Introduction
The three main components of CBM (Jardine et al., 2006) are data acquisition (i.e. collection and storage of machine health information), data processing (i.e. conditioning and data feature extraction/selection received) and decision-making (i.e. recommending maintenance actions through diagnosis and/or prognosis). Data processing can be divided into subsets such as feature extraction, feature selection, and machine fault diagnosis (i.e., training and testing data).
Machine Fault Diagnosis
The use of SVM and its extension for machine fault diagnosis has been summarized in a review article (Widodo and Yang, 2007). However, the use of SVM for the machine condition monitoring and fault diagnosis is still rare.
Feature-based Diagnosis
SVM has excellent generalization performance, so it can produce high classification accuracy for machine condition monitoring and diagnosis. In the process of feature-based diagnosis, after defining the features (statistical characteristics) from the original data, a big problem of dimensionality of the features appears.
Gear Box Condition Monitoring and Fault Diagnosis
Their relative efficiency in classifying bevel gearbox faults was compared and found that the average efficiency of ANN was more compared to PSVM. Until 2006, it was observed that SVM in machine condition monitoring and diagnosis tended to develop the expertise-oriented and problem-oriented domain.
Aim and Objective of the Present Work
The motivation for the present work arose from the above discussions and observations. The calculation of the overall prediction accuracy would be shown from the prediction results obtained from different domains and classifiers.
Organization of the Thesis
Chapter 5 discusses the generation of frequency domain statistical features from the FFT data. The classification performance of the classifier using frequency domain statistical features is presented at the same rotation rate as well as at interpolation and extrapolation rotation rates.
Introduction
Experimental Set-up and Experimentation
- Experimental Setup
- Tri-axial Accelerometer
- Constant DC Power Source Unit
- Data Acquisition System
The bearing housings were mounted rigidly on a rigid plate and with channels in turn connected to the rotor, the speed of the gearbox could be controlled in a manner shown (see Figure 2.4). The gearbox and its assembly are mounted on the base plate and are illustrated in Figure 2.5.
Measurement Procedure
The data acquisition system also had the ability to convert the time domain signal to the frequency domain. The time domain plot with three different colors (white, red and green) shows the response of the x, y and z axis direction. a) CT case in the time domain (e) CT case in the frequency domain.
Summary
Excel files were downloaded from the hard drive and copied to the computer running the classification code.
Introduction
SVM Classifier
Parameter Selection
The amount of nectar that can be obtained from the food source corresponds to the quality of the solution represented by that food source. The food source represents a possible solution of design variables (C, ,ν γ) and the nectar amount of a food source corresponds to the value of objective function (i.e. the accuracy of classification).
Summary
Introduction
Statistical Feature Extraction
The peak factor (T3) is defined as the ratio between the peak value and the RMS of the signal. If spikes occur in the time domain signal, this will result in an increase in the crest factor.
Simulation Results
Training and Testing at the Same Rotational Speed
The cross-validation accuracies in the network search technique for 30 Hz rotation speed for C-SVC and ν -SVC are shown in Figure 4.3(a-b) respectively. Fault prediction capability: After parameter optimization, the total datasets are used for final fault classification testing for GA and ABCA, and.
Training at Two Different Rotational Speeds and Testing at an
In the case of an intermediate speed range of 10 Hz, the classification accuracy (all classes) for the lowest speed is almost 35% and increases at the high speed to 94.17%. At a speed range of 5 Hz, the classification for various errors is over 76.67%.
Training at Two Different Rotational Speeds and Testing at Extrapolated
It is observed that at test rates of 15 Hz and 30 Hz for the 5 Hz range case, the GSM shows marginally good results, which is 1.16% and 0.83% more than the GA and ABCA techniques respectively . The lowest prediction accuracy (all classes) and the highest prediction accuracy (all classes) are the lowest and highest test accuracy (all classes) found in Table 4.1 and Table 4.4, respectively.
Summary
In the interpolation and extrapolation error predictions, it has been observed that in many cases the GA and the ABCA show similar results to the GSM. It is also observed that the prediction accuracy increases gradually with the increase of rotation speed.
Introduction
The error classification accuracy documented here shows better prediction accuracy than found in the time domain dataset in Chapter 4.
Statistical Feature Extraction
These total data sets are divided and used for training, testing or optimizing the parameters and for final testing in the simulation.
Simulation Results
Training and Testing at the Same Rotational Speed
The variation of the initial and final fitness values (i.e., the percent accuracy) with the population for C-SVC using GA and ABCA for 30 Hz rotation rates are shown in Figure 5.2(a) and Figure 5.2(b), respectively. The cross-validation accuracy in GSM for 30 Hz rotation speed for C-SVC and ν - SVC is shown in Figure 5.3(a−b) respectively.
Training at Two Different Rotational Speeds and Testing at an
The lowest prediction accuracy (all classes) and the highest prediction accuracy (all classes) are the lowest and highest testing accuracy (all classes), respectively, according to Table 5.1 and Table 5.2. For the case of speed extrapolation with time domain data, 4 cases (out of 6 cases) show better results compared to the frequency domain.
