete wavelet transform (DWT) is combined with fast Fourier transform (FFT) to trace the sidebands of the gear mesh frequencies (GMF). In [10], amplitude demodulation and frequency demodulation are applied to the current drawn by the induction motor to detect the rotating shaft frequencies and GMFs. Then, DWT is applied for denoising.
The effects induced by gear tooth surface damage faults on the torque oscillation profile are investigated in [11]. In particular, it is shown that such effects produce fault-related frequencies in the stator current spectrum and specific harmonics as integer multiples of the rotation frequency in the stator current space vector instantaneous frequency spectrum. In [12], an efficient method is presented for extracting defective bearing characteristics from the stator current of a loaded machine using the continuous wavelet transform (CWT) based on Morlet’s complex wavelet.
In particular, a feature extraction technique is proposed to cope with the fact that the peaks in the frequency spectrum corresponding to the fault components have very low amplitudes and are usually obscured by noise. For this reason, 2D and 3D scalograms of stator current signatures for both healthy and damaged bearings were used to characterize the defects in the time-frequency domain. A health indicator that exploits both a time domain feature (peak-to-peak values) and a frequency domain feature (maximum value of the FFT) is proposed in [13]. The obtained indicator is then used to classify gear and bearing faults through an adaptive neuro-fuzzy inference system.
In [14], data acquired from multiple current sensors are fused using a 2D convolutional neural network (CNN), with features extracted from the FFT spectra of the current signals. This paper proposes a method based on autoregressive (AR) spectral estimation of wavelet-processed motor currents.
Autoregressive modeling is a very popular tool for time series analysis and spectral estimation. It is well known that AR models provide more accurate spectral estimates with respect to classical Fourier-transform-based methods, like periodograms and the Blackman-Tukey method. This is mainly due to the fact that, unlike FFT-based methods, the estimated AR spectra (power spectral densities) do not have the sin(x)/x transform response characteristic of conventional windowed spectra and therefore, they do not suffer from sidelobe leakage and spectral smearing.
Moreover, the AR power spectral density (PSD) has been shown to be highly sensitive to system changes, making it particularly effective for fault diagnosis applications. The AR PSD procedure to perform online condition monitoring requires first estimating a nominal AR model from data gathered under healthy conditions and its PSD. Afterwards, online condition monitoring is achieved by continuously or periodically estimating the AR model from the online data, as well as its PSD, and eventual departures from the nominal AR PSD are detected through several possible methods, in our case the symmetric Itakura-Saito distance, also referred to as COSH distance.
An alternative procedure to quantify the dissimilarity between the current and nominal behavior could involve the Fourier spectra of the motor currents; however, this diagnostic process presents more difficulties compared to AR-based methods. As mentioned above, AR modeling is not applied directly to the raw current signals. Instead, a multiresolution analysis (MRA) of the motor currents is first performed using the discrete wavelet transform with Daubechies filters.
Each current is decomposed into approximation and detail components. In particular, suitable levels are selected based on the approximate entropy criterion, and the associated detail components are the signals to which AR spectral estimation is applied. The MRA preprocessing phase enables the separation of noise, disturbances, and variable torque effects from the current signals. Consequently, the effects of faults are amplified, and more robust detection thresholds can be defined.
In summary, the entire procedure combines the useful properties of the discrete wavelet transform (DWT) with the aforementioned advantages of AR spectral estimation. It is worth highlighting that the proposed method belongs to the so-called data-driven fault diagnosis approaches, as it does not rely on any physical knowledge of the electromechanical systems to be monitored.
Implementation and Validation
The proposed method has been validated by considering the effects of two very different situations on motor currents: unbalanced load and bearing failure. For the first case, real data were collected while performing electric cam movement tasks in a laboratory experiment. This is a very common situation, as many industrial machines rely on electric cams to perform complex tasks that require synchronization between the various mechanisms involved.
In the second case, data were obtained from a public domain database. The findings obtained from both scenarios substantiate the efficacy of the method in detecting and isolating faults. It is worth highlighting that the proposed approach differs from those previously described as it does not reference a specific machine component or fault type. Moreover, the parameters required to determine the reference model are computed by using only data collected under healthy conditions.
Online Condition Monitoring Phase
The first phase (parameter setting) is performed by using only current data collected under normal (healthy) operating conditions. A multiresolution analysis of the three-phase motor healthy current signals I1(t), I2(t), I3(t) is performed using the discrete wavelet transform (DWT). Each current is then decomposed into approximation and detail components. A suitable wavelet type dbM and detail level j are selected based on the approximate entropy criterion. The associated details djk(t),(k=1,2,3) are the signals from which features are extracted for fault diagnosis.
