163 Table 5.5 Fault diagnosis for various operating conditions of IM for the same speed and load case. 186 Table 6.3 Fault diagnosis for various operating conditions of IM for the same speed and load case.
Introduction
Importance of Study
Therefore, condition monitoring of rotary machines in industry is becoming more and more important for early detection of faults in machine components that are starting to fail and to avoid the possibility of catastrophic machine failure. Condition monitoring provides a continuous assessment of the health of machines and their components throughout their lifetime.
Condition Monitoring and Fault Diagnosis of Rotating Machines
The third step of CBM, i.e. making maintenance decisions, is an essential step in implementing a CBM program for the diagnosis and prognosis of machine failures. Xian and Zeng (2009) carried out intelligent fault diagnosis of the rotating machines based on the WPT and the SVM.
Condition Monitoring and Fault Diagnosis of Induction Motors
Various Types of Faults in Induction Motor
- Mechanical Faults in Induction Motors
 - Electrical Faults in Induction Motors
 
Electrical faults include stator winding fault (such as turn-to-turn fault, coil-to-coil fault, phase-to-phase fault, phase-to-ground fault, etc.), rotor-related faults (such as rod broken rotor, broken end ring, etc.) and faults related to the power supply (such as phase or voltage imbalance, single-phase, etc.). Rotor misalignment or misalignment (MR) is of two types, i.e., parallel and angular misalignments.
Condition Monitoring Techniques of Induction Motors
This monitoring technique can be used to detect air gap eccentricity and defects associated with the stator structure. By using appropriate data analysis algorithms, it is possible to detect changes in the vibration signal caused by faulty components and make decisions about the engine's status.
Artificial Intelligence based Fault Diagnosis of Induction Motors
2001) used the wavelet packet based features and ANN for fault diagnosis of broken rotor bar and air gap eccentricity in IMs. In other work, Widodo and Yang (2008) used the transient current CBM and wavelet SVM for fault diagnosis of IMs. 2011) presented a practical industrial application of SVM for mechanical fault diagnosis of IMs based on frequency domain vibration signals.
Outcome of the Literature Review
A single fault diagnosis system for the simultaneous detection of all possible faults in mechanical and electrical components and their severity levels is still unusual in the literature. Based on the existing literature, it can be summarized that several native waveforms are available that can be used in intelligent fault diagnosis. Fault diagnosis of IMs, when training and testing data available under different operating conditions used for diagnosis are not available in the literature.
Aim and Objective of the Present Work
In the case of waveform analysis, different mother waveforms will be used to investigate their impact on the performance of the current IM fault diagnosis methodology. The aim is to check the performance of the proposed diagnostic when limited information is available for training the classifier. Therefore, fault diagnosis will be performed for a wide range of IM operating conditions (i.e., load and speed) to check the robustness of the current methodology.
Organization of the Thesis
In Chapter 6, multi-fault diagnosis of IM is given based on frequency domain data. In Chapter 7, multi-fault diagnosis of IM is given based on CWT data. Multi-error diagnosis of IM is documented based on WPT data in Chapter 8.
Introduction
SVM Classifier
Binary SVM
The basic model of the SVM is a maximum margin classifier, which only works for linearly separable data in the feature space. The basic principle of the SVM is demonstrated in the two-dimensional plane as shown in Figure 2.1 and Figure 2.2. For the non-linear separation case, a hyperplane can be created by the SVM that allows the linear separation in the higher dimension.
Multiclass SVM
DAG training is similar to OVO; however, he used a rooted binary DAG for testing, which has k k( 1) / 2 internal nodes and (k1) leaves. Therefore, the objective is to select a pair (C,) which provides the best prediction performance of the SVM. Therefore, each instance of all training data is predicted once, so CV accuracy is the percentage of data that is correctly validated.
Wrapper Model for Feature Selection
It then uses a predefined learning algorithm to evaluate the performance of each feature one by one, returning the prediction accuracy corresponding to the feature.
Fault Diagnosis Procedure through Developed Methodology
In this study, 80% of the total data is used for training and the remaining 20% is used for testing. Now the training of the SVM is performed using the training data and the RBF kernel to build a model for further testing or final prediction of errors by selecting optimized SVM parameters. The final result of the fault diagnosis is obtained in the form of testing or prediction accuracy, i.e.
Summary
The second case of fault diagnosis is considered here to take care of situations where limited information is available about the operating conditions of the IM. In this study, the LIBSVM toolbox (Chang and Lin, 2011) was used in the MATLAB environment to implement multi-class SVM. CPU time for training and testing in different fault diagnosis cases is added in the respective chapters.
Introduction
Experimental Setup
- Machine Fault Simulator
 - Speed Controller or Variable Frequency Drive
 - Torque Controller or Magnetic Clutch
 - Fault Specifications and their Genration Procedure
 - Measurement Sensors .1 Tri-axial Accelerometer .1 Tri-axial Accelerometer
 - AC Current Probes
 - Tachometer or Photovoltaic Sensor with a Constant DC Power Source
 - Data Acquisition System
 
Mechanical faults include bearing fault (BF), unbalanced rotor (UR), bent rotor (BR) and faulty rotor (MR), while electrical faults include broken rotor bar (BRB), stator winding faults ( SWF) and phase unbalance fault (PUF) is considered (as shown in figure 3.6 and illustrated in table 3.2). MR can be considered as parallel misalignment or angular misalignment as shown in Figure 3.12. The tachometer or photovoltaic sensor was mounted near the coupling in the MFS to measure the angular velocity of the shaft as shown in Figure 3.15.