Summary
Introduction
It has been noted that there is an opportunity to investigate the classification of gear faults using the time-frequency data set. The extraction of time-frequency dataset (wavelet) from the time domain data is discussed here.
Feature Extraction
CWT Based Feature Extraction
For example, the best scale 4 (with CWT. coefficient 9.5591) is selected for the first data point in x-direction (as shown in Figure 6.4 (a)) for the Mexican hat wavelet family. The CWT-based functions for the CT case at 30 Hz are shown in Figure 6.7 (a-c) for the Mexican hat family and in Figure 6.8 (a-c) for the Morlet wavelet family.
WPT Based Feature Extraction
In this filtering method which is based on discrete wavelet transform "Db4" selects low-frequency high-frequency components for high-pass filtering and high-frequency low-frequency components for low-pass filtering. Once the best power level is obtained, various statistical characteristics (ie, standard deviation, skewness, and kurtosis) can be calculated for the WPT coefficients.
Simulation Results
Training and Testing at Same Rotational Speed
The variation of initial and final fitness values (i.e. the percent accuracy) with the population for C-SVC at rotation rates of 30 Hz is shown in Figure 6.12(a), (c) and (e) with the GA and Figure 6.12(b) , (d) and (f) with the ABCA for the CWT and WPT-based functions respectively). Hz rotational speeds on the C-SVC housing. a) 30 Hz (C-SVC with Morlet wavelet family using GA optimization).
Training at Two Different Rotational Speeds and Testing at an
For the intermediate rotation speed range of 5 Hz, at the lowest rotation speed the classification accuracy (all classes) is almost 68.33% and it increases at the high rotation speed to 100%. In the case of intermediate rotation speed range of 10 Hz, for the lowest rotation speed the classification accuracy (all classes) is almost 21.67% and it increases at the high rotation speed to 98.33%.
Training at two Different Rotational Speeds and Testing at an
The lowest prediction accuracy (all classes) and the highest prediction accuracy (all classes) are the lowest and highest test accuracy (all classes), respectively, as shown in Table 6.2 and Table 6.3. Initially, for each of the four classification cases (i.e., CT, MT, WT, and ND), the training data was provided at running speeds from 10 Hz to 30 Hz with an interval of 2.5 Hz, and then the multi-class classification possibility of two for these running speeds was classes of SVM were noted.
Summary
It is observed that at the same speed and interpolation speed range, the Morlet wavelet family has shown good prediction results. Again for the extrapolation speed range, the Morlet wavelet family performed well in three cases and one case WPT.
Introduction
The best prediction accuracies (all classes) for a rotational speed are plotted for a particular domain in rows, which is the best error prediction found in Chapters 4–6 by three optimization techniques (GSM, GA, ABCA). Overall uniform prediction accuracy (all classes) is equal to or better than the best accuracy (all classes) for all speeds considered except at 15 Hz, where the overall uniform prediction accuracy (all classes) is 96.25% and the time, frequency and time - frequency domain.
Comparison of Predictions in Different Domains
Estimation of Unified Fault Prediction Accuracies Based on Voting Strategies… 123
The first row of Table 7.1 describes the best prediction accuracies (all classes) found in Chapters 4-6 using time, frequency, and time-frequency domain data. The overall unified prediction accuracies (all classes) obtained from two classifiers and three domains are compared with the best prediction accuracies in time, frequency and time-frequency (wavelet) (which is shown in Table 7.7).
Summary
From the above discussion, it can be seen that the uniform prediction accuracies (all classes) found from the voting strategy are relatively good if we take the effect of the average prediction accuracies found using the three domains and two classifiers, with the exception of one case for an extrapolation rate range of 10 Hz. with a rotation speed of 25 Hz. This is an indication that the voting strategy is one of the promising strategies, instead of simply averaging the best prediction accuracies (all classes) to find the uniform prediction accuracies for better decision making.
Overview of the Present Work
Major Conclusions of the Present Work
Li Y., Chai Y.Z., Yin R.P., Xu X.M., 2005, “Fault diagnosis based on support vector machine ensemble”, in: Proceedings of the Fourth International Conference on Machine Learning and Cybernetics, 3309–3314. Gear fault detection using artificial neural networks and support vector machines with genetic algorithms”, Mechanical Systems and Signal Processing, 18, 625–.
Contribution of Present Work
Applicability and Limitations
The algorithm developed in the present work has been used for classification of the gear faults with vibration data in time domain or frequency domain or both. A lot of post-processing is required for the vibration data and therefore it can only be used in offline mode.
Future Scopes of Work
Xu Y., Wang L., 2005, “Fault diagnosis system based on rough set theory and support vector machine”, Lecture Notes in Computer Science, 980–988. Yuan S.F., Chu F.L., 2007, "Fault diagnosis based on support vector machine with parameter optimization by artificial immunization algorithm", Mechanical System and Signal Processing.