Suitable features for fault diagnosis are extracted through AR modeling of the selected details. A proper AR order p is first selected by means of some model selection criteria. Then, reference AR models are estimated, and the corresponding reference AR power spectral densities (PSDs) SAR0k(f),(k=1,2,3) are computed from the AR coefficients.
The second phase (online condition monitoring) is performed during online working conditions. The measured data are segmented into (possibly overlapped) frames. Each computed PSD SARk(f) is then compared with the corresponding reference PSD SAR0k(f) computed in the first phase to check if a system change is occurring. The comparison is made by means of the symmetric Itakura-Saito spectral distance (SISSD).
Wavelet Transform with Daubechies Filters
This section provides an overview of the first data processing phase, which employs the Wavelet Transform with Daubechies filters. The signal to which multiresolution analysis is applied is denoted by s(t), where s(t)∈{I1(t),I2(t),I3(t)}.
A multiresolution analysis (MRA) is defined as a collection of nested subspaces denoted by Vj, with j∈Z, which satisfy a set of properties. In particular, an MRA involves the successive projection of the signal s(t) to be studied in each of the approximation subspaces Vj. Since Vj⊂Vj+1, the approximation in Vj is coarser than that in Vj+1. The detail is the information removed when going from one approximation to another.
In the context of MRA theory, it is shown that there exists a function ψ(t), designated as the mother wavelet, which serves as a Riesz basis for Wj and is derived from the scaling function ϕ(t), which represents the basis for the subspace Vj. where gn=(−1)nh1−n and the sequences {hn}n∈Z and {gn}n∈Z can be interpreted as the impulse response of a low-pass digital scaling filter and a high-pass wavelet filter, having the same cutoff frequency.
The application of these filters to the original signal results in the generation of two new signals. The first of these comprises the lower frequencies, while the second comprises the higher frequencies. An additional pair of filters can reconstruct the original signal if they meet specific complementarity constraints. For this work, we have chosen Daubechies wavelets dbM, which have compact support and the highest number of evanescent moments for a given support width, making DWT analysis particularly attractive and feasible.
The above decomposition scheme (3) can be repeated more than once. At each iteration, the scheme decomposes the approximation resulting from the previous iteration. If the original sampled signal is s(t)=a0(t), t=0,…,N−1, then the trend and the detail series, after J iterations, are aJ(t) and dj(t), j=1,…,J,t=0,…,N−1 respectively. In other words, the information conveyed by a signal can be expressed as a set of details at varying resolutions (or scales) and a low-resolution approximation (or trend). In practice, these are computed by a fast recursive algorithm with a low computational cost, as described by Mallat. The MRA framework allows DWT to be executed with only the use of simple digital filters, thereby enabling its use in real-time applications as well.
Wavelet Filter Selection
The most important step of the preprocessing phase shown in Figure 1 is the procedure described in this subsection. The first practical challenge faced when undertaking a wavelet analysis is the selection of a wavelet filter among all possible ones. The selection of an appropriate wavelet depends on the specific analysis goal and the properties of the wavelet that are most suitable for achieving that goal.
In this particular case, the decision-making process is influenced by two factors. Firstly, the use of very short wavelet filters has the potential to introduce unintended artefacts into the analysis results. Secondly, the application of high-order wavelet filters may result in a reduction in the localization property of DWT coefficients while concurrently...
Conclusion
Motor current signature analysis (MCSA) has proven to be a highly valuable predictive maintenance tool. Although it is a relatively young, rarely utilized technology, it is rapidly gaining acceptance in industry today. MCSA is simply the process by which motor current readings are recorded and analyzed in the frequency domain.
Table 1: Comparison of Condition Monitoring Techniques
| Technique | Advantages | Disadvantages | Applications |
|---|---|---|---|
| Vibration Analysis | Widely used, well-established | Sensitive to sensor placement, susceptible to noise | Rotating machinery, gearboxes |
| Motor Current Signature Analysis (MCSA) | Non-intrusive, cost-effective, simple sensor installation | Requires substantial data, complementary to other techniques | Electric motors, driven mechanical systems |
| Infrared Thermography | Non-contact, visual representation | Affected by surface emissivity, limited to surface temperatures | Electrical components, thermal insulation |