Experimental Procedure
Among the many analyzers (or signal acquisition module), the time capture module was used to capture signals in the time domain and the frequency capture module was used to capture signals in the frequency domain. In Experiment-1, data was acquired for 30 seconds, so 300 raw data sets (300 data sets 2000 sample points) were generated in the time domain for a particular fault and operating condition of the IM. Finally, raw data sets in the time and frequency domains from both experiments were stored on the system hard disk with individual speeds and loads for each IM fault condition, from which they could be retrieved for later processing.
Observations and Discussions
Time Domain Analysis of IM Faults
In other words, changes in the raw time-domain vibration and current signals are too small to detect, so comparing the raw time-domain signals of defective and healthy IMs is not effective to determine whether the component is behaving normally or not. shows signs of failure. Therefore, in order to reveal important information from time domain signals, a signal processing method that converts the raw signals into a suitable condensed form is required. Useful features that can be extracted from the vibration and current signals of IMs include RMS value, standard deviation, skewness, kurtosis, etc.
Frequency Domain Signal Analysis of IM Faults
These sidebands can be visualized in the vibrational spectrum; however, they are of small amplitude. In addition, SWF also produces defect frequencies such as (pfr, 2pfr, 4pfr,..) in the vibration spectrum. This is one of the main problems in SWF detection based on current spectrum analysis.
Challenges in the Time and Frequency Analyses of Vibration and Current for IM Faults Faults
However, for a complex system comprising various components, it is a challenging task to accurately estimate the harmonic component of the error and locate them in the spectrum. Furthermore, for electrical fault the current spectrum is very sensitive to the accuracy of motor slip and supply frequency measurement, because the fault frequencies or sidebands are a function of motor slip and line frequency. It is noted that the excitation frequency of the stator will dynamically change the position of the current harmonics that appear in the spectrum of the stator current due to electrical faults. It is highly dependent on the mechanical load of the motor and the excitation frequency, which affects the slip frequency.
Summary
Introduction
Fault diagnosis is performed by training and testing the SVM with signals collected at the same IM speed and load. Therefore, to verify the effectiveness of this method in fault prediction, fault diagnosis is considered for a wide range of IM operating conditions (i.e., three different loads with seven IM speeds). In addition, fault diagnosis performance is compared for high sampling rate data with low sampling rate data.
Fault Feature Extraction
The standard deviation is the second standardized moment (or normalized central moment) of the data and is expressed as It is clear from scatterplots of the vibration and current characteristics that values of characteristics are linearly inseparable, i.e. Now in the next section, a comparative analysis of the vibration and current signals is performed in the mechanical and electrical fault diagnosis based on the SVM.
Fault Diagnosis based on SVM
Fault Diagnosis for the Same Speed and Load Case
Mechanical failure prediction based on vibration signal alone: Figure 4.5 shows the prediction of mechanical failures for different operating conditions of IM using vibration. Electrical fault prediction based on vibration signal alone: Figure 4.9 shows electrical fault prediction for different operating conditions of IMs using vibration signals. Electrical fault prediction based on current signal alone: Figure 4.10 shows electrical fault prediction for different operating conditions of IM using current signals.
Comparison of Fault Diagnosis based on Data of Experiment 1 and Experiment 2
The features extracted from the two experimental time domain data are added into feature clusters, and it is found that the more overlapping features appear in the feature space for data from Experiment-1 compared to data from Experiment-2, as shown in Figure 4.18 and Figure 4.19. Therefore, when the characteristics of the data from Experiment-1 are used for fault diagnosis, the prediction performance significantly decreases. Moreover, it increases the implementation costs, which is very undesirable in fault diagnosis.
Summary
In addition, this chapter compares the error prediction performance of high-sampling data with low-sampling data. Fault prediction and detection time were found to be slightly improved with low sampling rate data obtained from experiment 2. Fault diagnosis using low sampling rate time domain data is further expanded in the next section.
Introduction
To verify the robustness of the present fault diagnosis methodology, the diagnosis is performed for a wide range of IM speeds and loads. Furthermore, in this chapter, the work is extended to perform the diagnosis at different speeds and loads as a classifier training. The objective of this is to verify the predictive ability of the proposed diagnostics for limited training information.
Fault Feature Extraction
The odd moments reflect the position of the peak of the PDF relative to the mean, and the even moments are proportional to the spread of the distribution. Crest factor (CF): Crest factor is defined as the ratio of the peak value to the RMS of the signal. Ratio of Mean to Standard Deviation (Rmsd): It is defined as the ratio of the mean value to the standard deviation of the signal.
Fault Diagnosis based on SVM
SVM Parameter Selection and Training
The CV accuracy in the grid search technique for a typical case of 40 Hz and T3 load, when the training is performed with all the features individually or in sets, is shown in Figure 5.1 and Figure 5.2. The minimum CV accuracy (69.5 %) occurs when the model is trained with the crest factor (CF) as shown in Figure 5.2 (b). The maximum accuracy (100 %) comes with the standard deviation used individually, and in combination with the skewness and the kurtosis as shown in Figure 5.2 (d) and Figure 5.2 (g) respectively.
Clog3C
Fault Feature Selection
In the previous sections, fourteen features were initially extracted from the vibration and current signals representing IM faults. Therefore, a silent function should be selected from the feature set to provide error information in the IM. The result shows that all considered features successfully perform IM fault diagnosis with more than 90% accuracy, except the eighth moment (88.8%) and CF (68.
Fault Diagnosis for the Same Speed and Load Case
The matrix shows that for T1 loading and 15 Hz, 20% of the SWF2 data is misclassified with the SWF1 data. The confusion matrix shows that 20% of the PUF2 data is misclassified with the PUF1 data. In addition, the two severities of the phase unbalance as well as stator winding fault are also successfully classified under most operating conditions.