4 METHODS OF ANALYSIS
4.3 Signal processing methods
4.3.1 General
With advances in digital technology over the last several years, adequate data processing capability is now available on cost-effective, microprocessor-based, protective-relay platforms to monitor motors for a variety of abnormalities in addition to the normal protection functions.
Various electromechanical operational parameters able to provide rich information for diagnosis are analysed in steady state but also in transient state (especially at no-load starting) using approaches such as:
- frequency-domain analysis, - time-domain analysis,
- time-frequency domain analysis, - neural network,
- model-based techniques.
4.3.2 Analysis in steady state
The frequency-domain analysis for the detection and localisation of abnormal electrical and
mechanical conditions in electrical machines is widely used and accepted in the industrial environment. This popularity is due to the availability of the Fourier transform technique, as
the characteristics of the studied signals are more easily noticed in the frequency domain than in the time domain.
Advanced signal processing techniques, such as high-resolution spectral analysis are also
used since they may lead to a better interpretation of the spectra characteristic to specific machine abnormalities. In this respect, it was experimentally proved that stator current high- resolution spectral analysis is very sensitive to induction motor faults modifying main spectral components, such as voltage unbalance and single-phasing effects (Benbouzid et al. 1999).
Bi-coherence spectra were used by Li et al. (1995) to derive features that relate to the
condition of a bearing. The application of bi-spectral and tri-spectral analysis to the vibration
monitoring was also discussed by McCormick and Nandi (1999).
4.3.3 Transient state analysis
Elder et al. (1989) developed a technique which allows the detection and identification of specific faults within induction motors during the starting transient. During this short period the machine is under conditions of severe and electrical stress. This type of analysis at transient state is important especially for the detection of eccentricity since any UMP is a maximum at starting.
Since under transient conditions the frequency components indicative of faulty operations are non-stationary in both the time and frequency domains, other signal processing strategies applicable to time variants need to be applied. Time-frequency domain techniques use both time and frequency domain information, allowing for the investigation of transient features. A number of time-frequency domain techniques that have been proposed in literature are summarised by Ocak and Loparo (2001) as following:
- Short Time Fourier Transform (STFT),
- Spectogram- a representation of frequency components versus time (Burnett et al. 1995),
- Wigner Ville Distribution (Burnett et al. 1995),
Since in both the measurements and simulations the signals studied as possible fault indicators were captured in steady state, only the traditional signal processing technique based on Fourier transformation (FFT) was used in this thesis.
4.3.4 Artificial intelligence techniques
The condition monitoring and fault detection of electrical machines have moved in recent years from traditional techniques to Artificial Intelligence (AI) techniques. In the AI-based systems, several quantities are utilized as input signals such as, stator currents and voltages, electromagnetic fluxes, frame vibrations, etc. The AI techniques have numerous advantages over conventional fault diagnostic approaches; besides giving improved performance, these techniques are easy to extend and modify, and can be made adaptive by the incorporation of new data or information. The AI-based techniques may use expert systems, artificial neural networks, fuzzy logic, fuzzy-neural networks, genetic algorithms, support vector machines,
etc.
A review of the developments in the field of AI-based diagnostic systems in electrical machines and drives is provided by Vas (1999) and Filippetti et al. (2000b). Siddique et al. (2003) focus their review of various AI techniques to the induction motors, aiming more specifically to the stator winding fault detection.
From the multitude of AI-based diagnostic systems, the neural networks have a wide industrial applicability (Meireles et al 2003). A representative work dealing with the application and design of artificial neural networks for electrical motors fault detection is the one of Chow et al. (1993). Following, there will be presented few references to contributions dealing with the detection of various faults using the neural network approach.
Salles et al. (2000), show that a neural network approach is able to distinguish between load anomalies (oscillation torque, repetitive dip of torque) and rotor asymmetries (broken bars) of
cage induction motors. This is a very important issue since a load anomaly causes a sequence of speed and current spectral components that can be confused with the one produced by a cage-related fault.
Filippetti et al. (2005) report an induction machine rotor fault diagnosis based on a neural
network approach. After the neural network was trained using data achieved through experimental tests on healthy machines and through simulation in case of faulted machines, the diagnostic system was found able to discern between “healthy” and “faulty” machines. Tallam et al. (2000), present an on-line neural network based diagnostic scheme, for induction machine stator winding turn fault detection. This scheme is claimed to be insensitive to
unbalanced supply voltages or asymmetries in the machine and instrumentation. In addition, it is claimed that a turn fault can be detected in the early stage of development.
Huang et al. (2004) propose a scheme to monitor voltage and current space vectors simultaneously in order to monitor the level of air gap eccentricity in an induction motor. For
the amplitudes of eccentricity related components that change non-monotonically with the operating conditions, an artificial neural network is used to learn the complicated relationship and estimate corresponding signature amplitudes over a wide range of operating conditions. Li et al. (2000) use neural networks to perform motor bearing fault diagnosis based on the
extracted bearing vibration features. Computer-simulated data were first used to study and design the neural network motor bearing fault diagnosis algorithm. Actual bearing vibration data collected in real-time were then applied to perform initial testing and validation of the
approach. The results show that neural networks can be effectively used in the diagnosis of various motor bearing faults through appropriate measurement and interpretation of motor bearing vibration signals.
Support Vector Machine (SVM) is a novel machine learning method introduced in early 90's and it has been successfully applied to numerous classification and pattern recognition problems such as text categorization, image recognition and bioinformatics. SVM based classification scheme were designed for different tasks in cage induction motor fault diagnostics and for partial discharge analysis of insulation condition monitoring and were found highly competitive with, e.g., neural networks, which are widely studied also in the area of condition monitoring and fault diagnosis of electrical machines (Pöyhönen et al. 2002a, Pöyhönen et al. 2002b, Pöyhönen 2004).